# Black Box Personalization > Opening the black box of personalization. The longer companion to [llms.txt](https://www.blackboxpersonalization.com/llms.txt). Each public page is listed once, with its canonical URL. The public origin is https://www.blackboxpersonalization.com. Pages teach with the live mechanism. Black Box Personalization is a Black Box Site Design company. Black Box Site Design builds the sites. This site makes them smarter. This file is generated when the site is built, from the public page catalog, the lab groups, and the markdown notes in content/blog. It is not edited by hand. This file lists public pages only. It omits /admin, /admin/usage, /account, /api, sign-in, registration, and the embed frame. A read-only MCP server answers POST /api/mcp with server/discover, tools/list, and tools/call. The tools are list_labs, get_lab, search_blog, example_report, inspect_datalayer, journey_catalog, journey_next. Catalog results are public fields from this site. inspect_datalayer returns dataLayer events this origin collected for the caller's own browser session when that session id is passed, and explains how to get one when it is not. It does not return another visitor's events. example_report is seeded and is not a live GA4 property. journey_catalog and journey_next name the advanced labs. The question is not logged and writes are refused. robots.txt disallows /api except /api/mcp and the public status document at /api/status. The public card and catalog are linked below. Ask and AI copy open in manual mode. AI-assisted is optional and needs OPENAI_API_KEY. The Location lab maps the city from Vercel request headers even when the database cannot store the beacon. When the database is over quota, stored counts say live data is temporarily unavailable. The page still renders. The status page lists whether the application, the database, and the internal APIs are operational. The same report is JSON at the /api/status path, which is not a listed page. Public pages other than the notes and the white paper are also published in Spanish under the /es prefix, with hreflang alternates. The notes and the white paper stay in English, and their canonical URL is the unprefixed one. English URLs stay unprefixed. The language choice is a first-party cookie. Google Tag Manager loads only after Accept. Google Analytics 4 is a tag inside that container. Microsoft Clarity (project z1z7ayg6c7) loads on that same Accept and stays off before it. Navigations push virtual_page_view. Advertising storage stays denied. Global Privacy Control forces measurement off. Vercel Speed Insights is cookieless and runs on the public pages without waiting for Accept. The script and the vitals post stay on this host. Resend sends contact and white-paper mail only when it is configured. OpenAI, IndexNow, and Search Console are server-side and optional. Admin routes, account pages, API routes, and the embed frame are not listed. ## Related - [Black Box Site Design](https://www.blackboxsitedesign.com): Parent company. It designs and builds fast, accessible, measurable marketing sites. Tagline: Websites engineered to perform. This site makes those sites smarter. The organization graph points at it with parentOrganization. ## Start here - [Black box, Google Analytics](https://www.blackboxpersonalization.com): Opening the black box of Google Analytics. A free site that shows how a visit is measured, and how those same signals can change the page you see. Spanish: [Caja negra y Google Analytics](https://www.blackboxpersonalization.com/es) - [GA4 realtime active users](https://www.blackboxpersonalization.com/live): GA4 realtime for the last 30 minutes: users, events, key events, pages, cities, countries, and devices. Google's third-party view of the same visits. Spanish: [Usuarios activos de GA4 en tiempo real](https://www.blackboxpersonalization.com/es/live) - [Tracking & personalization notes](https://www.blackboxpersonalization.com/blog): Tracking & personalization notes: what a beacon sends, what a browser reveals, and how those signals change a page. Free, with the labs beside each idea. Spanish: [Notas de medición y personalización](https://www.blackboxpersonalization.com/es/blog) - [Opening the black box, free](https://www.blackboxpersonalization.com/white-paper): A field guide to measurement and personalization in the browser, the middleware, and the server. Read it free, then request the 30 October 2026 PDF. - [Black box of Google Analytics](https://www.blackboxpersonalization.com/about): The black box of Google Analytics, opened in public. A free hands-on look at collection and personalization, and a place to ask for the next lab. The about page lists resume roles, highlights, and core competencies, plus certifications from the public LinkedIn profile, hobbies, and links to LinkedIn and Instagram. Phone, personal email, and street address are not published. Spanish: [Caja negra de Google Analytics](https://www.blackboxpersonalization.com/es/about) - [Contact us about the labs](https://www.blackboxpersonalization.com/contact): Contact the site about the labs, a page, or an idea for the next one. The form is stored on this site and does not depend on measurement consent. Spanish: [Escribe sobre los laboratorios](https://www.blackboxpersonalization.com/es/contact) - [Live production status](https://www.blackboxpersonalization.com/status): Live production status for this site: the application, the database, and the internal APIs, each marked operational, degraded, or down, with its response time. Public production status. The page and /api/status report the application, the database, and the internal APIs. A database quota refusal is degraded. A database that does not answer is down. The page still renders. Checks are reused for about 45 seconds. No connection string or host name is included. Spanish: [Estado de producción en vivo](https://www.blackboxpersonalization.com/es/status) ## Getting started - [Free Google Analytics labs](https://www.blackboxpersonalization.com/lab): Free Google Analytics labs in this browser. See what a page collects, what a beacon sends, and how those signals personalize what you see. Open one. Spanish: [Laboratorios gratis de Google Analytics](https://www.blackboxpersonalization.com/es/lab) - [Using the Chrome Sources panel](https://www.blackboxpersonalization.com/lab/sources): Using the Chrome Sources panel to pause a practice score, step into the helpers, and watch the offer stay off. Free in the Chrome you have open. A practice script calls one fixed sample, hour 14 and width 1280, and draws the return in this tab. The button does not post that object. How to use it: - The practice file loads on its own. - The button calls that one fixed sample. - Compare the return with 40 and with the offer. - The result stays in this tab. It is not posted. Spanish: [Usar el panel Sources de Chrome](https://www.blackboxpersonalization.com/es/lab/sources) - [Call an API from this page](https://www.blackboxpersonalization.com/lab/calls): Call a public API from the page you have open. Edit a small script, ask this site for data, and read the response. The practice calls do not write anything. Two same-origin GETs, the data-center search and the collect list, run only after the saved JSON sets callApis to true and both paths pass the guard. Anything else is refused. How to use it: - Open the two-request document on the page. - Each call is a relative path, not an outside host. - A path that is not those two GETs is refused. - The network stays quiet until callApis is true. Spanish: [Llamar a una API pública desde la página](https://www.blackboxpersonalization.com/es/lab/calls) - [Your GA4 property, read-only](https://www.blackboxpersonalization.com/lab/your-data): Your GA4 property, read-only. Journey, attribution, and a small twin use your data when Google sign-in is set. Until then, the page shows labeled example data. You want the journey and the channel credit from your own property. Sign-in is a limited preview, the tables stay an example until you connect, and the list stores an email only if you check the box. How to use it: - Connect Google, or read the example tables. - Pick a property and run the three views. - Revoke clears the cookie. The email box is opt-in. - Sign-in is a limited preview. The list is for everyone else. Spanish: [Tu GA4, solo lectura](https://www.blackboxpersonalization.com/es/lab/your-data) - [Why am I seeing this page](https://www.blackboxpersonalization.com/lab/why): Why am I seeing this page? The header control labels each personalized slot with the rule and the signals that fired it. Free to turn on across the site. Labels for the whole site can be turned on, and after Accept that choice is stored in this browser. The event is queued only when the labels turn on, and the four slot explanations are not sent. How to use it: - The stored choice is read only after Accept. - Turning the labels on writes the key, and only after Accept. - why_labels_view is queued only on that turn-on. - Four slots are explained on the page and not sent. Spanish: [Por qué veo esta página](https://www.blackboxpersonalization.com/es/lab/why) ## Tracking & data layer - [What dataLayer.push does](https://www.blackboxpersonalization.com/lab/datalayer): What dataLayer.push does, and how window.dataLayer becomes a Google hit. Watch the queue, the tag, and the collect request with consent on or off. Try it walks one push through the action, the queue, the GA4 tag, the consent check, and the collect request. Consent off keeps the event on the page, and consent on uses the same queue as the rest of the site. How to use it: - Pick an event the site already knows, then press Try it. - Five stages advance one at a time. - Consent off shows the push and stops before the hit. - Consent on decodes the collect hit. Advertising storage stays off. Spanish: [Qué hace dataLayer.push](https://www.blackboxpersonalization.com/es/lab/datalayer) - [Browser and server-side](https://www.blackboxpersonalization.com/lab/server-side): Browser and server-side collection of the same test event, from gtag and from a first-party endpoint, and what blockers drop. Free to send both. The browser pushes server_side_test, and only the consent flag is posted to this site, which hashes the first-party visitor cookie and forwards the event when a Measurement Protocol secret is set. After Accept the transport is queued, not the client id. How to use it: - Consent off refuses the copy. - The client id is a number built from the first-party cookie. - Google is called only when the server secret is set. - After Accept, server_side_view carries the transport, not the client id. Spanish: [Navegador y lado servidor](https://www.blackboxpersonalization.com/es/lab/server-side) - [Triggers, tags, and variables](https://www.blackboxpersonalization.com/lab/tag-manager): Triggers, tags, and variables in a simulated container. A click shows what would fire. No Google Tag Manager script is loaded. Free to step through. A local click is matched against three triggers, and demo_click and file_download payloads are built in memory. Google Tag Manager is not loaded, and after Accept the count of tags is queued. How to use it: - Three variables are filled on the page. - Click and watch which triggers match. - The tags are built and not fired. The container is not loaded. - After Accept, tag_manager_view carries how many tags matched. Spanish: [Activadores, etiquetas y variables](https://www.blackboxpersonalization.com/es/lab/tag-manager) - [What GA4 collects on its own](https://www.blackboxpersonalization.com/lab/enhanced-measurement): What GA4 collects on its own: scroll, outbound clicks, file downloads, site search, and video. This site sends scroll itself. Free to compare the two. Scroll, click, download, search, and video shapes are written into a log on this page. After Accept the lab queues one view with no parameters, and it does not push those log names. How to use it: - Scroll inside the frame. Ninety percent is noted there. - Click, download, search, and video shapes log without leaving. - The automatic scroll push is dropped. The log stays local. - After Accept, only the lab view is queued. Spanish: [Lo que GA4 recoge por su cuenta](https://www.blackboxpersonalization.com/es/lab/enhanced-measurement) - [GA4 ecommerce events, demo store](https://www.blackboxpersonalization.com/lab/ecommerce): GA4 ecommerce events, demo store. View, add to cart, begin checkout, and purchase show the payloads. Nothing is charged. Free to click each step. A 24 dollar notebook walks through view, cart, checkout, and a demo purchase. Nothing is charged, and after Accept each step is pushed as a GA4 ecommerce event and as ecommerce_demo. How to use it: - Step through view, add to cart, checkout, and purchase. - The purchase is the step that adds a transaction id. - Consent off still shows the payload, then does not push it. - Consent on pushes the GA4 event and ecommerce_demo. Ads stay denied. Spanish: [Eventos de ecommerce de GA4](https://www.blackboxpersonalization.com/es/lab/ecommerce) - [SPA tracking for route changes](https://www.blackboxpersonalization.com/lab/spa-tracking): SPA tracking for route changes: a client route can miss page_view, and a history listener sends the next one. Free to compare both paths side by side. This lab keeps its own route stack and writes a page_view line only in history mode, without changing this page's address and without pushing that demo page view. After Accept it queues the mode. How to use it: - Switch routes. The address of this lab does not change. - A page_view line is written only in history mode. - That demo page view is not pushed to the dataLayer. - After Accept, spa_tracking_view carries the mode. Spanish: [Seguimiento SPA de cambios de ruta](https://www.blackboxpersonalization.com/es/lab/spa-tracking) - [Form analytics without answers](https://www.blackboxpersonalization.com/lab/forms): Form analytics without answers: focus, time, skips, abandonment, and corrections. The characters you type are not stored. Free to try the demo. Focus, skip, correction, submit, and abandon are counted as step names. The characters you type are not sent, and after Accept the lab view has no parameters. How to use it: - Focus a field. That starts the step count. - A correction and a submit are separate step names. - Leaving the form counts as abandon. - After Accept, form_lab_view is queued with no fields. Spanish: [Analítica de formularios sin las respuestas](https://www.blackboxpersonalization.com/es/lab/forms) - [GA4 event quality checker](https://www.blackboxpersonalization.com/lab/event-quality): GA4 event quality checker. Paste a dataLayer push and see bad names, missing parameters, personal data, and size limits. The paste is not forwarded. Paste a JSON event and this page lists findings about the name, the parameters, and personal data. After Accept it queues the finding count, not the paste. How to use it: - Paste an object or an array. Parsing stays in this browser. - Watch the name check and the size check. - Personal data and missing fields are flagged on the page. - After Accept, finding_count is queued. The paste is not pushed. Spanish: [Revisor de calidad de eventos GA4](https://www.blackboxpersonalization.com/es/lab/event-quality) - [Web pathing for one visit](https://www.blackboxpersonalization.com/lab/pathing): Web pathing for one visit: the order of pages in this tab, and a count of hops from consented sessions. Nothing here is a third-party profile. This tab's first-party trail of page views, clicks, scroll milestones, and form field names is listed, then page-to-page hops are counted. The lab does not queue an analytics event of its own. How to use it: - The trail waits until consent is on. - Events in this tab are numbered in order. - Field values stay out of the record. Names can appear. - The hop count is page to page, on this page. Spanish: [El recorrido web de una visita](https://www.blackboxpersonalization.com/es/lab/pathing) - [Scroll and click heatmap](https://www.blackboxpersonalization.com/lab/heatmap): Scroll and click heatmap from consented beacons: depth, click counts, and a labeled sample when the stored set is thin. Free to read on this page. Consented scroll depths and clicks are totaled by path. Stored views have no x and y, so thin data switches to sample dots, and after Accept only the data source is queued. How to use it: - Only granted rows are kept. - Scroll is bucketed by percent. Clicks are counted by path. - A thin set switches to the labeled sample dots. - After Accept, the event carries the source, not the dots. Spanish: [Mapa de calor de scroll y clics](https://www.blackboxpersonalization.com/es/lab/heatmap) - [Session replay with masking](https://www.blackboxpersonalization.com/lab/session-replay): Session replay with masking: what a replay records, how field values are hidden, and a recorder that stays in this tab. Free, and nothing is uploaded. A short timeline of this tab is masked before it is stored, and at most 20 beats are kept. After Accept the queued flag is whether masking was on, not the text. How to use it: - Type in the demo. The timeline stores the masked value. - The log keeps at most 20 beats, in this tab. - The beats are not sent as the event payload. - After Accept, session_replay_view carries the mask flag. Spanish: [Reproducción de sesión con enmascarado](https://www.blackboxpersonalization.com/es/lab/session-replay) - [Video events from one player](https://www.blackboxpersonalization.com/lab/video-tracking): An original eight-second clip plus an optional YouTube embed after consent. Milestones, watch time, and dataLayer events stay on this page. No model. An eight-second clip emits start, milestone, complete, and engagement through the dataLayer, and resume seconds are kept in localStorage only after Accept. The lab does not call Adobe. How to use it: - Play the clip. video_start is sent once. - A milestone fires once. A seek is classified on the page. - Attention is scored locally. - A resume second is saved only after Accept. Spanish: [Eventos de un reproductor](https://www.blackboxpersonalization.com/es/lab/video-tracking) - [Cross-domain tracking links](https://www.blackboxpersonalization.com/lab/cross-domain): Cross-domain tracking links: what happens to a visitor id, a UTM, a referrer, and a _gl parameter across a redirect or another domain. Free to step through. One simulated hop shows whether a cookie, a UTM query, the referrer, and a toy linker arrive. After Accept the page queues cross_domain_view with no parameters, not the hop itself. How to use it: - Start from one lab URL. - Change the host or the path and watch what still arrives. - The referrer policy decides whether the referrer is kept. - A toy linker decorates the URL. The view is not the hop. Spanish: [Enlaces de seguimiento entre dominios](https://www.blackboxpersonalization.com/es/lab/cross-domain) - [How a headless browser sees](https://www.blackboxpersonalization.com/lab/puppeteer): How a headless browser sees this site: what Puppeteer is, how a Node script drives Chrome, and the GA4 hits one real run captured. Free to read the trace. This page shows the script and the saved output of one headless Chrome run that opened this site, clicked Accept, and read the dataLayer. The page does not launch a browser. How to use it: - Read the launch in the script. It is not running now. - The script listens before navigation, then clicks Accept. - The saved output is the dataLayer read inside that page. - Nothing on this visit starts Chrome. Spanish: [Cómo ve el sitio un navegador headless](https://www.blackboxpersonalization.com/es/lab/puppeteer) - [Bot traffic filtering explained](https://www.blackboxpersonalization.com/lab/bots): Bot traffic filtering explained: crawler, headless, and high-rate beacons stay stored, and the Understand reports skip them. Free to inspect the rules. Stored beacons are split into kept rows and rows marked crawler, headless, or too fast. After Accept the page queues bots_view with the data source. How to use it: - A known crawler name is one mark. - A headless browser is a second mark. - A fast repeat is a third mark. - Four counters stay on the page. After Accept, the source is queued. Spanish: [Filtrar el tráfico de bots](https://www.blackboxpersonalization.com/es/lab/bots) - [What are webhooks used for](https://www.blackboxpersonalization.com/lab/webhooks): What are webhooks used for: a signed POST when something happens, how to verify the signature, and a live delivery log for events from this site. Free to try. The demo receiver checks the signature on this site. A custom https destination is refused when it points at a private address, localhost, or this host. A JSON body is signed, the HMAC and a five-minute timestamp are checked, and a forced first failure is retried once with the same delivery id. After Accept a send queues the event name, not the URL. How to use it: - Send signs the raw body. - A stale timestamp or a bad signature is rejected. - The first built-in failure retries once, with the same delivery id. - After Accept, webhook_lab_view carries the event name, not the URL. Spanish: [Para qué sirven los webhooks](https://www.blackboxpersonalization.com/es/lab/webhooks) ## Consent & privacy - [Cookies, storage, and consent](https://www.blackboxpersonalization.com/lab/cookies): Cookies, storage, and consent on this origin: who owns each key, how long it lasts, and which services are not cookies. Free, and Speed Insights sets none. Vercel Speed Insights sets no cookie. The page also names Google Analytics, Vercel geo headers, Neon, Resend, optional OpenAI, IndexNow, and Search Console. This page lists storage names in this browser and matches them to a catalog without printing the values. After Accept it queues cookies_view with no parameters, once the list has loaded. How to use it: - Read the names script can see, not the values. - Cookie names from the server are matched to the catalog. - HttpOnly names are marked even though script cannot read them. - After Accept, the view is queued with no fields. Spanish: [Cookies, almacenamiento y consentimiento](https://www.blackboxpersonalization.com/es/lab/cookies) - [Consent Mode and Global Privacy](https://www.blackboxpersonalization.com/lab/consent): Consent Mode and Global Privacy Control on this site: what is sent before Accept, what changes after, and how a GPC signal keeps measurement off. Google Analytics waits for Accept. Speed Insights does not, and it sets no cookie. Global Privacy Control forces measurement off. Before Accept, the Google tag and Clarity stay unloaded. After Accept, analytics storage is granted, advertising storage stays denied, and a sample click is pushed only when consent is on. How to use it: - Decline leaves analytics and ads storage denied. - Accept grants analytics storage only. - Global Privacy Control blocks that grant. - Watch the requests that leave, and the lab view after Accept. Spanish: [Consent Mode y Global Privacy Control](https://www.blackboxpersonalization.com/es/lab/consent) - [Local storage personalization](https://www.blackboxpersonalization.com/lab/storage): Local storage personalization: the keys this site writes, what each one changes on the page, and how to edit or delete them. Free to inspect in this browser. The keys this origin actually has are listed, including the visitor and the session written in two stores. Consent off leaves those two unread for personalization, and this lab sends no event of its own. How to use it: - Read the stores this origin actually has. - The visitor and the session are two different stores. - Consent off skips those records for personalization. - This lab does not queue an event. Spanish: [Personalización con almacenamiento local](https://www.blackboxpersonalization.com/es/lab/storage) - [What a fingerprint could see](https://www.blackboxpersonalization.com/lab/fingerprint): What a fingerprint could see without cookies: traits calculated in this tab and not stored. A free reading of the browser, dropped when you close the page. A small canvas drawing is hashed, and a few local traits are read, all inside this tab. After Accept the page queues fingerprint_view with no parameters. How to use it: - Watch a few font widths sampled in this browser. - The canvas hash comes from a fixed drawing. - The other traits are read from this window. - After Accept, the view is queued with no fields. Spanish: [Lo que podría ver una huella](https://www.blackboxpersonalization.com/es/lab/fingerprint) - [When the Google tag is blocked](https://www.blackboxpersonalization.com/lab/blocking): When the Google tag is blocked, this page says so for this browser and shows the share of lab visitors whose tag did not load. Free, and the check stays here. Resource Timing shows whether gtag.js came back, and a failed request is treated as blocked. After Accept the page posts that result and queues blocking_view so the totals can include beacons and loaded tags. How to use it: - Look at Resource Timing for a gtag.js response. - A failed request counts as blocked. - The totals are beacons and loaded tags. - After Accept, blocking_view carries the blocked flag. Spanish: [Cuando la etiqueta de Google se bloquea](https://www.blackboxpersonalization.com/es/lab/blocking) - [Consent the agent must honor](https://www.blackboxpersonalization.com/lab/privacy-negotiator): See what an agent would and would not do at each consent level. A location trade is reversible. Sensitive inferences stay off. No model call. Measurement, personalization, and a location trade become a list of what an agent may do. Sensitive categories and identity stay off, and after Accept the choice can be stored. How to use it: - Change a trade and watch the allowed list rewrite. - Sensitive categories and identity are not offered. - The choice is stored only after Accept. - After Accept, lab_interaction is queued for that choice. Spanish: [Consentimiento del agente](https://www.blackboxpersonalization.com/es/lab/privacy-negotiator) - [App events and a webview](https://www.blackboxpersonalization.com/lab/app-tracking): App events and a webview: how a screen view differs from a web hit, and how an app instance id reaches a page only if the app puts it on the URL. A pretend phone log keeps screen names on the page, and opening the site appends a webview URL that can carry a simulated app instance id or a demo user id. After Accept the view is queued with no parameters. How to use it: - The log starts on this page. Screen views are prepended. - An id is passed only on the webview URL. - The screen names stay on the page. - After Accept, app_tracking_view is queued with no fields. Spanish: [Eventos de app y un webview](https://www.blackboxpersonalization.com/es/lab/app-tracking) ## Personalization & experiments - [Same page, five people](https://www.blackboxpersonalization.com/lab/personas): Same page, five people, plus you. Each card names the signals and the rule that changed the hero, the offer, and the notes. Tweaks stay in this browser. Five fixed visitors and a You card run the same homepage rules. Tweaks stay in the browser and are not stored. Consent off uses device, referrer, and local time only. The same homepage rules run for five fixed visitors and for a You card whose device, referrer, hour, pages, and visits you can change. Consent off drops the stored history from that card. How to use it: - Compare the five fixed visitors. Their inputs do not change. - Edit the You card. Device, referrer, hour, pages, and visits apply. - The homepage rules are what write each card's result. - Consent off clears stored history from the You card. Spanish: [La misma página, cinco personas](https://www.blackboxpersonalization.com/es/lab/personas) - [See live audience segments](https://www.blackboxpersonalization.com/lab/segments): See live audience segments built from signals already in this browser, plus a count of consented visitors. The rules are on the page. No third-party profile. Engaged reader and Night owl can change the home page call to action. After Accept, only the on/off flags are stored, not a third-party profile. How to use it: - On and Off match this browser right now. - Resize. Mobile and Desktop flip with the width. - Consent off leaves New visitor and West Coast unread. - The bars below are anonymous counts, not this visit. Spanish: [Segmentos de audiencia en vivo](https://www.blackboxpersonalization.com/es/lab/segments) - [Next-page recommendations](https://www.blackboxpersonalization.com/lab/recommendations): Next-page recommendations from consented page-to-page transitions, with the counts behind each pick. Free to see why a page was chosen or marked default. Consented page-to-page sessions are counted, then the next pages and the pages from the same session are listed. The matrix stays on the page. How to use it: - The counts are consented page views only. - One column is the next page other visits opened. - The other column is pages seen in the same session. - The matrix is not sent as an analytics event. Spanish: [Recomendaciones de la página siguiente](https://www.blackboxpersonalization.com/es/lab/recommendations) - [An A/B test and a bandit](https://www.blackboxpersonalization.com/lab/experiments): An A/B test and a bandit on two headlines. One splits traffic in half. The other spends more impressions on the headline that converts. Free aggregate counts. An A/B arm is pinned with a hash, and a bandit arm is drawn once per tab with Thompson sampling. After Accept both arm ids are queued, and impression clicks are counted on this site. How to use it: - The A/B arm stays put for this first-party key. - The bandit arm is drawn once in this tab. - A click on an impression counts as a conversion here. - The simulate controls rerun the draw on the page. Spanish: [Un test A/B y un bandit](https://www.blackboxpersonalization.com/es/lab/experiments) - [Progressive profiling explained](https://www.blackboxpersonalization.com/lab/profiling): Progressive profiling explained: role, then goal, then industry. The answers are zero-party data and can be reset. A free look at how the headline changes. Role, then goal, then industry build the headline from those three answers. After Accept the answers can be stored, and the queued event carries the step name only. How to use it: - Answer one question at a time. - The headline is written from the three answers. - Reset clears the answers. - After Accept, the event names the step, not the answers. Spanish: [Perfilado progresivo, explicado](https://www.blackboxpersonalization.com/es/lab/profiling) - [Visitor intent scoring explained](https://www.blackboxpersonalization.com/lab/intent): Visitor intent scoring explained. Scroll, dwell, return visits, and page depth add up. The white paper offer appears past 40. Free to watch the weights move. Scroll, dwell, return visits, and pages in this session make a score, and a link appears at 40. After Accept only a boolean is queued, not the number. How to use it: - Scroll and dwell are read even before Accept. - Return visits and page history are read only after Accept. - The link stays hidden while the score is under 40. - After Accept, intent_ready is queued. The score is not. Spanish: [Puntuación de intención del visitante](https://www.blackboxpersonalization.com/es/lab/intent) - [Exit-intent popups, done right](https://www.blackboxpersonalization.com/lab/exit-intent): Exit-intent popups, done right: a small note when the pointer leaves at the top, or a phone scrolls up fast. Capped for 12 hours. Free to try. A mouse leave, a fast scroll, or a touch can count as exit intent, and a live match is kept for 12 hours. After Accept the trigger name is queued. How to use it: - With consent off, the listeners stay quiet. - One of the three gestures can light a match. - The 12-hour cap is written, then the event. - The buttons can skip that cap so you can try again. Spanish: [Avisos de intención de salida](https://www.blackboxpersonalization.com/es/lab/exit-intent) - [Personalization by time and day](https://www.blackboxpersonalization.com/lab/context): Personalization by time and day: the local hour, the weekday, and a holiday table choose the headline. Free rules that do not need a profile. The clock on this device can pick a holiday, the weekend, or a time of day. Try buttons stay on the page, and after Accept only the live rule id is queued. How to use it: - Watch the holiday, weekend, or hour from this clock. - A listed date wins over the weekend, then the hour. - Try buttons preview a rule without replacing the live clock. - After Accept, the queued id is the live rule only. Spanish: [Personalizar por hora y por día](https://www.blackboxpersonalization.com/es/lab/context) - [Weather and time of day](https://www.blackboxpersonalization.com/lab/weather): Weather and time of day pick the hero, the theme, and a suggested read from a city-level forecast and the local hour. Free, and the Open-Meteo request is shown. Open-Meteo is called from the server and cached about 10 minutes. Coordinates are not stored. Consent off skips any precise point and still uses the city on the request. If the forecast fails, the page says weather is unavailable and keeps the local hour. A forecast is matched to the first rule in a fixed list, then the local hour finishes the scene. A preset can skip the forecast, and after Accept the matched rule id is queued. How to use it: - The first matching weather row wins. - The local hour finishes the scene. - A preset shows a scene without calling the forecast. - After Accept, weather_personalization_view carries the rule id. Spanish: [Clima y hora del día](https://www.blackboxpersonalization.com/es/lab/weather) - [Traffic source, referrer decoded](https://www.blackboxpersonalization.com/lab/source): Traffic source, referrer decoded: campaign parameters first, then the referrer host for search, social, another site, this site, or direct. Free to read. This visit is classified from campaign parameters first, and otherwise from the referrer host: search, social, other, internal, or direct. After Accept, and only for the live preset, the rule id is queued. How to use it: - The three campaign parameters are read before the referrer. - The referrer host becomes search, social, other, internal, or direct. - The rows are shown and not stored. - After Accept, only the live preset queues the rule id. Spanish: [Fuente de tráfico, referrer descifrado](https://www.blackboxpersonalization.com/es/lab/source) - [Write a headline from a segment](https://www.blackboxpersonalization.com/lab/ai-copy): Write a headline from a segment and a role. A template you can read fills the lines. AI-assisted copy is optional. The free lab shows the inputs and the result. The page opens in manual mode and fills the headline from a template. AI-assisted copy runs only when the visitor opts in and OPENAI_API_KEY is set. A headline and a summary are filled from a template for the segment and the role. Generate can ask the server for a model draft, and that draft falls back to the same template. How to use it: - Change the segment and the role. The template fills without a model. - Generate is the only control that asks the server. - If the model is off, the same template is what you see. - After Accept, ai_copy_view carries the engine and the segment. Spanish: [Un titular a partir de un segmento](https://www.blackboxpersonalization.com/es/lab/ai-copy) - [People nearby, opted in](https://www.blackboxpersonalization.com/lab/nearby): Who is near you right now, only if they opted in. Locations snap to one-mile cells, expire in 30 minutes, and cells under three people stay off the map. Opt-in only. The browser snaps a point to a one-mile cell and the precise point is not sent. Rows expire after 30 minutes. Cells with fewer than three people are omitted from the map and from the radius count. Remove me deletes that browser's row. Global Privacy Control refuses the write. Apply drizzle/0015_nearby_presence.sql. A labeled sample draws when live cells are hidden. An opt-in point is snapped to a one-mile cell, and a cell is published only when at least three people share it. Rows expire after 30 minutes. How to use it: - Opt in before a point is snapped to a cell. - A cell with fewer than three people stays hidden. - Your own hidden count can still be shown to you. - A row is gone 30 minutes after it was written. Spanish: [Quién está cerca, con aviso](https://www.blackboxpersonalization.com/es/lab/nearby) - [A page shaped by place](https://www.blackboxpersonalization.com/lab/geo-page): The same page changes with a coarse place: city, local time, weather, and a nearby landmark. A panel names every signal that was used. Nothing is stored. The page reads Vercel IP geo headers and, only after a separate opt-in, a browser point snapped to a one-mile cell. Weather comes from Open-Meteo. Landmarks are a built-in teaching set. A postal code on the request is ignored. Nothing is stored. A labeled sample is marked as a sample. The page can be shaped from the request headers, a labeled sample, or a browser point snapped to a cell. Weather and landmarks fill in for that point, and this lab queues no event of its own. How to use it: - The coarse place comes from the edge headers. - Choose the sample, this browser, or the request. - Weather and landmarks are asked for that point. - A browser point is used only after you say yes. Spanish: [Página según el lugar](https://www.blackboxpersonalization.com/es/lab/geo-page) - [Radius versus drive time](https://www.blackboxpersonalization.com/lab/isochrones): Compare a straight-line radius with a drive-time shape for the same point. That gap is why delivery zones and local ads miss on a circle. Nothing is stored. Radius rings are drawn in the browser. Drive-time shapes are requested from a public Valhalla demo, then an OSRM table if that fails, and otherwise the page keeps the rings. The pin is rounded before the request. Nothing is stored. The pin is rounded to two decimals, then this page asks the public Valhalla demo for 10, 20, and 30 minute drive shapes. If that fails it estimates from OSRM, and the radius rings remain either way. How to use it: - Move the pin. It is rounded before any request. - Valhalla is asked for the three drive-time shapes. - If that fails, an OSRM time table estimates a ring. - The radius rings stay on the map either way. Spanish: [Radio o tiempo en coche](https://www.blackboxpersonalization.com/es/lab/isochrones) ## Agentic simulations - [The Agentic Journey Lab](https://www.blackboxpersonalization.com/lab/agentic-journey): The Agentic Journey Lab is a transparent simulation of sense, infer, decide, act, and learn. Rules and an in-browser bandit. No model calls. Four goals are scored from the signals this page is allowed to read, then an explore or exploit arm picks the next page. After Accept the outcome is queued, not the scores. How to use it: - Watch the four goal scores from the signals on the page. - The draw is explore or exploit, then one next page. - Consent off shortens the path. - After Accept, lab_interaction carries the outcome, not the scores. Spanish: [Laboratorio de recorrido agéntico](https://www.blackboxpersonalization.com/es/lab/agentic-journey) - [Friction the page can see](https://www.blackboxpersonalization.com/lab/friction-detector): See rage clicks, dead clicks, scroll thrash, and a stalled field on a demo widget. Rules intervene with help, a simpler version, or a shortcut. Rage clicks, dead clicks, scroll thrash, and a stalled field are rules on a demo widget. The log stays in this tab, and after Accept the page queues lab_interaction rather than the log. How to use it: - Three clicks in a short burst count as rage. - A click that hits nothing is a dead click. - Direction changes count as thrash. A focused field can stall. - After Accept, the lab event is queued, not the log lines. Spanish: [Fricción que la página ve](https://www.blackboxpersonalization.com/es/lab/friction-detector) - [A site agent answers JSON](https://www.blackboxpersonalization.com/lab/agent-handoff): A simulated visitor assistant asks this site for a guide. Keyword rules return schema.org JSON and one real page. No model call and nothing is stored. A question is matched to the first keyword list, and the result is a schema.org object for a real page on this site. After Accept the page queues lab_interaction with the rule id. How to use it: - Type a question. Matching is lowercase keywords, not a model. - No match comes back as empty fields. - A match attaches a schema.org object for one real page. - After Accept, the queued value is the rule id. Spanish: [Un agente responde en JSON](https://www.blackboxpersonalization.com/es/lab/agent-handoff) - [A path from a goal sentence](https://www.blackboxpersonalization.com/lab/concierge): Type a goal sentence. Keyword rules build a path of real pages on this site, with a reason for each stop. No model call and the sentence is not stored. A sentence is matched to the first keyword list, and that list's three real pages are laid out with a reason id. After Accept the page queues lab_interaction with the path id. How to use it: - Type a goal. The first matching word list wins. - The stops are real pages on this site. - Each stop shows the reason beside the link. - After Accept, the queued id is the path, not the sentence. Spanish: [Ruta desde una frase](https://www.blackboxpersonalization.com/es/lab/concierge) - [One action, then hold back](https://www.blackboxpersonalization.com/lab/next-best-action): One action for this session: a guide, a lab, a consult nudge, or nothing. A frequency cap and a rushed visit hold it back. No model is called. One guide, lab, or contact link is picked from labs already viewed. The offer is held when consent is off, the visit is rushed, or that action was shown twice. How to use it: - Consent off or a rushed visit holds the offer. - The pick is one action, from labs already viewed. - A second showing is the cap. The third does not appear. - After Accept, lab_interaction is queued for that action. Spanish: [Una acción y luego espera](https://www.blackboxpersonalization.com/es/lab/next-best-action) - [Blocks assembled per visitor](https://www.blackboxpersonalization.com/lab/content-assembly): The same four blocks, ordered differently for two labeled personas. The proof block links to a lab and does not cite a number. No model call. The same four blocks are shown in two preset orders, one for a marketer and one for an engineer. After Accept the page queues lab_interaction for the page view only. How to use it: - Switch the persona and watch the block order change. - The two orders are fixed. The blocks are the same. - After Accept, only the page view is queued. Spanish: [Bloques por visitante](https://www.blackboxpersonalization.com/es/lab/content-assembly) - [Adaptive tests by segment](https://www.blackboxpersonalization.com/lab/goal-seeking-test): Simulated traffic for two segments. Planted conversion rates are printed on the chart. Epsilon-greedy allocation. Not a measurement of this site. Two charts use planted conversion rates for new and returning visitors, always from the same seed. After Accept the page queues lab_interaction for the view. How to use it: - Each segment has two planted rates. - The draw explores, then exploits. - Every eighth arrival is recorded on the chart. - After Accept, only the view is queued. Spanish: [Tests adaptativos por segmento](https://www.blackboxpersonalization.com/es/lab/goal-seeking-test) - [Memory a journey may keep](https://www.blackboxpersonalization.com/lab/cross-channel-memory): A scripted path from a paid-social ad to this site, an email, and a return visit. Categories carry forward. Raw ids and the address are forgotten. A scripted ad, site, email, and return show which categories are carried and which ids are forgotten. Nothing here is a live profile, and after Accept the page queues lab_interaction. How to use it: - Step through the fixed ad, site, email, and return. - A category can carry forward while an id is dropped. - The story stays inside the script. It does not read this visit. - After Accept, a view is queued, then a demo click. Spanish: [Memoria que un viaje guarda](https://www.blackboxpersonalization.com/es/lab/cross-channel-memory) - [The agent checks its own plan](https://www.blackboxpersonalization.com/lab/self-critique): The agent proposes an action. Fairness, sensitive categories, a frequency cap, and consent can veto it. One veto holds the action. No model call. Four checks run on a proposal, and it is allowed only when fairness, sensitive categories, frequency, and consent all pass. After Accept the page queues lab_interaction. How to use it: - Read the four checks. All four have to pass. - Income and a precise place are a veto. - The consult nudge has a frequency cap. - After Accept, lab_interaction is queued for the result. Spanish: [El agente revisa su plan](https://www.blackboxpersonalization.com/es/lab/self-critique) ## AI agents & MCP - [What MCP is, in practice](https://www.blackboxpersonalization.com/lab/mcp): What MCP is in practice. A host opens one client per server. Tools, resources, and prompts move as JSON-RPC. This page cites revision 2026-07-28. You want to know what MCP is before you connect a server. It is a protocol between a host and a tool server, not a model. How to use it: - Open a section. Each one stays closed until you tap it. - Read the host, the client, and the server. - Copy a config only after you trust the server. - Open a lab from the list to try a trace. Spanish: [Qué es MCP en la práctica](https://www.blackboxpersonalization.com/es/lab/mcp) - [How an agent reads with MCP](https://www.blackboxpersonalization.com/lab/browser-mcp): How an agent reads this site with Browser MCP. Example calls stay on this origin, the question is not stored, and the answer links back to the page. You want an agent that can read a page and show the tool it called. The question should stay here, and the browser should stay on this site. How to use it: - Accept, then press Next through the example. - Watch the example tool calls, the snapshot, and the answer. - Open a source link. It is a page on this site. - Open How we built it for the code and the event map. Spanish: [Cómo un agente lee con MCP](https://www.blackboxpersonalization.com/es/lab/browser-mcp) - [Ask a report over MCP](https://www.blackboxpersonalization.com/lab/analytics-mcp): Ask which lab was viewed most by calling a report tool over MCP. The table is example data. A live GA4 property is not connected, and the question is not stored. You want an agent to answer which page was viewed most. The numbers on this page are an example, not a live property. How to use it: - Accept, then press Next through the report. - Read the dimensions, the metric, and the rows. - Notice the example label on the table. - Open How we built it for the event name. Spanish: [Pide un informe por MCP](https://www.blackboxpersonalization.com/es/lab/analytics-mcp) - [How an agent audits tags](https://www.blackboxpersonalization.com/lab/tag-audit-mcp): How an agent audits tags on this site. It reads pushes before and after Accept, stays on this origin, and does not submit a form. The trace is an example. You want to see whether a tag waited for Accept. The agent only opens a page on this site. How to use it: - Press Next from the open step to the answer. - Compare the before-Accept and after-Accept notes. - Open the dataLayer lab and the consent lab. - Check that the event has a tool name only. Spanish: [Cómo un agente audita etiquetas](https://www.blackboxpersonalization.com/es/lab/tag-audit-mcp) - [How a site audit works](https://www.blackboxpersonalization.com/lab/site-audit): How a site audit works. One public page in a sandboxed browser, with address checks and a rate limit. The scanner runs on Black Box Site Design. You want a page audit that cannot reach a private network. This lab names the sandbox, the checks, and the rate limit, then points at the scanner. How to use it: - Read the sandbox, the address checks, and the cap. - Open the Site Design audit. It may not be live yet. - This page does not fetch a URL. Spanish: [Cómo funciona una auditoría](https://www.blackboxpersonalization.com/es/lab/site-audit) - [Drafts for the next site fix](https://www.blackboxpersonalization.com/lab/self-improving): Drafts for the next site fix. Drop-off pages and not-found answers are the evidence. A pull request opens only when a flag is on, and a person approves it. You want the site to notice a weak page and a missing answer. The draft cites those counts, and a person still has to approve any change. How to use it: - Read whether the numbers are an example or from GA4. - Check the not-found count. The question is absent. - Leave pull requests off until the flag is set. Spanish: [Borradores del próximo arreglo](https://www.blackboxpersonalization.com/es/lab/self-improving) - [A read-only MCP server](https://www.blackboxpersonalization.com/lab/mcp-server): A read-only MCP server for this site. list_labs, get_lab, and search_blog return public fields. inspect_datalayer returns this browser's dataLayer session when that session id is passed. The question is not logged and writes are refused. You want this site to answer tools/list from a host you already use. The route is public, read-only, and rate-limited. How to use it: - Press Next to see tools/list and one call. - Copy the config into Cursor or Claude Desktop. - Restart the host and approve the tool. - Open What is MCP for the build steps. Spanish: [Servidor MCP de solo lectura](https://www.blackboxpersonalization.com/es/lab/mcp-server) - [Try a site MCP tool call](https://www.blackboxpersonalization.com/lab/mcp-playground): Try a site MCP tool call. Pick a tool from tools/list, fill the form from inputSchema, and read the raw JSON-RPC. Typed text is not logged. inspect_datalayer reads this browser's session. You want to call one tool yourself and read the JSON-RPC. Catalog tools return public fields from this site. How to use it: - Pick a tool. The form follows its schema. - Run it. Read the request and the response. - Accept if you want the tool name queued. - Leave the question text out of the event. Spanish: [Prueba una llamada MCP](https://www.blackboxpersonalization.com/es/lab/mcp-playground) - [How to build a site chatbot](https://www.blackboxpersonalization.com/lab/chatbot): How to build a site chatbot that answers from your own labs and notes, links the source, waits for consent, and does not store the question. You want an assistant that answers from pages you already published. It should show the source and not store the question. How to use it: - Accept measurement, then ask a starter or type a question. - Watch the stages, then open Track for the event list. - Read the excerpts and open a source link. - Open How we built it for the code and the event map. Spanish: [Cómo construir un chatbot](https://www.blackboxpersonalization.com/es/lab/chatbot) ## Advanced architecture - [Advanced architecture labs](https://www.blackboxpersonalization.com/lab/advanced): Sixteen architecture labs you can try in this browser. Tagging, consent, ranking, contracts, edge tests, agents, vitals, privacy, and proofs. These labs show an architecture choice in this browser, with example numbers marked as examples. After Accept, a lab may queue one view event and still does not call an outside account. How to use it: - Open one lab from the list. - Try the control and read any example label. - Accept only if you want the view event. - Open How we built it for the code. Spanish: [Arquitectura avanzada](https://www.blackboxpersonalization.com/es/lab/advanced) - [Server-side tagging](https://www.blackboxpersonalization.com/lab/advanced/server-side-tagging): Follow one event from this browser through a first-party /collect route. The route redacts, enriches, and builds a GA4 payload it does not forward. An event can leave the browser, land on your own server, and only then be shaped for GA4. This page posts to a route on this origin. The route redacts an email flag, adds the server time, and shows the payload it would forward. It does not call Google. A live forward needs a Measurement Protocol secret and a place to store the outcome. This route returns the payload and does not store it. You want the hit shaped on your server before any vendor sees it. After Accept, the page can queue a view event and the route still does not forward. How to use it: - Pick an event and mark whether an email was present. - Send it. Read the redacted list and the server time. - See forwarded false on the GA4 payload. Spanish: [Etiquetado en servidor](https://www.blackboxpersonalization.com/es/lab/advanced/server-side-tagging) - [Consent time machine](https://www.blackboxpersonalization.com/lab/advanced/consent-time-machine): Replay one scripted visit under Accept, Decline, and partial consent. Compare cookies, storage keys, and network hits side by side. The same visit does not have to mean the same hit. Pick Accept, Decline, or partial and step through one scripted visit. Each column shows the cookies, storage keys, and network hits that policy would allow. The visit is not this site's traffic. You want to see one visit under three consent choices before you pick a policy. The columns are a script, and Accept does not turn those hits into real calls. How to use it: - Choose Accept, Decline, or partial. - Step through the visit with the scrubber. - Compare cookies, storage, and network hits. Spanish: [Máquina del consentimiento](https://www.blackboxpersonalization.com/es/lab/advanced/consent-time-machine) - [On-device personalization](https://www.blackboxpersonalization.com/lab/advanced/on-device-personalization): A tiny ranking model in this browser scores the next step from the path you just clicked. The scores stay on the device. The next step can be scored without a server. Click through a short path. A small model in this page multiplies that history by fixed weights and shows the score for each suggestion. Nothing is uploaded. You want a next-step score that never leaves the browser. Click a path, read the scores, and know that Accept only queues a view event. How to use it: - Start on Collection, then click another page. - Read the score beside each suggestion. - Confirm the note that nothing is uploaded. Spanish: [Personalización local](https://www.blackboxpersonalization.com/es/lab/advanced/on-device-personalization) - [dataLayer contracts](https://www.blackboxpersonalization.com/lab/advanced/datalayer-contracts): Each event has a JSON contract. Edit the payload and the panel flags missing fields, wrong types, and forbidden keys in red. A dataLayer push is a contract, not a free-form bag. Pick an event, edit the JSON, and the page checks required fields, types, and keys that should never be on the hit. Violations show in red. The JSON stays in this tab. You want a bad payload caught before it becomes a hit. Edit the JSON, read the red rows, and leave the payload unpushed. How to use it: - Pick page_view, add_to_cart, or purchase. - Break a type or add an email key. - Read each red violation in the panel. Spanish: [Contratos del dataLayer](https://www.blackboxpersonalization.com/es/lab/advanced/datalayer-contracts) - [Edge A/B testing](https://www.blackboxpersonalization.com/lab/advanced/edge-ab-testing): See an edge bucket split, example conversion counts, and a Bayesian probability that one arm is ahead. The counts are not live traffic. Edge middleware can pin a visitor to a bucket before the page renders. This lab shows that split, then a Bayesian probability that one arm is ahead. Add example visits and conversions and the probability moves. The counts are not this site's traffic. Production would assign the bucket in edge middleware and store real impressions. This page only demonstrates the split and the math. You want the bucket and the odds visible before you trust a winner. The counts are examples, and adding a visit only changes this page. How to use it: - Read which bucket a key lands in. - Add an example visit or conversion. - Watch the probability that B is ahead. Spanish: [Test A/B en el edge](https://www.blackboxpersonalization.com/es/lab/advanced/edge-ab-testing) - [Agent API](https://www.blackboxpersonalization.com/lab/advanced/agent-api): Read agents.json and the journey tools on /api/mcp. Compare how an agent would walk the labs with how a person would. An agent needs a document and a tool, not a secret account. This page shows agents.json, calls the read-only journey tools on /api/mcp after Accept, and puts an agent walk beside a human walk. Agent visits is a channel name, not a traffic report. The tools answer with the public lab list. A real agent-visits channel in an analytics property is not connected. You want an agent to find the next lab without an account. Accept to call the read-only tools, and compare that walk with a click. How to use it: - Read agents.json on the page. - Accept, then press Call journey tools. - Compare the agent steps with the human steps. Spanish: [API para agentes](https://www.blackboxpersonalization.com/es/lab/advanced/agent-api) - [Vitals and behavior](https://www.blackboxpersonalization.com/lab/advanced/vitals-vs-behavior): Example Core Web Vitals joined to an example engagement rate. Filter the rows. They are not this site's field data. A slow page and a quiet page are different problems. This lab joins example LCP, INP, and CLS to an example engagement rate so you can see the two side by side. Filter the rows. They are not field data from this site. A live join would read CrUX or the site's own RUM plus consented engagement. Those sources are not queried here. You want speed and engagement on the same row before you blame the page. Every number is an example, so filter to one path and read both sides. How to use it: - Read LCP, INP, CLS, and engagement. - Filter the table by a path. - Keep the example label in view. Spanish: [Vitales y conducta](https://www.blackboxpersonalization.com/es/lab/advanced/vitals-vs-behavior) - [Live event graph](https://www.blackboxpersonalization.com/lab/advanced/live-event-graph): A D3 map of this visit. Pages are nodes. Events are edges. Add a step and the drawing updates in this tab. A visit is a graph, not a list of totals. Pages are nodes and events are edges. Add a page view, a scroll, or a click and D3 redraws this visit. The drawing stays in the tab. You want the current visit as a map instead of a counter. Add a page or an event and D3 redraws it in this tab. How to use it: - Start from this page as the first node. - Add a scroll, a click, or another page. - Read the event name on the edge. Spanish: [Grafo del recorrido](https://www.blackboxpersonalization.com/es/lab/advanced/live-event-graph) - [Differential privacy](https://www.blackboxpersonalization.com/lab/advanced/differential-privacy): Move an epsilon slider and watch Laplace noise change an example count. The privacy and accuracy tradeoff is drawn on the page. A published count can hide whether one person was in it. Move epsilon. Laplace noise is added to an example count, and the page shows how far the noisy answers drift. Lower epsilon is more private and less accurate. A production query would track a privacy budget across every release. This slider spends nothing outside the tab. You want to see how much noise a private count needs. Move epsilon, compare the example count with the noisy one, and publish nothing. How to use it: - Read the example count of 120. - Drag epsilon from private toward accurate. - Read the noisy count and the average error. Spanish: [Privacidad diferencial](https://www.blackboxpersonalization.com/es/lab/advanced/differential-privacy) - [Federated learning](https://www.blackboxpersonalization.com/lab/advanced/federated-learning): Simulated browsers take one local step. The server averages the weights. A chart shows the loss falling. No device calls this site. Several browsers can train on their own rows and send only a weight update. Press Train. Three simulated browsers take a local step, the page averages the weights, and the loss chart updates. No browser outside this tab takes part. Real federated learning needs devices, a secure average, and a limit on what an update can reveal. This chart is a local simulation. You want the average of local updates without pooling the rows. Press Train and watch the loss fall on these simulated browsers. How to use it: - Read the three simulated browsers. - Press Train once and read the new weight. - Follow the loss line across rounds. Spanish: [Aprendizaje federado](https://www.blackboxpersonalization.com/es/lab/advanced/federated-learning) - [Causal attribution](https://www.blackboxpersonalization.com/lab/advanced/causal-attribution): A Markov removal model on example journeys. Remove a page and watch the credit and the conversion probability move. Last click gives the whole sale to the last page. This lab uses a Markov chain on example journeys. Remove a page and the conversion probability drops by that page's share. The journeys are not stored visits. A live model would read consented paths from storage. These journeys are written into the page. You want credit that changes when a page is removed, not a last-click guess. Pick a page and read the new shares on these example journeys. How to use it: - Read the baseline conversion probability. - Remove home, lab, consent, or packages. - Watch the other pages' credit move. Spanish: [Atribución causal](https://www.blackboxpersonalization.com/es/lab/advanced/causal-attribution) - [Predictive intent](https://www.blackboxpersonalization.com/lab/advanced/predictive-intent): Scroll, dwell, and path depth on this page become a leave score and a packages score. The note on the page changes with them. A page can react before a click. Scroll this lab and wait. Two scores update from scroll depth, dwell, and how many steps you add. The note changes when one score leads. The math stays in the tab. The weights are fixed in the page. A trained model and a decision service are not called. You want the page to react to this visit before a click. Scroll, wait, or add a step, and the note follows the higher score. How to use it: - Scroll the page and watch both scores. - Add a step to raise the path depth. - Read the note that the page is showing. Spanish: [Intención prevista](https://www.blackboxpersonalization.com/es/lab/advanced/predictive-intent) - [Agent to agent](https://www.blackboxpersonalization.com/lab/advanced/agent-to-agent): A visitor agent and the site agent exchange next-step messages. The rules are on the page. No hosted model is called. Two agents can negotiate the next step without a chat product. Choose a goal. A visitor agent asks, and the site agent answers with a lab on this site and a consent check. The transcript is the exchange. No hosted model is called. The turns are rules in this page. A hosted visitor agent and a hosted site agent are not connected. You want the negotiation written down, not hidden in a chat product. Pick a goal and read each turn, and no hosted model is called. How to use it: - Pick measure, privacy, or packages. - Press Negotiate and read the transcript. - Open the lab the site agent named. Spanish: [Agente con agente](https://www.blackboxpersonalization.com/es/lab/advanced/agent-to-agent) - [Verifiable analytics](https://www.blackboxpersonalization.com/lab/advanced/verifiable-analytics): After Accept, events you add are hashed into a Merkle tree. Check a leaf against the root with an inclusion proof. A consented event can be committed without publishing the event itself. Add an event after Accept. The page hashes it into a Merkle tree, shows the root, and checks an inclusion proof. Decline and the tree does not take the event. The root lives in this tab. A public transparency log that other people can read is not published. You want a root that commits to events without listing them in public. Accept, add an event, and check the proof before you trust the root. How to use it: - Accept, then add one of the named events. - Read the root hash. - Check the inclusion proof for a leaf. Spanish: [Analítica verificable](https://www.blackboxpersonalization.com/es/lab/advanced/verifiable-analytics) - [Digital twin](https://www.blackboxpersonalization.com/lab/advanced/digital-twin): Scripted personas walk journey rules. Compare the predicted funnel before and after one rule change. The people are not visitors. You can rehearse a rule on simulated people before you ship it. Four personas walk a short path. Turn on one rule and the predicted funnel changes. The people are written on the page. They are not visitors. A production twin would sample consented behavior. These four personas are fixed. You want to see a rule change on simulated people before you ship it. Turn the offer on and compare the two funnels, because these people are not visitors. How to use it: - Read the four personas and who converts. - Turn on the offer after the lab. - Compare the done count before and after. Spanish: [Gemelo digital](https://www.blackboxpersonalization.com/es/lab/advanced/digital-twin) - [Tracking time travel](https://www.blackboxpersonalization.com/lab/advanced/tracking-time-travel): A dataLayer and storage timeline for this tab, with a scrubber. Move it and read the state at that step. Debugging a tag means seeing the state at one moment, not the latest push only. Add events, then drag the scrubber. The dataLayer list and the storage keys rewind to that step. The timeline is this tab. You want the dataLayer at an earlier step, not only the latest push. Add events and drag the scrubber so the storage keys rewind with it. How to use it: - Add a page view, a click, or consent. - Drag the scrubber to an earlier step. - Read the dataLayer and the storage keys. Spanish: [Viaje en el rastreo](https://www.blackboxpersonalization.com/es/lab/advanced/tracking-time-travel) ## Phone & sensors - [IP geolocation accuracy](https://www.blackboxpersonalization.com/lab/location): City-level IP location from the request, an optional precise point, and a follow view with Standard or Prefer high accuracy. Follow saves no trail. The map draws the city from the request even when the database refuses the write. A precise point is stored only after Share precise location. Follow this browser starts only after Start location, offers Standard or Prefer high accuracy, stops when the page hides or closes, and does not save a movement history. A labeled simulated example runs when permission is denied or unavailable. A location_sample beacon is posted to this site, and the page draws the browser, the edge, the function, and the database from the stored row. A precise point is optional. How to use it: - The beacon is first-party. It posts, then can fall back to sendBeacon. - The stored row is read back onto the four hops. - Share precise location is the only way a precise point starts. - Decline leaves the precise point unread. Spanish: [Precisión de la geolocalización por IP](https://www.blackboxpersonalization.com/es/lab/location) - [Compass, heading, and a waypoint](https://www.blackboxpersonalization.com/lab/compass): Compass and a waypoint: orientation and location stay in this browser, with one point and a straight-line distance. Not a hiking navigator app. Heading, a waypoint, and a straight-line distance stay in the browser. The page is an educational aid. It is not a hiking navigator, and it does not convert a magnetic heading to true north. Two coordinates become a straight-line distance and an initial bearing. Live mode can read this device when you press Start, and a simulated walk is labeled as not your phone. How to use it: - Set a waypoint and read the distance and the initial bearing. - Start asks before any live watch. Stop ends it. - Play simulated example runs a scripted walk, not this device. - Clear removes the waypoint and the trail from this page. Spanish: [Brújula y un punto](https://www.blackboxpersonalization.com/es/lab/compass) - [Tilt a level with the phone](https://www.blackboxpersonalization.com/lab/tilt): Tilt a level with the phone after Start, or with the keyboard, a drag, or a labeled simulated example when the sensor is missing or denied. Nothing is stored. The level uses device orientation after a click when the browser offers it. Keyboard, touch, and a labeled simulated example work without a sensor. Stopping or leaving the page ends the reading. A ball moves from this device's orientation, or from the keyboard, a drag, or a scripted example. Permission is asked only after Start, and nothing is saved. How to use it: - Start is what asks for device orientation. - The latest beta and gamma move the ball, clamped inside the well. - Arrow keys, a drag, or the simulated example are labeled as not hardware. - Stop ends the listeners. Nothing is saved. Spanish: [Inclina el nivel](https://www.blackboxpersonalization.com/es/lab/tilt) - [Scan a QR code on this phone](https://www.blackboxpersonalization.com/lab/qr-scanner): Scan a QR code in this browser. The camera starts only after a click, nothing is uploaded, and the sample code only says hello. The demo still works. The camera starts only after a click, prefers the rear camera, and decodes in the browser. The sample code says hello and is not a link. Nothing is uploaded. Camera frames are drawn to a canvas and read with BarcodeDetector, or with jsqr when that API is missing. The text stays on the page, and a code is not opened or uploaded. How to use it: - Start camera asks for video only, and prefers the rear camera. - Stop, hiding the page, or leaving turns the camera off. - A denied camera still lets you decode a picture you choose. - The decoded text is shown here. This page does not open it. Spanish: [Escanear un código QR](https://www.blackboxpersonalization.com/es/lab/qr-scanner) - [What your IP gives away](https://www.blackboxpersonalization.com/lab/ip-vs-gps): Set the network's city beside a browser point you choose to share. The map draws the gap and the distance. The point stays in this tab and is not stored. IP headers are read from the request and not stored. A browser point is compared in the tab and is not sent. The gap is why a precise point waits for consent. The city and coordinates already on the request are rounded to two decimals, and miles are measured to a browser point that stays in this tab. Neither point is stored. How to use it: - The network point is the edge headers, rounded to two decimals. - Ask for a browser point. It stays in this tab. - The gap between the two points is measured here. - Neither point is written to storage. Spanish: [Qué dice la IP y el GPS](https://www.blackboxpersonalization.com/es/lab/ip-vs-gps) - [Draw a zone, drag a pin](https://www.blackboxpersonalization.com/lab/geofence): Draw a zone on the map and drag a test pin through it. Enter, exit, and dwell events change the content in this browser only. Nothing is sent or stored. The zone, the pin, and the enter, exit, and dwell events stay in the browser. Nothing is sent or stored. A ring and a draggable pin show enter, exit, and dwell, written only in this tab. Nothing is posted, and the lab does not queue an event of its own. How to use it: - Drag the pin and watch the ring test. - Enter, exit, and dwell are written in this tab. - The test does not call the network. Spanish: [Dibuja una zona de prueba](https://www.blackboxpersonalization.com/es/lab/geofence) - [How apps and speakers profile](https://www.blackboxpersonalization.com/lab/apps-and-voice): A simulated phone, a local microphone meter, and official links for apps, speakers, and privacy law. Not legal advice. No real advertising id is read. A simulated tracking choice swaps a scripted SDK list, and a channel class is carried across app, web, speaker, and TV. Microphone samples stay in this browser, and after Accept the page queues lab_interaction. How to use it: - Change the tracking choice and watch the SDK rows swap. - Three facts carry forward. Three ids are forgotten. - A third tap is marked as rage. - Microphone samples stay in this browser. Spanish: [Apps y altavoces](https://www.blackboxpersonalization.com/es/lab/apps-and-voice) ## Ads & attribution - [Custom campaign link builder](https://www.blackboxpersonalization.com/lab/utm-builder): A custom campaign link (UTM) builder. Name the source, medium, and campaign, then see how this site and GA4 read them. Copy the URL when the fields pass. A path and three required slugs become utm_source, utm_medium, and utm_campaign on a URL that stays on this origin. After Accept the view is queued with no parameters. How to use it: - A value that is not a slug is rejected. - The path stays on this site. - The three required parameters are set, then the two optional ones. - After Accept, utm_builder_view is queued with no fields. Spanish: [Constructor de enlaces de campaña](https://www.blackboxpersonalization.com/es/lab/utm-builder) - [Paid and social measurement](https://www.blackboxpersonalization.com/lab/ad-measurement): Paid and social measurement for every ad platform: one taxonomy, UTM rules, aligned tags, and a free dry run that sends nothing and stores nothing. The UTM builder, click id detector, tag fan-out, and landing preview run in the browser. Nothing is sent to an ad platform and nothing is stored. The page works with no database. One canonical event becomes a browser snippet and a server body for each platform, and the first matching campaign rule picks a landing variant. Those previews stay on the page. How to use it: - The event is named once and shown per platform. - A card is withheld when that platform's consent is off. - An email is hashed in this browser before it appears. - The landing variant stays hidden until Accept. Spanish: [Medición de paid y social](https://www.blackboxpersonalization.com/es/lab/ad-measurement) - [Meta Conversions API explained](https://www.blackboxpersonalization.com/lab/capi): Meta Conversions API explained: a server event beside the Pixel, SHA-256 email and phone, a shared event id, and a free dry run that sends nothing to Meta. The dry run hashes email and phone in the browser, shows the JSON, and sends nothing to Meta. Raw email and phone are not stored. The page works with no database. An email and a phone are normalized and hashed in this browser, then shown as a sample server payload and a pixel line. Both stay on the page, and user data is omitted when consent is off. How to use it: - Watch the email and phone get normalized before the hash. - Consent off omits the user data from the sample. - The match keys are the documentation names. - The payload and the pixel line stay on this page. Spanish: [La API de conversiones de Meta](https://www.blackboxpersonalization.com/es/lab/capi) - [LinkedIn attribution, compared](https://www.blackboxpersonalization.com/lab/attribution): Attribution models compared on one journey, including a LinkedIn-style path. Last click, first click, linear, time decay, and position, side by side. One journey is scored six ways: last click, first click, linear, time decay, position, and Markov removal. Short stored paths fall back to four invented journeys, and after Accept the data source is queued, not the shares. How to use it: - Read one journey as a list of touches. - Five rules score in one pass. Markov removes a channel. - Stored paths that are too short switch to the labeled sample. - After Accept, attribution_view carries the source, not the shares. Spanish: [Modelos de atribución, explicados](https://www.blackboxpersonalization.com/es/lab/attribution) - [Offline conversions, explained](https://www.blackboxpersonalization.com/lab/offline-conversions): Offline conversions, explained: tie a demo lead to a closed outcome and send it with the Measurement Protocol when the secret is set. Free to walk through. A demo lead can be closed through the Measurement Protocol when that server secret is set, and otherwise the page shows the body without calling Google. After Accept it queues the transport, not the lead. How to use it: - Record a demo lead. The id is made on this page. - A send is refused while consent is off. - Google is called only when the server secret is set. - After Accept, offline_conversion_view carries the transport. Spanish: [Conversiones offline, explicadas](https://www.blackboxpersonalization.com/es/lab/offline-conversions) ## Search, SEO & AI visibility - [How search engines see this](https://www.blackboxpersonalization.com/lab/seo): How search engines see this page: title, canonical, structured data, and the HTML before JavaScript, for any URL on this site. Free to fetch and compare. HTML is fetched for a URL on this site, and the title, canonical, robots, and Open Graph tags are read from that response. After Accept the path and the warn count are queued. How to use it: - Pick a URL on this site. The raw document is fetched. - Title, canonical, robots, and Open Graph are read from the head. - Raw text and rendered text are compared. - After Accept, seo_inspect carries the path and the warn count. Spanish: [Cómo ven esta página los buscadores](https://www.blackboxpersonalization.com/es/lab/seo) - [JSON-LD this site publishes](https://www.blackboxpersonalization.com/lab/json-ld): JSON-LD this site publishes: Person, Organization, and WebSite on every page, a TechArticle on each lab, and breadcrumbs on nested pages. Free to inspect. Organization and WebSite are on every page. The organization names Black Box Site Design as parentOrganization. A lab adds a LearningResource. A note adds a BlogPosting. The about page adds a Person. The labs index is a CollectionPage. The organization sameAs stays empty. Pick a path and this page builds the site graph and the page graph, then checks context, types, dates, and sameAs. After Accept it queues the path, the script count, and the warn count. How to use it: - Pick a path. The scripts are built for that path. - A lab or a note is expected to carry an article. - Context, types, dates, and sameAs are checked on the page. - After Accept, jsonld_view carries the path and the two counts. Spanish: [El JSON-LD que publica este sitio](https://www.blackboxpersonalization.com/es/lab/json-ld) - [Google Search Console data, live](https://www.blackboxpersonalization.com/lab/search-console): Google Search Console data, live: clicks, impressions, click-through rate, average position, top queries, and top pages, once access to the property is set. This page asks Google for clicks, impressions, queries, and pages over the report window, then caches that report. It does not queue an analytics event of its own, and it stops when the account is not connected. How to use it: - If the account is not connected, the page stops. - The report is four slices: clicks, impressions, queries, and pages. - The window is summed, then cached, including a failure. - This lab does not queue its own analytics event. Spanish: [Datos en vivo de Search Console](https://www.blackboxpersonalization.com/es/lab/search-console) - [Google Trends and its limits](https://www.blackboxpersonalization.com/lab/trends): What Google Trends measures: a sampled index scaled 0 to 100 for the request, not a search count. Official alpha, pytrends, and SerpApi, with a live chart. The live chart calls the Trends website endpoint, or SerpApi when that key is set. If the call fails, the page says Trends is unreachable. The labeled example is separate and is not a measurement. Load interest asks this server for a relative interest series, trying a third-party engine when a key is set and otherwise the website endpoint. After Accept the data source is queued, not the index. How to use it: - Load interest calls this site's own route. - A configured key tries the third-party engine first. - Otherwise the website endpoint is used. A short cache can answer. - The latest point is compared with a short mean. The index is not sent. Spanish: [Google Trends y sus límites](https://www.blackboxpersonalization.com/es/lab/trends) - [Trends beside Search Console](https://www.blackboxpersonalization.com/lab/trends-search-console): Pair Google Trends interest with Search Console clicks, impressions, CTR, and position. A gap is rising interest with no strong rank in the fetched rows. Search Console rows are read only when the service account is set. Otherwise the join is a labeled example. A query missing from the returned rows is not proof of zero impressions. A Trends spike is joined to the top query rows for this property, and a gap is labeled when interest is up and the query is missing or ranked worse than 10. After Accept the source word is queued, not the queries. How to use it: - If the property is not set, Compare stops. - Compare asks for the top queries. - A gap needs the spike plus a missing or poorly ranked query. - After Accept, the event carries the source, not the query text. Spanish: [Trends y Search Console](https://www.blackboxpersonalization.com/es/lab/trends-search-console) - [Sitemap changes and submission](https://www.blackboxpersonalization.com/lab/sitemap): Sitemap changes and submission: a timeline of URL and lastmod edits, and how this site tells Search Console and IndexNow. Free to read the diff. When the database is offline, the diff uses the snapshot bundled with the deploy and says history is unavailable. The page still renders. The public URL list is diffed into added, removed, and lastmod changes, and an admin submit can tell Search Console and IndexNow about that diff. After Accept the counts are queued, not the URL list. How to use it: - Read added, removed, and lastmod from the two lists. - The page says which URLs IndexNow would hear. - A fetched sitemap is checked against the list. - Submit is a dry run unless the session is an admin. Spanish: [Cambios del sitemap y su envío](https://www.blackboxpersonalization.com/es/lab/sitemap) - [Get your site found in search](https://www.blackboxpersonalization.com/lab/get-found): Get your site found in search with indexing, discovery, and measurement, using this site as the worked example. A free checklist of steps already served. The checklist marks indexing steps that are already done in the repo and two the owner still has to do. After Accept only the done count is queued, not the checklist text. How to use it: - Read each row as done or still the owner's. - The evidence for a done row is on the page. - After Accept, done_count is queued. The text is not. Spanish: [Haz que busquen tu sitio](https://www.blackboxpersonalization.com/es/lab/get-found) - [What an llms.txt file is](https://www.blackboxpersonalization.com/lab/llms-txt): What an llms.txt file is: an emerging convention, not a standard. This site builds /llms.txt and /llms-full.txt at build from its pages, labs, and notes. The files at /llms.txt and /llms-full.txt are generated at build from the page catalog, the lab groups, and the notes. robots.txt comments name both files, and the sitemap lists them. The file you see is built from the public page catalog, the lab groups, and the notes, and each line is labeled. After Accept the queued number is how many fetches were logged, not the file body. How to use it: - Read the file. Each line says where it came from. - A sitemap draft is offered, and it is not the served file. - After Accept, llms_txt_view carries the logged fetch count. Spanish: [Qué es un archivo llms.txt](https://www.blackboxpersonalization.com/es/lab/llms-txt) - [Measuring AI referrals](https://www.blackboxpersonalization.com/lab/ai-referrals): Measuring AI referrals, crawler user agents, and a citation log you fill in by hand. A free look at how this site tells those three sources apart. A referrer host or utm_source is checked against a fixed assistant list, and stored page views that match are counted. After Accept, ai_referral_detected is queued only when an assistant matches. How to use it: - utm_source is checked before the referrer host. - A match names one assistant from the fixed list. - Stored page views that match are counted on the page. - After Accept, the queued id is the assistant, and only on a match. Spanish: [Medir las referencias de IA](https://www.blackboxpersonalization.com/es/lab/ai-referrals) - [AI referral clicks and crawlers](https://www.blackboxpersonalization.com/lab/ai-referral-traffic): AI referral traffic: referrer hosts, campaign source, and crawler agents, with example charts. Not this site's traffic. This tab reads only this browser. A teaching page for AI referral clicks versus crawler user agents. The charts are an invented 14-day log. The this-tab box reads the current referrer, campaign source, and User-Agent and does not store them. The CDN Logs lab is the other mockup for example edge rows. The charts are a seeded example of clicks and crawlers, not this site's traffic. The this-tab box classifies this browser's referrer, utm_source, and User-Agent, and it does not send them. How to use it: - The charts share the seeded clicks by day. - Filter one source and watch that series. - This tab separates a click from a crawler User-Agent. - The example log is not this site's requests. Spanish: [Tráfico de referidos IA](https://www.blackboxpersonalization.com/es/lab/ai-referral-traffic) - [Ask analytics in plain English](https://www.blackboxpersonalization.com/lab/ask): Ask analytics in plain English. Preset questions become SELECT statements over the log. AI-assisted matching is optional, and the SQL stays on the page. The page opens in manual mode. AI-assisted matching runs only when the visitor opts in and OPENAI_API_KEY is set. If that call fails, the keyword rules still run. A question is scored against four SQL templates, or you can run a preset, and the rows are shown. After Accept the queued fields are the data source and the engine, not the question text. How to use it: - Ask a question or run a preset. - The winner is the template with the highest keyword score. - The rows are the result of that template. - After Accept, ask_view carries the source and the engine. Spanish: [Preguntar analítica en castellano](https://www.blackboxpersonalization.com/es/lab/ask) ## Reporting & charts - [Funnel drop-off analysis](https://www.blackboxpersonalization.com/lab/funnels): Funnel drop-off analysis across visit, lab page, white paper, and lead for the last 28 days, with device and source filters. Counts come from stored events. Four steps over 28 days count any page, a lab page, the white paper, then a stored lead. The first three counts are distinct visitors, and after Accept the data source is queued, not the counts. How to use it: - The window is the last 28 days. - The first three steps count visitors, and they can filter device and source. - The lead step shows only when device and source are All. - Sample bars are labeled. After Accept, the source is queued, not the counts. Spanish: [Análisis de caída en el embudo](https://www.blackboxpersonalization.com/es/lab/funnels) - [Cohort retention analysis](https://www.blackboxpersonalization.com/lab/cohorts): Cohort retention analysis by Monday-start week. Each cell is that group share with a later page view. A free table built from stored first visits. Page views are grouped by the Monday of each visitor's first week, and later weeks count how many of those visitors came back. After Accept the data source is queued, and a thin grid is labeled. How to use it: - Only page views are kept for the grid. - Each visitor is bucketed by that first Monday. - A later column is how many of that cohort returned. - A thin grid is labeled. After Accept, the source is queued. Spanish: [Análisis de retención por cohorte](https://www.blackboxpersonalization.com/es/lab/cohorts) - [Traffic spikes and dips](https://www.blackboxpersonalization.com/lab/anomalies): Traffic spikes and dips against a rolling mean and a two-standard-deviation band. A day is flagged when the count sits outside the prior seven days. A day is a spike or a dip when its count sits outside two standard deviations of the previous seven days. After Accept the data source is queued, not the counts. How to use it: - Missing days are filled with zero before the test. - The band is the seven days before the day you are looking at. - A labeled sample is there when stored rows are thin. - After Accept, anomaly_view carries the source, not the counts. Spanish: [Picos y caídas de tráfico](https://www.blackboxpersonalization.com/es/lab/anomalies) - [A/B test significance calculator](https://www.blackboxpersonalization.com/lab/significance): A/B test significance calculator for p-value, confidence interval, probability to beat, and sample size, from the experiments lab counts. Free to run. Two arm counts become a two-proportion p-value, an unpooled interval, a probability that B is ahead, and a sample-size estimate. After Accept the data source is queued, not the p-value. How to use it: - Use the stored arms, or switch to the labeled sample. - The test compares the two rates. - Beta draws estimate the chance that B is ahead. - The sample-size line is impressions per arm. The p-value stays here. Spanish: [Calculadora de significancia A/B](https://www.blackboxpersonalization.com/es/lab/significance) - [Analytics data quality checks](https://www.blackboxpersonalization.com/lab/data-quality): Analytics data quality checks on stored page views, web vitals, and form submits. Each check is a field this site expects. Visitor ids stay off the list. Stored beacons are checked for a path, a device class, a referrer class, a web vital, and a form name. The counts stay on the page, and after Accept only the source and the fail count are queued. How to use it: - Hits are split by event name before the checks run. - A missing required value is a failure. - Eight failure types stay on the page, plus a labeled sample. - After Accept, the event carries the source and the fail count. Spanish: [Controles de calidad de la analítica](https://www.blackboxpersonalization.com/es/lab/data-quality) - [Example path tree in the D3 lab](https://www.blackboxpersonalization.com/lab/d3): D3 lab: an example tidy tree of visits, from the entry source through the path to a page and the rows behind a metric. Not this site's traffic. An example tidy tree of invented sessions. It drills from the entry source to the path, then one page, then the rows behind a metric. Color is LCP. Not this site's traffic. No device reading. The tree is a few thousand invented sessions across 40 pages, from a fixed seed. It does not read this browser, and it does not send those rows to Google or Adobe. How to use it: - Link thickness is sessions. Color is LCP. - Open a source, then a page, from the branch you tap. - Each metric has its own columns. - The example stays on the page. It is not this site's traffic. Spanish: [Laboratorio D3](https://www.blackboxpersonalization.com/es/lab/d3) - [Visualization workbench](https://www.blackboxpersonalization.com/lab/visualization): Visualization workbench: an example table, calculated fields, and a channel, page, and event drill-down. The rows are invented, not real traffic. An editable example table calculates share and pace, and a drill goes from channel to page to event. The rows are not real visitors. Heatmap, map, and flow views are named as not built. An in-memory example table groups by channel, then page, then event, and a timer can count visibility in component state. The rows are invented, and this section sends no event of its own. How to use it: - Switch Example visitors and Example walk. Both are invented. - Drill channel, then page, then event, without leaving the page. - Reset example rows puts the planted table back. - Demo metrics count only while that control is on. Spanish: [Banco de visualización](https://www.blackboxpersonalization.com/es/lab/visualization) ## Performance & infrastructure - [Core Web Vitals, measured](https://www.blackboxpersonalization.com/lab/performance): Core Web Vitals, measured for this visit beside the 75th percentile of stored web vital events. LCP, INP, CLS, and TTFB, free to read on the page. LCP, INP, CLS, and TTFB for this visit sit beside the nearest-rank 75th percentile of the latest stored web vital rows. After Accept the page queues performance_view with no parameters. How to use it: - TTFB is read from this navigation. - LCP, CLS, and INP are watched on this visit. - This visit is rated against the usual thresholds. - The stored percentile loads beside it. The view has no fields. Spanish: [Core Web Vitals, medidos](https://www.blackboxpersonalization.com/es/lab/performance) - [Free page speed test lab](https://www.blackboxpersonalization.com/lab/pagespeed): Free page speed test lab for a page on this site: Lighthouse scores, field data when Chrome has it, and a history of stored scores. Run it in the browser. This server runs Lighthouse for a public path on this site, reuses a result for 15 minutes, and reads the category scores from the response. After Accept a finished run queues pagespeed_test. How to use it: - Pick a path on this site. Other hosts are not run. - A result from the last 15 minutes is reused. - New calls are limited. Scores and field data come from the response. - After Accept, a finished run queues pagespeed_test. Spanish: [Laboratorio gratuito de velocidad](https://www.blackboxpersonalization.com/es/lab/pagespeed) - [CDN logs and cache status](https://www.blackboxpersonalization.com/lab/cdn-logs): CDN logs, an example: edge and origin, cache HIT, MISS, and STALE, status, TTFB, bytes, and region. The rows are invented. Not this site's traffic. An invented 24-hour edge log. HIT and STALE are answered at the edge. MISS means the edge asked the origin. The page does not read a Vercel log drain. Filters run over a seeded example edge log and summarize hit ratio, edge ratio, median time to first byte, errors, and bots. The rows are an example, not this site's requests. How to use it: - The log is one planted day of example rows. - Change a filter. The summary uses only the rows that remain. - Hit ratio, edge ratio, median TTFB, errors, and bots update together. - Nothing here opens a log drain or this site's traffic. Spanish: [Registros de CDN](https://www.blackboxpersonalization.com/es/lab/cdn-logs) - [What an API request is](https://www.blackboxpersonalization.com/lab/api): What an API request is: one program asking another for data, or asking it to do something. This page calls Wikidata and shows both the request and the response. Country, city, and name filter a Wikidata query for data centers, and the request sits beside the rows. This page does not call Adobe. How to use it: - Type a country, a city, or a name. The filters stay plain text. - The page builds one SPARQL select from those three. - The call goes to query.wikidata.org. - A saved response can be reused instead of calling again. Spanish: [Cómo funciona una petición API](https://www.blackboxpersonalization.com/es/lab/api) - [Data APIs this site calls](https://www.blackboxpersonalization.com/lab/data-apis): Data APIs this site calls, grouped: Trends, Search Console, PageSpeed, Wikidata, Open-Meteo, and sitemap submission. Each card opens that lab. A list of the labs that call an outside data API, and which credential each one needs. This page lists the labs that read an outside service and links to each one. It does not call those services, and it does not queue an event of its own. How to use it: - Open a card to go to that lab. - The card names the service, not a secret. - Nothing on this page calls Trends, Search Console, or the others. - llms.txt is described here and served from its own lab. Spanish: [APIs de datos del sitio](https://www.blackboxpersonalization.com/es/lab/data-apis) - [Embed a browser test card](https://www.blackboxpersonalization.com/lab/embed): Embed a browser test card on another site. It reads this browser and links back to the full lab report. The frame itself is not indexed. Free to copy. Two controls fill a script tag and an iframe for the browser-test card. The card reads this frame and posts the host so this site can store the host and the time, without its own analytics event. How to use it: - Change the two controls and watch the snippet fill in. - The card reads this frame, not a profile. - The host and the time can be stored. The hint stays on the card. - This lab does not queue an analytics event of its own. Spanish: [Insertar una tarjeta de prueba del navegador](https://www.blackboxpersonalization.com/es/lab/embed) ## Pages - [Public experiment ledger](https://www.blackboxpersonalization.com/experiments): A public ledger of experiments on this site: the hypothesis, the split, a Bayesian read, the stopping rule, and a Merkle root anyone can check. Spanish: [Libro público de pruebas](https://www.blackboxpersonalization.com/es/experiments) - [Privacy policy for this site](https://www.blackboxpersonalization.com/privacy): What this site keeps: analytics cookies only after you accept, chatbot questions out of analytics, path history on this device, and read-only Google access. Spanish: [Política de privacidad](https://www.blackboxpersonalization.com/es/privacy) - [What changed on this site](https://www.blackboxpersonalization.com/changelog): What changed on this site. Dated notes for the pages that shipped here. A suggestion stays a draft until a person approves the pull request. Spanish: [Qué cambió en este sitio](https://www.blackboxpersonalization.com/es/changelog) - [What your device gives away](https://www.blackboxpersonalization.com/showcase): What your device gives away without asking: hardware, settings, network, and in-tab behavior this browser already exposes, and how to limit each one. Spanish: [Lo que revela tu dispositivo](https://www.blackboxpersonalization.com/es/showcase) - [Measurement package options](https://www.blackboxpersonalization.com/packages): One destination for the lab next-step test. Nothing is sold on this page. The test stays off until it is approved, and a visit without consent is not counted. Spanish: [Paquetes de medición web](https://www.blackboxpersonalization.com/es/packages) ## Notes ## From a push to a Google hit - [From a push to a Google hit](https://www.blackboxpersonalization.com/blog/from-a-push-to-a-google-hit): The same page view can go to two places. Here is how a line pushed onto window.dataLayer becomes a request to google-analytics.com, and what that request carries. [What a beacon actually sends](https://www.blackboxpersonalization.com/blog/what-a-beacon-actually-sends) follows that request to this site's own API. That is the first-party path: a body I wrote, sent to a route I run, enriched by a server I control. Most sites have a second path, and so does this one. The same page view can also go to Google Analytics. That path starts with an ordinary JavaScript array and ends with a request to a host this site doesn't own. The [dataLayer lab](https://www.blackboxpersonalization.com/lab/datalayer) shows each step live. Open it in another tab and follow along. ## The dataLayer is a queue, not a report `window.dataLayer` is just an array on the page. Pushing onto it sends nothing. It's a to-do list that Google Tag Manager reads later. The `gtag()` function is still only a few lines, and it runs before any Google code arrives. This site keeps it as a consent shim. Consent Mode reads an arguments object, so a plain array would be ignored: ```js window.dataLayer = window.dataLayer || []; function gtag() { dataLayer.push(arguments); } ``` A page view does not call that function. The site pushes an object. That appends the object to the array and returns. It does not call Google: ```js dataLayer.push({ event: "virtual_page_view", page_path: "/" }) ``` ## The container reads the queue The real work starts when `gtm.js` loads from `googletagmanager.com`, for container `GTM-NCP2JHJR`. This site does not load `gtag.js`. Google Analytics 4, measurement id `G-JDE6BF33RY`, is a tag inside that container. The container replaces `dataLayer.push` with its own function, then works through whatever is already waiting, in order. Each command means something different: - `consent` changes the gate. The shim pushes it. After Accept, analytics storage is granted and advertising storage stays denied. - A navigation is a `virtual_page_view` object, not a `config` call and not `gtag('event', 'page_view')`. The container maps that event to a GA4 `page_view`. The GA4 tag should not send a page view of its own, or the visit would be counted twice. - Other events are objects with an `event` key, and each one becomes a measurement. - `segment_update` carries `segments`, the comma-separated list of segments this browser currently matches. In the container, that value is the GA4 user property `segments`. This site does not call `gtag('set')`. One more small decision: the page drops the container's automatic `gtm.scrollDepth` event, so it doesn't double the site's scroll events, which fire the first time a page crosses 25, 50, 75, and 90 percent. ## Consent decides whether the container loads at all On this site, a visitor who hasn't accepted measurement doesn't get Google Tag Manager. `gtm.js` loads only after Accept, and the site holds back event pushes until then too. A denied visit produces no collect request, and no `_ga` cookie gets written. The lab still lets you see what *would* happen. Press one of the **Try it** buttons with consent off and the pipeline shows the push, then stops at the consent check with a blocked network step. Those denied rows never reach the real queue, so the container never sees them. That's the same idea the beacon notes ended on. Building an event, showing it to you, and sending it are separate decisions. ## What the hit carries Accept, press **Try it** again, and a request goes to `https://www.google-analytics.com/g/collect`. A `virtual_page_view` push should arrive on that request as `page_view`. The lab captures it by wrapping `sendBeacon`, `fetch`, and `XMLHttpRequest`, with a `PerformanceObserver` as a backup, then decodes it beside the push that caused it. Most of the meaning sits in a handful of short keys, in the query string, or in the body when several events are batched together: | Key | What it is | | --- | --- | | `en` | Event name: `page_view`, `click`, `scroll`, and the rest. A `virtual_page_view` push is mapped to `page_view`. | | `tid` | The measurement ID of the property. On this site it's `G-JDE6BF33RY`. | | `cid` | Client ID. After consent, Google keeps it in the `_ga` cookie so later hits from this browser can be tied together. | | `sid` | Session ID for the current visit. | | `gcs` | Consent state. | | `dl` | Document location, the page URL the hit is about. | | `dt` | Document title. | | `ep.*` | Event parameters, such as `ep.page_path`, `ep.percent_scrolled`, or `ep.link_url`. | | `up.*` | User properties, such as `up.segments`, copied from the `segment_update` event. | `gcs` is the one worth learning to read. It starts with `G1`, then one digit for ad storage and one for analytics storage, where `1` is granted and `0` is denied. After Accept on this site you'll see `G101`: analytics on, advertising off. That's the whole consent policy, written into every request. ## What the hit leaves out The lab ends with the event schema, the full list of events this site sends to Google and the parameters each one is allowed to carry. The rule behind it is short: emails, names, and form values are never parameters. A few examples show what that looks like in practice. When the contact form or the white paper form stores a submission, `generate_lead` carries one parameter, `form_name`, set to `contact` or `white_paper`. It doesn't carry the name, email, company, or message. When you click a recommended-next card, `select_content` says which post you picked and why it was recommended. `login` and `sign_up` only ever report `method: password`. The schema is the contract. If a value isn't on that list, it doesn't leave the page through this path. ## Two paths, side by side Go back to the [Collection Lab](https://www.blackboxpersonalization.com/lab) with consent on and look at the network panel. You'll see both paths in one list. The `POST /api/collect` beacon is a first-party row. `googletagmanager.com` and `google-analytics.com`, or a regional host like `region1.google-analytics.com`, are labeled third-party Google Analytics. They describe the same activity, but they end up in different places. The first-party record is one I can open, join, and explain row by row. The Google hit becomes part of someone else's model, and the [Live](https://www.blackboxpersonalization.com/live) page shows what that third-party view of the same visits looks like. Neither is a black box once you can read the request. That's the point of putting both on one screen. ## Ghosts in the beacon log - [Ghosts in the beacon log](https://www.blackboxpersonalization.com/blog/ghosts-in-the-beacon-log): Not every visit is a person. Some are crawlers, some are headless browsers, and some are scripts firing too fast to be human. Here is how this site marks those ghosts, why it keeps the rows instead of deleting them, and how the bots lab shows the count. Halloween is the one night a year when it is fine to have a few ghosts in the house. Your analytics is different. Ghost visits show up every night: a crawler reading every page, a headless browser running a test script, a loop that fires the same event forty times a minute. None of them is a person, and all of them can land in your reports looking like one. The [bots lab](https://www.blackboxpersonalization.com/lab/bots) shows how this site handles them. It counts the latest stored beacons, splits them into four buckets, and lists the exact rules that put a beacon in each one. This post walks through those rules, why the marked rows are kept instead of deleted, and what that means if you are building your own collection. ## How many ghosts are out there The scale is not small. The [2025 Imperva Bad Bot Report](https://www.imperva.com/resources/reports/2025-Bad-Bot-Report.pdf) found that automated traffic made up 51% of all web traffic in 2024, the first time in a decade it passed human activity. It put malicious bots at 37% of all internet traffic, up from 32% in 2023, and good bots at 14%. That is one vendor's view of its own network, not a measurement of your site, but it is a fair warning that "a visit" and "a person" are not the same thing. Not every bot is bad, either. Search crawlers and AI crawlers read pages so they can be found. You want them on the site. You just do not want them counted as customers. ## What Google Analytics does for you, and what it hides If you use Google Analytics, some of this is already handled. Google's help page on [known bot-traffic exclusion](https://support.google.com/analytics/answer/9888366) says traffic from known bots and spiders is excluded automatically, identified with a combination of Google research and the [IAB/ABC International Spiders and Bots List](https://iabtechlab.com/software/iababc-international-spiders-and-bots-list/). The same page says you cannot turn that exclusion off and cannot see how much known bot traffic was excluded. That is a reasonable default. It is also a black box. You get a cleaner number, but you do not get to see what was removed or why. The point of this site is to open that box, so its own first-party beacon pipeline does the filtering out in the open. ## The three rules this site actually runs Every consented beacon goes through three checks when it is stored. The [bots lab](https://www.blackboxpersonalization.com/lab/bots) lists them, and they match the code. 1. **Crawler.** The server reads the request's User-Agent with the Next.js [`userAgent` helper](https://nextjs.org/docs/app/api-reference/functions/userAgent), which returns an `isBot` flag for requests from known bots. If it is true, the row is marked `crawler`. That list covers names such as Googlebot and GPTBot. 2. **Headless.** If the browser string contains `HeadlessChrome`, `PhantomJS`, `Puppeteer`, `Playwright`, or `Selenium`, the row is marked `headless`. This is a separate check because `HeadlessChrome` is not on the Next.js crawler list. A headless browser is a real browser engine driven by a script, often a test runner or a scraper. 3. **Rate.** If the same anonymous visitor key already stored 40 beacons in the previous 60 seconds, the next row is marked `rate`. A person reading pages does not fire that fast. A loop does. Anything that matches none of these is `included`. Those four counters are the boxes at the top of the lab. ## What the rules do not do The lab is just as plain about the limits. This site does not read `navigator.webdriver`, mouse paths, or screen size to decide. It does not block the request, and it does not delete the beacon. That list matters, because it tells you what slips through. The User-Agent is a string the client sends, and a script can send any string it likes. A headless browser that reports itself as ordinary Chrome passes rule 2. A scraper that paces itself under 40 beacons a minute passes rule 3. The [W3C WebDriver spec](https://www.w3.org/TR/webdriver2/) defines `navigator.webdriver` as a way for a browser under automation to say so, but a script that wants to hide can avoid it too. So these rules catch the honest ghosts and the clumsy ones. They do not catch a bot that is working hard to look human, and no simple rule set does. That is fine for a demo site, and it is a useful baseline for a small one. Commercial bot management adds behavior signals, IP reputation, and challenges on top. You can add those later. Start by marking the obvious cases so they stop skewing your numbers. ## Mark, don't delete Here is the design choice worth copying. A marked row stays in the collection log. The reports that answer "what did people do?" skip it. On this site that means [attribution](https://www.blackboxpersonalization.com/lab/attribution), [funnels](https://www.blackboxpersonalization.com/lab/funnels), [cohorts](https://www.blackboxpersonalization.com/lab/cohorts), [anomalies](https://www.blackboxpersonalization.com/lab/anomalies), and the other Understand counts leave out `crawler`, `headless`, and `rate` rows. Why keep them at all? - **You can audit the rule.** If a real visitor gets marked by mistake, the row is still there to find. Deleting it would hide the error forever. - **You can change the rule later.** Tighten the rate limit or add a pattern, and you can re-run it over history instead of only fixing the future. - **The ghosts are data too.** A spike in `crawler` rows after a launch tells you search and AI crawlers found the new pages. A spike in `rate` rows can point to a broken script or someone hammering a form. The lab is also honest about where the mark does not reach. [Pathing](https://www.blackboxpersonalization.com/lab/pathing), [recommendations](https://www.blackboxpersonalization.com/lab/recommendations), and Google Analytics do not read it. Google Analytics has its own exclusion, described above. The other two are places where a determined bot could still leave footprints, and the lab says so instead of pretending the filter is universal. ## Ghosts and personalization This matters beyond reports. If you personalize from behavior, a bot can teach your model the wrong lesson. A scraper that walks every product page in a category looks like a visitor with a strong interest in that category. A test script that loads the pricing page a hundred times looks like a buyer. The [segments lab](https://www.blackboxpersonalization.com/lab/segments) and the home page's ["Why am I seeing this?" panel](https://www.blackboxpersonalization.com/lab/why) work from first-party signals in the browser. Marking bot rows at storage time is what keeps the aggregate side of that picture, the counts and funnels that tell you which segments matter, from being tuned by things that are not people. You do not need an expensive vendor to get this first layer right. You need a few clear rules, applied when the beacon arrives, with the evidence kept. ## Try it 1. Open the [bots lab](https://www.blackboxpersonalization.com/lab/bots) and look at the four counters for the latest stored beacons. A zero means the rule matched nothing in that set, not that the rule is off. 2. If fewer than eight beacons are stored, the lab says the mix is not a rate yet. Click **Sample data** to see example counters. The sample is labeled as not coming from the database. 3. Read the line under the counters. It tells you how many of these beacons are left out of the Understand reports. 4. Compare that with what you know about your own analytics. If you use Google Analytics, ask yourself where you would see the bots it removed. On this site, you can see them. The takeaway for anyone building collection: you will always have a few ghosts. Name them, mark them when they arrive, keep the evidence, and make sure your reports know the difference between a visitor and a costume with nobody inside. Happy Halloween. ## Sources - [2025 Bad Bot Report, Imperva](https://www.imperva.com/resources/reports/2025-Bad-Bot-Report.pdf) - [Known bot-traffic exclusion, Google Analytics Help](https://support.google.com/analytics/answer/9888366) - [IAB/ABC International Spiders and Bots List, IAB Tech Lab](https://iabtechlab.com/software/iababc-international-spiders-and-bots-list/) - [userAgent, Next.js documentation](https://nextjs.org/docs/app/api-reference/functions/userAgent) - [WebDriver, W3C](https://www.w3.org/TR/webdriver2/) ## Why a first-party cookie forgets you - [Why a first-party cookie forgets you](https://www.blackboxpersonalization.com/blog/why-a-first-party-cookie-forgets-you): A first-party cookie written by JavaScript can still disappear after about a week in Safari. Here is how that differs from a cookie set by the server, what it does to a client id, and how the cookies lab shows it. "First-party" names the cookie jar. In Safari, a cookie JavaScript wrote on your own site can be gone after about a week away, and the next hit starts a new client id. The [cookies lab](https://www.blackboxpersonalization.com/lab/cookies) lists every cookie and storage key on this origin: owner, purpose, lifetime. Open it in Safari and in Chrome. The rows can match. A week later, they will not. ## Two ways to write a cookie Script sets a cookie by assigning [`document.cookie`](https://developer.mozilla.org/en-US/docs/Web/API/Document/cookie). You can add `expires` or `max-age`, plus `path`, `domain`, `secure`, and `samesite`. You cannot mark it `HttpOnly`, and other script on that origin can read the value. The server sets one with [`Set-Cookie`](https://developer.mozilla.org/en-US/docs/Web/HTTP/Reference/Headers/Set-Cookie). [MDN](https://developer.mozilla.org/en-US/docs/Web/HTTP/Guides/Cookies) says `HttpOnly` forbids JavaScript from reading the cookie. The browser still sends it later. [ITP 2.1](https://webkit.org/blog/8613/intelligent-tracking-prevention-2-1/) states the same limit: a cookie created through `document.cookie` cannot be `HttpOnly`. ## Seven days, and one shorter case [ITP 2.1](https://webkit.org/blog/8613/intelligent-tracking-prevention-2-1/) caps a persistent `document.cookie` cookie at seven days, on a first-party page and in a third-party iframe. A session cookie stays a session cookie. A shorter expiry is kept. [WebKit's tracking-prevention page](https://webkit.org/tracking-prevention/) says ITP now deletes JavaScript-created cookies and other script-writable storage after 7 days with no click, tap, or typing. Scrolling does not count. Link decoration is the one-day case. [ITP 2.2](https://webkit.org/blog/8828/intelligent-tracking-prevention-2-2/) applies it when a classified cross-site tracker navigated in and the landing URL has a query string or a fragment. The current page calls that 24 hours, on that page only. CNAME cloaking is seven days, on `Set-Cookie`. [WebKit](https://webkit.org/blog/11338/cname-cloaking-and-bounce-tracking-defense/) caps cookies from a subdomain whose CNAME differs from the site and from the top frame's own CNAME. An edge host in front of the site is not that case. The tracking-prevention page caps IP-address cloaking at 7 days. ## The client id [Google's cookie page](https://support.google.com/analytics/answer/11397207?hl=en): the GA4 tags use first-party cookies. `_ga` distinguishes users. `_ga_` plus the container id persists session state. Both default to 2 years. No return visit means at most 400 days in Chrome and 7 in Safari. [`cookie_expires`](https://developers.google.com/analytics/devguides/collection/ga4/reference/config) is `63072000` seconds and refreshes each hit while `cookie_update` is true. Safari still caps the write. A week away, the id is new, and returning-visitor and attribution numbers restart. Rules do not read `_ga`. See [From a push to a Google hit](https://www.blackboxpersonalization.com/blog/from-a-push-to-a-google-hit) and [When the Google tag never arrives](https://www.blackboxpersonalization.com/blog/when-the-google-tag-never-arrives). ## localStorage [MDN](https://developer.mozilla.org/en-US/docs/Web/API/Window/localStorage): `localStorage` has no expiry, and `sessionStorage` ends with the page session. [WebKit, March 2020](https://webkit.org/blog/10218/full-third-party-cookie-blocking-and-more/), still deletes both, plus IndexedDB, media keys, and service workers, after seven days of Safari use without interaction. The rules read `site.visitor.v1`, the record behind [Personalization from the same signals](https://www.blackboxpersonalization.com/blog/personalization-from-the-same-signals). Safari can drop it while `bbp.visitor` remains. ## What the server can keep `bbp.visitor` is `Set-Cookie` from the collect route when a beacon is stored and the cookie is missing: `HttpOnly`, `SameSite=Lax`, `Path=/`, `Secure` on https, `Max-Age` 400 days (34,560,000 seconds). Later beacons leave that date. Ordinary events need granted consent, and `Sec-GPC: 1` is refused. The gate is [What a beacon actually sends](https://www.blackboxpersonalization.com/blog/what-a-beacon-actually-sends). The seven-day cap is the `document.cookie` rule. This origin's response is outside it. A cloaked vendor subdomain is not, and consent still has to be granted. [Since Chrome 104](https://developer.chrome.com/blog/cookie-max-age-expires), `Expires` and `Max-Age` stop at 400 days for a server-set cookie too. Script cannot read `bbp.visitor`. The rules use `site.visitor.v1`. ## Other browsers [Firefox Total Cookie Protection](https://blog.mozilla.org/en/mozilla/firefox-rolls-out-total-cookie-protection-by-default-to-all-users-worldwide/) is on by default: each site has its own cookie jar. That confines a cookie to the site where it was set, which is a different control from Safari's script-cookie cap. [Chrome's 22 April 2025 update](https://privacysandbox.com/news/privacy-sandbox-next-steps/) kept third-party cookie choice in Privacy and Security settings, with no new prompt. Incognito already blocks third-party cookies. The first-party date limit is 400 days. ## What the lab shows Names come from `/api/cookie-inventory` and from the page. Values stay hidden. A `document.cookie` name says "Readable by script." A server-only name says "HttpOnly — script cannot read the value." Party is the identifier's owner, so `_ga` says third party. Speed Insights sets none of these keys. The lifetime is the catalog's, because `Max-Age` is not returned. Unknown names say so. Cookies, this site, first party: - `bbp.visitor`. Anonymous id for `collect_event.visitor_key`. HttpOnly. 400 days, SameSite=Lax, Secure on https. - `bbp_experiment`. HttpOnly counts already recorded, so a refresh does not double-count. 30 days, SameSite=Lax. - `bbp_form`. HttpOnly form steps already counted. No field text. 12 hours, SameSite=Lax. - `bbp_blocking`. HttpOnly. [Blocking lab](https://www.blackboxpersonalization.com/lab/blocking) counts this browser once per 12 hours. Value `1`. 12 hours, SameSite=Lax. - `bbp.reset`. Names only, after Reset my data. Removed when dismissed. 120 seconds. Not HttpOnly. - `bbp.track`. Event name for a dataLayer push after login, sign-up, or logout, then cleared. 120 seconds. - `session-token`. Auth.js session. Value hidden. Account id inside is `user.id`. 30 days from sign-in. HttpOnly, SameSite=Lax. - `csrf-token`. Auth.js. Pairs the sign-in form with this browser. Value hidden. Sign-in attempt. HttpOnly. - `callback-url`. Auth.js. Return path after a magic link. Not a personalization signal. Sign-in attempt. Google Analytics, party "Third party": - A `_ga` name. Client id and GA4 session cookie from `gtag.js` after analytics storage is granted. Cell: `_ga` up to 2 years, `_gid` about 24 hours, `_gat` about 1 minute, `_ga_` is the session cookie. Google lists `_ga` and `_ga_` at 2 years. Here: `_ga_JDE6BF33RY`. Rules skip it. localStorage, this site, first party, until you clear site data. The value stays in this browser. - `site.consent.v1`. `granted` or `denied`. Gates beacons and the dataLayer. - `site.consent.choice.v1`. Choice, time, and version. No account id. - `site.visitor.v1`. Visit count and topic order for the rules. Not `bbp.visitor`. Written only with consent. - `site.zero.v1`. Profiling answers (role, goal, industry), until reset or a clear. - `site.exit.v1`. Time of the last exit note. The note hides for 12 hours. The key remains until cleared. - `site.experiment.visitor`. [Experiments-lab](https://www.blackboxpersonalization.com/lab/experiments) key. Not the server visitor id. - `site.why.v1`. Whether Why labels are on. Not written while consent is off. sessionStorage, this site, first party. This tab. Closing the tab removes it. - `site.session.v1`. Pages, scroll, and dwell. The rules read it. - `site.path.v1`. Page order on the [pathing lab](https://www.blackboxpersonalization.com/lab/pathing). - `site.experiment.bandit`. Thompson-sampling arm, so a refresh keeps the headline. - `white-paper-lead:` prefix. Stops a second `generate_lead` on the thank-you page. Value `1`. The [storage lab](https://www.blackboxpersonalization.com/lab/storage) shows values, sizes, and the page request's cookies, HttpOnly included. Consent off: visitor and session records are not written and do not change the page. A consented beacon stores `visitor_key` from `bbp.visitor` with `user_id` null. On sign-in, `claimVisitor` sets Auth.js `user.id` on unclaimed rows for this key and skips rows that already have an account. Later beacons store both. `_ga` is not joined. ## Try it 1. In Safari, open the [cookies lab](https://www.blackboxpersonalization.com/lab/cookies), accept measurement, and reload after a beacon. `bbp.visitor` is HttpOnly. If the tag ran, `_ga` and `_ga_JDE6BF33RY` are readable by script, and the cell still says up to 2 years. 2. Open it in Chrome. Compare the expiry each browser stored. Chrome stops at 400 days. Safari's script cookie is seven days without interaction, or one day after link decoration. `bbp.visitor` can keep the longer date. 3. In the [storage lab](https://www.blackboxpersonalization.com/lab/storage), the returning-visitor line is `site.visitor.v1`. In Safari it shares the script-storage timer. The server cookie can still arrive after that key is gone. ## Sources - [Intelligent Tracking Prevention 2.1, WebKit](https://webkit.org/blog/8613/intelligent-tracking-prevention-2-1/) - [Intelligent Tracking Prevention 2.2, WebKit](https://webkit.org/blog/8828/intelligent-tracking-prevention-2-2/) - [Full Third-Party Cookie Blocking and More, WebKit](https://webkit.org/blog/10218/full-third-party-cookie-blocking-and-more/) - [CNAME Cloaking and Bounce Tracking Defense, WebKit](https://webkit.org/blog/11338/cname-cloaking-and-bounce-tracking-defense/) - [Tracking Prevention in WebKit](https://webkit.org/tracking-prevention/) - [Document.cookie, MDN](https://developer.mozilla.org/en-US/docs/Web/API/Document/cookie) - [Set-Cookie, MDN](https://developer.mozilla.org/en-US/docs/Web/HTTP/Reference/Headers/Set-Cookie) - [Using HTTP cookies, MDN](https://developer.mozilla.org/en-US/docs/Web/HTTP/Guides/Cookies) - [Window.localStorage, MDN](https://developer.mozilla.org/en-US/docs/Web/API/Window/localStorage) - [GA4 cookie usage on websites, Google Analytics Help](https://support.google.com/analytics/answer/11397207?hl=en) - [GA4 configuration fields, Google for Developers](https://developers.google.com/analytics/devguides/collection/ga4/reference/config) - [Cookie Expires and Max-Age upper limit, Chrome for Developers](https://developer.chrome.com/blog/cookie-max-age-expires) - [Next steps for Privacy Sandbox and tracking protections in Chrome](https://privacysandbox.com/news/privacy-sandbox-next-steps/) - [Firefox rolls out Total Cookie Protection by default, Mozilla](https://blog.mozilla.org/en/mozilla/firefox-rolls-out-total-cookie-protection-by-default-to-all-users-worldwide/) ## How an agent reads this site - [How an agent reads this site](https://www.blackboxpersonalization.com/blog/how-an-agent-reads-this-site): The Agentic column is ten simulations of a loop: read what this site already knows, pick a real page, and name the rule. The llms.txt lab is where that reading starts. An agent, in the sense these pages use the word, is a loop. It reads what it is allowed to read, picks a next step, changes the page, and can remember whether that step helped. The [lab menu](https://www.blackboxpersonalization.com/lab) leads with an Agentic column of ten simulations, and each page names the rule that fired. The [About page](https://www.blackboxpersonalization.com/about) maps the raw material, from device and referrer to consent and segments. These labs are what a careful agent would do with it. ## What the agent is allowed to read Before an agent acts, it needs a place to start. The [llms.txt lab](https://www.blackboxpersonalization.com/lab/llms-txt) is that place. [llmstxt.org](https://llmstxt.org) is a 2024 proposal by Jeremy Howard. A site may publish Markdown at `/llms.txt` so a model can begin with a name, a short summary, and a few chosen links. This site also serves [`/llms-full.txt`](https://www.blackboxpersonalization.com/llms-full.txt), which repeats the public pages and adds a short note where a page has one. The lab puts three files side by side, because they answer three different questions. `robots.txt` tells a polite crawler which paths it may fetch. The sitemap is the URL list. [`/llms.txt`](https://www.blackboxpersonalization.com/llms.txt) is the reading guide. Both llms files are built with the site, from the public page catalog, the lab groups, and these notes. Which AI tools read the file is unconfirmed, and the public evidence is limited. A fetch is a fetch. The [AI referrals lab](https://www.blackboxpersonalization.com/lab/ai-referrals) logs a document request whose User-Agent matches this site’s crawler list, and the [llms.txt lab](https://www.blackboxpersonalization.com/lab/llms-txt) shows that log for `/llms.txt` and `/llms-full.txt`. ![The llms.txt lab on the live site, captured 5 November 2026. robots.txt, the sitemap, and the generated llms.txt file sit side by side.](https://www.blackboxpersonalization.com/blog-image/how-an-agent-reads-this-site/llms-txt.png) ## Sense, decide, act, learn The [Agentic Journey Lab](https://www.blackboxpersonalization.com/lab/agentic-journey) runs the loop in five panels, on signals this site already has. Sense reads this browser, or marks a field missing. Stored visit history stays missing while consent is off. Infer intent ranks four goals: the basics, measurement, personalization, or help. Confidence is that goal’s points divided by the top goal’s points. Decide path highlights one real page, and ignore, override, or a returning visitor moves the highlight. Act reorders cards only when consent is granted. Learn is an epsilon-greedy bandit. The panel prints epsilon as 0.2. A draw under that line explores a random goal. Otherwise the highest stored reward rate wins, or the rules do if nothing has been clicked. The log stores category tokens in this browser, and only with consent. Reset my data clears it. ![The Agentic Journey Lab on the live site, captured 5 November 2026, with consent declined. The Sense panel lists signals this browser already has and marks stored history as missing.](https://www.blackboxpersonalization.com/blog-image/how-an-agent-reads-this-site/agentic-journey.png) ## A sentence in, one real page out Two labs take a sentence and answer with pages that already exist. Neither one calls a model, and neither one stores the sentence. The [agent handoff lab](https://www.blackboxpersonalization.com/lab/agent-handoff) is the short version. Keyword rules pick one page and return [schema.org](https://schema.org) JSON: a LearningResource for a lab, or a ContactPage for contact. An unmatched question chooses no page. The preset best beginner guide matches the beginner list and points at the [Collection Lab](https://www.blackboxpersonalization.com/lab). Go deep on GA4 points at the [dataLayer lab](https://www.blackboxpersonalization.com/lab/datalayer). I need help points at [contact](https://www.blackboxpersonalization.com/contact). ![The agent handoff lab on the live site, captured 5 November 2026. The preset best beginner guide matched a keyword rule and returned schema.org JSON for the Collection Lab.](https://www.blackboxpersonalization.com/blog-image/how-an-agent-reads-this-site/agent-handoff.png) The [concierge lab](https://www.blackboxpersonalization.com/lab/concierge) builds a path, and each stop names the reason. A beginner sentence goes to the [Collection Lab](https://www.blackboxpersonalization.com/lab), then the [consent lab](https://www.blackboxpersonalization.com/lab/consent), then the [showcase](https://www.blackboxpersonalization.com/showcase): consent comes before measurement, and the showcase names signals the browser already has. A measurement sentence goes to the [dataLayer lab](https://www.blackboxpersonalization.com/lab/datalayer), then the [server-side lab](https://www.blackboxpersonalization.com/lab/server-side), then the [consent lab](https://www.blackboxpersonalization.com/lab/consent). An unmatched sentence builds no path. ## One nudge, and the rules that hold it back The [next best action lab](https://www.blackboxpersonalization.com/lab/next-best-action) offers one thing: a guide, another lab, a consult nudge, or nothing. A personalization lab already opened leads to consult. A measurement lab already opened leads to another lab. Otherwise the offer is a guide. The same action stops after two showings. Dwell under 15 seconds counts as rushed, and the offer stays back. Consent off means nothing is offered. The [friction detector](https://www.blackboxpersonalization.com/lab/friction-detector) watches a demo widget. Three clicks on Continue inside one second count as rage, and the page offers a simpler version. A dead control offers help. Three direction changes in the sample scroller count as thrash, and the page offers a shortcut. A field focused for three seconds with no typing is stalled. The log names the rule and stays in memory for that view. The [self-critique lab](https://www.blackboxpersonalization.com/lab/self-critique) proposes an action, then fairness, sensitive categories, a frequency cap, and consent can veto it. One failed check holds the action. Inferred income and a precise place fail the sensitive check. A repeated consult fails the cap after two showings. Consent off vetoes every proposal. The [privacy negotiator](https://www.blackboxpersonalization.com/lab/privacy-negotiator) shows the same boundary as a trade. Off blocks stored history and reordering. Measurement allows stored visits. Personalization allows reordering. Local examples need an extra location trade, and that trade turns off again. Sensitive categories and identity stay off. The choice is written in this browser only when measurement consent is on. ## Same blocks, different order The [content assembly lab](https://www.blackboxpersonalization.com/lab/content-assembly) reorders four blocks: intro, proof, how, and offer. The first-time marketer preset leads with the intro, then how, the offer, and proof. The analytics engineer preset leads with proof, which points at the [consent lab](https://www.blackboxpersonalization.com/lab/consent), then how, intro, and the offer. Nothing on the page is a measured statistic. The [goal-seeking test](https://www.blackboxpersonalization.com/lab/goal-seeking-test) charts simulated traffic for two segments. Epsilon 0.2 explores. Otherwise the arm with the higher simulated win rate so far gets the next arrival. The planted rates are written on the chart. The [cross-channel memory lab](https://www.blackboxpersonalization.com/lab/cross-channel-memory) walks a script: an ad, then this site, then an email, then a return. A channel class and a topic carry forward. The click id, the advertising id, and the email address stay forgotten, including on the return. ## Where a model writes a line Those ten labs are rules. A model that writes a headline lives on the [AI copy lab](https://www.blackboxpersonalization.com/lab/ai-copy). It opens in manual mode and fills the headline and summary from a template, using a segment and a role. The site owner can turn on AI-assisted mode with an OpenAI key. If that call fails, the template is what you see. The inputs stay next to the line. ## A short walk 1. Read the brief on the [llms.txt lab](https://www.blackboxpersonalization.com/lab/llms-txt), then open [`/llms.txt`](https://www.blackboxpersonalization.com/llms.txt). 2. Run the [Agentic Journey Lab](https://www.blackboxpersonalization.com/lab/agentic-journey) once with consent off and once after Accept. Act stays still until consent is on. 3. Type a goal into the [concierge lab](https://www.blackboxpersonalization.com/lab/concierge). Ask the [agent handoff lab](https://www.blackboxpersonalization.com/lab/agent-handoff) the same idea in a shorter sentence. 4. Open the [privacy negotiator](https://www.blackboxpersonalization.com/lab/privacy-negotiator) and the [self-critique lab](https://www.blackboxpersonalization.com/lab/self-critique) before copying the pattern onto a site with customers. The [Data APIs page](https://www.blackboxpersonalization.com/lab/data-apis) lists labs that call an outside service. The llms files are generated here. The agentic labs stay in the browser and name the rule. ## What a beacon actually sends - [What a beacon actually sends](https://www.blackboxpersonalization.com/blog/what-a-beacon-actually-sends): What a browser reveals in one beacon: the page view body, the consent gate it has to pass, and the record that comes back from this site. Free. "Analytics" usually means a script you can't read sending a request you never see. On this site, one event is small enough to read in full. This note follows a single page view from the moment the browser builds it to the moment the server writes it down. The [Collection Lab](https://www.blackboxpersonalization.com/lab) shows the browser, the network, and the record side by side. Open it in another tab and follow along. The [dataLayer lab](https://www.blackboxpersonalization.com/lab/datalayer) shows one of those events as a push, and the [fingerprint lab](https://www.blackboxpersonalization.com/lab/fingerprint) names what this browser could reveal before anything is sent. The page view itself is sent by the site-wide tracker, which runs on every page, including the lab. ## Before anything leaves: the browser panel The first panel in the Lab is labeled **Browser**. It reads `navigator`, `screen`, `document`, and the Performance API: viewport and device pixel ratio, language and time zone, the Network Information API where it exists, navigation timing, Largest Contentful Paint, Cumulative Layout Shift, Interaction to Next Paint, the referrer, campaign parameters, and the *names* of cookies and localStorage keys. Navigation timing is the single entry from `performance.getEntriesByType("navigation")`: DNS, TCP, TLS, time to first byte, DOM interactive, DOM content loaded, and the load event. Those numbers are how this page experienced the network, not a guess filled in afterward. All of it stays on the device. The panel says so, and the code agrees. A beacon carries a smaller profile, not this list. Reading a value in the page and sending it are two separate decisions. ## The gate A beacon is only sent if two things are true: 1. You have turned measurement consent on. The choice is stored in this browser under one localStorage key, `site.consent.v1`. Only the value `granted` counts. Anything else, or no value at all, counts as off. 2. Your browser isn't sending Global Privacy Control. If `navigator.globalPrivacyControl` is true, the site treats consent as denied, whatever the stored choice says. The server also refuses the post when the request carries `Sec-GPC: 1`, even if the body claims consent was granted. Consent is off by default. The Google tag is not requested until you accept. Accepting loads it with analytics storage on and advertising storage off. Turning measurement off sets analytics storage back to denied. That denial is not pushed to the dataLayer. When consent is off, the Lab still *builds* the web vital it is about to show you, with `consent` set to `denied`. Then it drops it. The **Back end** panel marks it `suppressed` and shows "the raw payload that would have been sent". No request is made and the server has nothing to enrich. The site-wide tracker is stricter: with consent off, it doesn't build a page view, a click, or a scroll event at all. There is one deliberate exception. If you switch consent from on to off, one `consent_change` is sent so that decision is on the record. Its `consent` field is `denied`, and `properties.choice` is `denied`. The route stores that single row. The profile on it is the one already in memory when the switch flips, so `presentation` can still say `personalized`. The consent field is the part that changed. A first Decline, when consent was never granted, sends nothing. After the withdrawal, nothing else is sent. If the browser is sending Global Privacy Control, the server refuses this post too and stores nothing. The header, the lab switch, and the consent banner are marked so the click tracker skips them. Turning measurement off does not add a click beacon. The `consent_change` is the withdrawal. ## The body With consent on, here is a page view as it leaves the browser. These are the keys the site-wide tracker writes. The values are one returning desktop visit: morning in Los Angeles, a previous note open, scroll already at 40%, no campaign parameters. Nothing in the object is optional padding. ```json { "event_id": "3f6c1a2e-8b14-4c7a-9f02-6d1e0a4b7c93", "event_name": "page_view", "occurred_at": "2026-10-04T16:04:11.418Z", "consent": "granted", "page": { "path": "/lab", "title": "How data collection works | Black Box Personalization", "referrer": "/blog/personalization-from-the-same-signals" }, "campaign": {}, "properties": { "session_id": "7c2a9e14-6b31-4f08-9d2e-1a0b6c4d8e77", "sequence": 4, "timestamp": "2026-10-04T16:04:11.418Z", "path": "/lab", "referrer_path": "/blog/personalization-from-the-same-signals" }, "client_profile": { "presentation": "personalized", "presentation_reason": "rules", "visitor": "returning", "visit_count": 3, "timezone": "America/Los_Angeles", "local_time": "9:04 AM", "local_hour": 9, "time_of_day": "morning", "device_class": "desktop", "viewport_width": 1440, "viewport_height": 900, "connection": { "api": "available", "effective_type": "4g", "downlink_mbps": 10, "rtt_ms": 50, "save_data": false, "slow": false }, "referrer_class": "internal", "referrer_host": "www.blackboxpersonalization.com", "utm": {}, "pages_viewed": ["/", "/blog", "/blog/personalization-from-the-same-signals", "/lab"], "topics_viewed": ["personalization", "consent"], "scroll_depth": 40, "click_count": 2, "last_click": "Personalization from the same signals", "rules": [ { "slot": "hero", "id": "returning-visitor", "name": "Returning visitor" }, { "slot": "cta", "id": "scroll-depth", "name": "Scroll depth" }, { "slot": "recommendations", "id": "topic-affinity", "name": "Topic affinity" }, { "slot": "layout", "id": "standard-weight", "name": "Standard image weight" } ], "segments": [ { "id": "new-visitor", "on": false }, { "id": "returning-visitor", "on": true }, { "id": "mobile", "on": false }, { "id": "desktop", "on": true }, { "id": "engaged-reader", "on": false }, { "id": "lab-explorer", "on": false }, { "id": "west-coast", "on": true }, { "id": "other-region", "on": false }, { "id": "night-owl", "on": false }, { "id": "campaign-visitor", "on": false }, { "id": "signed-in", "on": false } ] } } ``` Read it top to bottom: - **`event_id`** is a random UUID made in the browser, so the record can be matched to the request that carried it. - **`event_name`** is what happened. The route accepts `page_view`, `click`, `scroll_depth`, `web_vital`, `form_start`, `form_submit`, `consent_change`, `sign_in`, `sign_out`, `outbound_click`, `location_sample`, and `precise_location`. - **`occurred_at`** is the browser's clock, not the server's. Hold onto that. It is the same instant as `properties.timestamp`. - **`page`** is the path, the document title, and the previous path in this tab. That previous path lives in `sessionStorage` under `site.path.v1`. It is not `document.referrer`. The title is cut at 180 characters. - **`campaign`** is the campaign parameters the tracker reads: `utm_source`, `utm_medium`, `utm_campaign`, `utm_term`, `utm_content`, and `utm_id`. It is an empty object when none of those are on the URL. - **`properties`** always carries a session id, a sequence number, a timestamp, the path, and the previous path. The session id ends when the tab closes. A scroll event adds `percent`, one of 25, 50, 75, or 100, and each of those is sent once per path. Extra string properties are cut at 200 characters. The route repeats that cut and keeps at most 20 properties. A web vital is a different body, built by the lab rather than this tracker: `metric`, a number (`value_ms` for LCP and INP, `value` for CLS), and a `rating`. It does not carry the session fields. - **`client_profile`** is the personalization profile this page had at that moment. `referrer_class` and `referrer_host` come from `document.referrer`, which is a separate fact from `page.referrer`. `rules` is the winning rule for the hero, the call to action, the notes, and the layout. `segments` is every segment flag, on or off. `topics_viewed` is the topic list from notes this browser has actually opened. That last field is the one worth reading slowly. A site that personalizes from your reading history and then reports on it is sending that history. Here it is sent in plain JSON, to the site's own origin, and only after you said yes. ## The request The body goes out as a single `fetch`: ``` POST /api/collect content-type: application/json ``` The site-wide tracker sets `keepalive`, so the request can finish even if you navigate away. It does not give up on a timer. The lab's own sender, the one that posts web vitals, aborts a request after twelve seconds and caps itself at twelve events in any ten-second window. That cap does not cover scroll events. The route still stores a fast series. Once the same visitor key has 40 stored beacons in 60 seconds, the next row is kept and marked with `bot_reason` set to `rate`. Then look at the **Network** panel. Once the request finishes, it shows up there as a first-party row, with the same host as the page. Google Analytics rows, from googletagmanager.com and google-analytics.com, are grouped separately and labeled third-party Google Analytics. The Resource Timing list itself is never uploaded. A row's transfer size is sometimes zero. That usually means the response was cached, or the other server didn't send `Timing-Allow-Origin` and the browser hid the body size. The request still happened. The missing number is part of the list. ## What comes back The route doesn't store the body as-is. It wraps it. The response holds the stored record, and the **Back end** panel shows both blocks side by side when the event is one the lab itself posted: **Raw payload sent** and **Enriched record stored**. A page view from the site-wide tracker raises the stored count on that panel. The raw JSON block there is the lab's own event. The record keeps your body as `raw_payload`, gives it a `record_id`, and stamps `received_at` and `stored_in`. Then it adds `server_enriched_profile`. These are the keys, for the same page view, on a request that carried Vercel headers. The address below is a documentation address, not a visitor. On a laptop most of those headers are null, and the reconciliation lines say so instead of inventing a place. ```json { "record_id": "a91e4c20-5d77-4b18-8c3f-2e6a0d9b1f44", "received_at": "2026-10-04T16:04:11.640Z", "stored_in": "database", "endpoint": "POST /api/collect", "raw_payload": { "event_id": "3f6c1a2e-8b14-4c7a-9f02-6d1e0a4b7c93", "event_name": "page_view", "occurred_at": "2026-10-04T16:04:11.418Z", "consent": "granted", "page": { "path": "/lab", "title": "How data collection works | Black Box Personalization", "referrer": "/blog/personalization-from-the-same-signals" }, "campaign": {}, "properties": { "session_id": "7c2a9e14-6b31-4f08-9d2e-1a0b6c4d8e77", "sequence": 4, "timestamp": "2026-10-04T16:04:11.418Z", "path": "/lab", "referrer_path": "/blog/personalization-from-the-same-signals" }, "client_profile": { "presentation": "personalized", "presentation_reason": "rules", "visitor": "returning", "visit_count": 3, "timezone": "America/Los_Angeles", "local_time": "9:04 AM", "local_hour": 9, "time_of_day": "morning", "device_class": "desktop", "viewport_width": 1440, "viewport_height": 900, "connection": { "api": "available", "effective_type": "4g", "downlink_mbps": 10, "rtt_ms": 50, "save_data": false, "slow": false }, "referrer_class": "internal", "referrer_host": "www.blackboxpersonalization.com", "utm": {}, "pages_viewed": ["/", "/blog", "/blog/personalization-from-the-same-signals", "/lab"], "topics_viewed": ["personalization", "consent"], "scroll_depth": 40, "click_count": 2, "last_click": "Personalization from the same signals", "rules": [ { "slot": "hero", "id": "returning-visitor", "name": "Returning visitor" }, { "slot": "cta", "id": "scroll-depth", "name": "Scroll depth" }, { "slot": "recommendations", "id": "topic-affinity", "name": "Topic affinity" }, { "slot": "layout", "id": "standard-weight", "name": "Standard image weight" } ], "segments": [ { "id": "new-visitor", "on": false }, { "id": "returning-visitor", "on": true }, { "id": "mobile", "on": false }, { "id": "desktop", "on": true }, { "id": "engaged-reader", "on": false }, { "id": "lab-explorer", "on": false }, { "id": "west-coast", "on": true }, { "id": "other-region", "on": false }, { "id": "night-owl", "on": false }, { "id": "campaign-visitor", "on": false }, { "id": "signed-in", "on": false } ] } }, "server_enriched_profile": { "source": "POST /api/collect", "received_at": "2026-10-04T16:04:11.640Z", "clock_skew_ms": 222, "client_reported": { "presentation": "personalized", "presentation_reason": "rules", "visitor": "returning", "visit_count": 3, "timezone": "America/Los_Angeles", "local_time": "9:04 AM", "local_hour": 9, "time_of_day": "morning", "device_class": "desktop", "viewport_width": 1440, "viewport_height": 900, "connection": { "api": "available", "effective_type": "4g", "downlink_mbps": 10, "rtt_ms": 50, "save_data": false, "slow": false }, "referrer_class": "internal", "referrer_host": "www.blackboxpersonalization.com", "utm": {}, "pages_viewed": ["/", "/blog", "/blog/personalization-from-the-same-signals", "/lab"], "topics_viewed": ["personalization", "consent"], "scroll_depth": 40, "click_count": 2, "last_click": "Personalization from the same signals", "rules": [ { "slot": "hero", "id": "returning-visitor", "name": "Returning visitor" }, { "slot": "cta", "id": "scroll-depth", "name": "Scroll depth" }, { "slot": "recommendations", "id": "topic-affinity", "name": "Topic affinity" }, { "slot": "layout", "id": "standard-weight", "name": "Standard image weight" } ], "segments": [ { "id": "new-visitor", "on": false }, { "id": "returning-visitor", "on": true }, { "id": "mobile", "on": false }, { "id": "desktop", "on": true }, { "id": "engaged-reader", "on": false }, { "id": "lab-explorer", "on": false }, { "id": "west-coast", "on": true }, { "id": "other-region", "on": false }, { "id": "night-owl", "on": false }, { "id": "campaign-visitor", "on": false }, { "id": "signed-in", "on": false } ] }, "added_by_server": { "ip": "203.0.113.24", "country": "US", "region": "CA", "city": "San Francisco", "latitude": 37.77, "longitude": -122.42, "timezone": "America/Los_Angeles", "edge_region": "sfo1", "function_region": "iad1", "database_region": "us-east-1", "database_label": "us-east-1 · Ashburn", "parsed_user_agent": { "raw": "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/131.0.0.0 Safari/537.36", "browser": "Chrome", "browser_version": "131.0.0.0", "os": "Mac OS", "os_version": "10.15.7", "device_type": null, "device_vendor": "Apple", "device_model": "Macintosh", "engine": "Blink", "is_bot": false }, "headers": { "x-forwarded-for": "203.0.113.24", "x-real-ip": null, "cf-connecting-ip": null, "true-client-ip": null, "fly-client-ip": null, "x-vercel-forwarded-for": "203.0.113.24", "cf-ipcountry": null, "x-vercel-ip-country": "US", "x-vercel-ip-country-region": "CA", "x-vercel-ip-city": "San Francisco", "x-vercel-ip-latitude": "37.77", "x-vercel-ip-longitude": "-122.42", "x-vercel-ip-timezone": "America/Los_Angeles", "x-vercel-id": "sfo1::iad1::example", "x-forwarded-proto": "https", "x-forwarded-host": "www.blackboxpersonalization.com", "host": "www.blackboxpersonalization.com" } }, "timings": { "request_start": "2026-10-04T16:04:11.418Z", "server_receipt": "2026-10-04T16:04:11.640Z", "db_write": "2026-10-04T16:04:11.710Z" }, "reconciliation": [ "Viewport class (desktop) agrees with the User-Agent reading (desktop, type unset). The page used the viewport.", "An IP was taken from a forwarding header. This app does not look the address up.", "A country header was present (US). It was not computed from the raw IP in this app.", "Vercel edge headers included a city-level point (37.77, -122.42). Latitude and longitude are stored to two decimal places.", "The client clock says 9:04 AM in America/Los_Angeles (hour 9). The server clock at receipt is 2026-10-04T16:04:11.640Z." ], "summary": [ "Client profile: returning visitor, morning in America/Los_Angeles, viewport desktop. Presentation is personalized (rules).", "Server parsed the User-Agent header as Chrome on Mac OS.", "Edge IP header: 203.0.113.24.", "Edge country header: US." ] } } ``` `bot_reason` is not on this record. The route adds it only when the user agent matches a crawler list, looks headless, or trips the rate mark above. The row is still stored. The panel shows `summary` as a short list, then `reconciliation`: lines that explain where the browser's view and the server's view disagree. The usual case is device type. The page decides mobile or desktop from the viewport width, because that's what the layout actually used. The User-Agent parser may say something else, especially for a desktop browser in a narrow window, where `device_type` is null and the server treats that as desktop. The record keeps both answers and says which one personalization trusted. It doesn't quietly overwrite one with the other. `clock_skew_ms` is the server clock minus `occurred_at`. `added_by_server` is the part the browser cannot see: the forwarding address, the edge place headers, the edge and function regions, the database region, and the parsed User-Agent. `client_reported` is the profile the browser sent, copied through so the join stays visible. `timings.db_write` is filled after the row is inserted. Until then it is null. The record is written to the database. In production that database is Postgres, and `GET /api/collect` reports `store` as `database` and `driver` as `postgres`. A local checkout without a Postgres URL uses SQLite and reports `driver` as `sqlite`. The lab reads new rows two ways, from a server-sent event stream and a poll of `GET /api/collect` every two seconds. The poll keeps running while the stream is open. The stream's in-memory listener only sees rows saved in that same process, so the stream also reads the database every two seconds. A row written on another serverless instance still shows up. ## Nothing hidden, on purpose Put together, one beacon is three things you can check: 1. What the browser **knows**: the Browser panel. It never leaves the device. 2. What the browser **sends**: the raw payload. Only with consent, and only to this origin, plus the one `consent_change` that records turning consent off. 3. What the server **adds**: the enriched record. Receipt time, a parsed User-Agent, and whatever headers the edge provided. Most analytics stacks blur those three into one opaque request. Taking them apart is the whole point of this site. You can't judge what a measurement is worth, or whether it's fair, until you can see which of the three a given field came from. Open the [Collection Lab](https://www.blackboxpersonalization.com/lab) and turn consent on. A `page_view` leaves for `POST /api/collect`, shows up as a first-party row in the network list, and increases the stored count. The raw JSON in the back end panel is the lab's web vital, or, with consent off, that same vital marked `suppressed`. Scroll past 25% and the site-wide tracker sends a `scroll_depth` event. Turn consent off after it was on. One `consent_change` is stored with `consent` set to `denied`, and the next scroll is not sent. ## Personalization from the same signals - [Personalization from the same signals](https://www.blackboxpersonalization.com/blog/personalization-from-the-same-signals): The same signals that measure a visit can also change the page. A rules engine, a consent gate, and a panel that names which rule won. Free to read. Collection and personalization are usually described as different products. In the page they are the same facts, used twice. A URL parameter, a referrer, the local hour, the width of the viewport, a save-data flag, a returning visit stored on the device, the notes you have already opened, how far you scrolled, what you clicked: each of those is available in the browser during the session. None of them requires a third-party profile. ## Rules, not a model This site keeps the logic as rules you can read. One rule wins for each slot: - The hero answers the most specific arrival signal it has: campaign parameters, then a search or social referrer, then a return visit, then the time of day. - The call to action changes when scroll depth passes 40%, and otherwise follows a return visit or a campaign. - Notes are reordered after you have actually opened a topic. Unread topics come first. - A slow connection, or Save-Data, drops the hero illustration so the file is never requested. A narrow viewport tightens the layout and still loads the image. The winning rule, and the signal values that triggered it, are written next to the page in “[Why am I seeing this?](https://www.blackboxpersonalization.com/lab/why)”. [Live audience segments](https://www.blackboxpersonalization.com/lab/segments) are the on-or-off form of those same signals, and [web pathing](https://www.blackboxpersonalization.com/lab/pathing) is the order of the pages in the visit. A second control shows the default experience without clearing the profile, so you can compare the two. ## Consent is the gate Turning consent off does not delete the idea of the rules. It stops them from changing what you see, and it stops beacons from being posted. Visit history in local storage is not read while consent is off. The panel says so, and it still lists the rule that would have fired. That is the distinction worth keeping. A signal can be observed, it can be applied to the page, and it can be sent. Those are three decisions. Mixing them is how measurement becomes a black box. ## A costume doesn't fool a fingerprint - [A costume doesn't fool a fingerprint](https://www.blackboxpersonalization.com/blog/a-costume-does-not-fool-a-fingerprint): Clearing cookies is like taking off a name tag at a Halloween party. A browser fingerprint recognizes the walk instead. Here is what a fingerprint reads, why it ignores a Decline, how browsers push back, and how the fingerprint lab shows it without keeping anything. Halloween week is a good time to talk about disguises. A visitor who clears cookies has taken off the name tag. The next page load has no `_ga`, no `bbp.visitor`, nothing to say "you were here before." A fingerprint does not need the name tag. It looks at the costume, the height, and the way you walk, and it can still guess who came back. The [fingerprint lab](https://www.blackboxpersonalization.com/lab/fingerprint) computes the traits a fingerprint could use, right in your tab, and then throws them away. This post explains what those rows mean, why this site refuses to use them, and what browsers now do about it. ## Name tag versus walk A cookie is state. The site writes a value, the browser stores it, and the browser sends it back. The [cookies lab](https://www.blackboxpersonalization.com/lab/cookies) lists every one this site sets. Clear site data and that state is gone. That is the control people expect to have. A fingerprint is stateless. Nothing is written. The site reads things the browser already exposes and combines them into a value that tends to stay the same from visit to visit. The W3C note [Mitigating Browser Fingerprinting in Web Specifications](https://www.w3.org/TR/fingerprinting-guidance/) splits this into two kinds: - **Passive** fingerprinting uses what arrives with every request, with no code on the page: HTTP headers like the User-Agent, the IP address, and other network details. The note says that for some populations the User-Agent and IP address together will often uniquely identify a browser. - **Active** fingerprinting runs JavaScript or CSS in the page to read more: window size, fonts, connected devices, performance, sensors, and how graphics render. The note points out that active techniques are at least potentially detectable on the client, while passive ones are not. So the walk is made of many small traits. None of them is a name. Together they can act like one. ## What the lab reads Open the [fingerprint lab](https://www.blackboxpersonalization.com/lab/fingerprint) and you get seven rows, all computed in the page: 1. **Screen.** `screen.width` by `screen.height`. 2. **Color depth.** `screen.colorDepth`. 3. **Time zone.** From `Intl.DateTimeFormat().resolvedOptions().timeZone`. 4. **Languages.** `navigator.languages`, in your preferred order. 5. **hardwareConcurrency.** The number of logical processors the browser reports. 6. **Font sample.** Seven common faces (Arial, Courier New, Georgia, Times New Roman, Verdana, Trebuchet MS, Comic Sans MS). Each is measured on a canvas against monospace, and a face is listed when its width differs. It is a sample, not your full font list. 7. **Canvas hash.** The page draws a small dark box with the words "open box" and an arc, exports it with `toDataURL()`, and shows the first 16 hex characters of a SHA-256 digest. Different GPUs, drivers, and font rendering can produce slightly different pixels, so the hash can differ between machines that otherwise look alike. Canvas readback as a tracking signal goes back to the 2012 paper [Pixel Perfect](https://hovav.net/ucsd/dist/canvas.pdf) by Keaton Mowery and Hovav Shacham, which showed that text and WebGL rendering differ enough between systems to fingerprint them. None of these rows is secret. The risk comes from combining them. ## How unique is a costume? The classic measurement is the EFF's [How Unique Is Your Web Browser?](https://coveryourtracks.eff.org/static/browser-uniqueness.pdf) by Peter Eckersley. In a sample of 470,161 browsers from privacy-aware volunteers, 83.6% had an instantaneously unique fingerprint. The paper notes the sample was biased toward people who cared about privacy. A larger, later study pulled the number down. [Hiding in the Crowd](https://dl.acm.org/doi/10.1145/3178876.3186097) (Gómez-Boix, Laperdrix, and Baudry, The Web Conference 2018) collected 2,067,942 fingerprints from one of the top 15 French websites and found only 33.6% were unique. The authors also found that non-unique fingerprints are fragile: change a few features and the browser often becomes unique again. They credit the removal of plugins for a big drop in unique desktop machines. That is the honest picture. Plenty of people show up in the same store-bought costume, and a fingerprint cannot always tell them apart. But "sometimes in a crowd" is not the same as "anonymous," and a site that fingerprints does not get to choose which visitors it can single out. ## Why Decline has to mean decline Here is the part that matters for anyone running collection. A visitor who clicks Decline, or clears cookies, is making a choice about being recognized. A fingerprint keeps recognizing them anyway, because nothing was stored for them to clear. The UK regulator said so plainly. When Google announced that from 16 February 2025 its advertising customers would no longer be prohibited from fingerprinting, the ICO [responded](https://ico.org.uk/about-the-ico/media-centre/news-and-blogs/2024/12/our-response-to-google-s-policy-change-on-fingerprinting/) that fingerprinting "is not a fair means of tracking users online because it is likely to reduce people's choice and control over how their information is collected." It added that PECR and data protection law still apply, that businesses must give users fair choices, including consent where necessary, and that for advertising this "is a high bar to meet." So this site keeps the name tag and skips the walk. The lab's own panel says it: the anonymous id is the `bbp.visitor` cookie, written only when a consented beacon runs and deleted when you clear cookies. The traits in the lab are not written to storage, not pushed to the dataLayer, and not posted to an API. With consent on, the lab sends one `fingerprint_view` event with no parameters, so the page view counts but none of the traits leave your tab. You can check the consent side in the [consent lab](https://www.blackboxpersonalization.com/lab/consent) and the stored keys in the [storage lab](https://www.blackboxpersonalization.com/lab/storage). ## Browsers are handing out masks Browser makers are not waiting for every site to behave. - **Safari.** WebKit's [Safari 26.0 features post](https://webkit.org/blog/17333/webkit-features-in-safari-26-0/) says Safari now prevents known fingerprinting scripts from reliably reading APIs that reveal device characteristics, including screen dimensions, hardware concurrency, speech synthesis voices, web audio readback, and 2D canvas. It also stops those scripts from setting long-lived script-written storage and from reading query parameters and `document.referrer`. Note the word "known": the rules target scripts Safari classifies as fingerprinting, not every page. - **Firefox.** Mozilla describes its [expanded fingerprinting protections](https://blog.mozilla.org/en/firefox/fingerprinting-protections/), which cover more APIs in Private Browsing and in Enhanced Tracking Protection's Strict mode. The [W3C note](https://www.w3.org/TR/fingerprinting-guidance/) names two of the goals in holiday terms, if you squint: a "decreased fingerprinting surface" means fewer costume details to notice, and an "increased anonymity set" means more people in the same costume. ## Try it 1. Open the [fingerprint lab](https://www.blackboxpersonalization.com/lab/fingerprint) in Chrome and write down the canvas hash and font sample. 2. Clear site data and reload. The cookies are gone, and the [cookies lab](https://www.blackboxpersonalization.com/lab/cookies) shows it. The fingerprint rows come back the same, because nothing about the walk changed. 3. Open the lab in a private Firefox window or in Safari and compare. Whether the values change depends on the browser, its settings, and whether it treats the script as a fingerprinter, so treat any difference as a hint, not a measurement. 4. Resize the window or change your language list and reload. The screen and languages rows move, which is the fragility the 2018 study describes. The takeaway for anyone building collection: the browser already tells you plenty, and you do not need a disguise-proof id to understand visitors. Ask for consent, store a first-party id the visitor can clear, and let a cleared name tag mean what the visitor meant by it. Happy Halloween. ## Sources - [Mitigating Browser Fingerprinting in Web Specifications, W3C](https://www.w3.org/TR/fingerprinting-guidance/) - [How Unique Is Your Web Browser?, Peter Eckersley, EFF](https://coveryourtracks.eff.org/static/browser-uniqueness.pdf) - [Hiding in the Crowd: an Analysis of the Effectiveness of Browser Fingerprinting at Large Scale, ACM](https://dl.acm.org/doi/10.1145/3178876.3186097) - [Pixel Perfect: Fingerprinting Canvas in HTML5, Mowery and Shacham](https://hovav.net/ucsd/dist/canvas.pdf) - [Our response to Google's policy change on fingerprinting, ICO](https://ico.org.uk/about-the-ico/media-centre/news-and-blogs/2024/12/our-response-to-google-s-policy-change-on-fingerprinting/) - [WebKit Features in Safari 26.0, WebKit](https://webkit.org/blog/17333/webkit-features-in-safari-26-0/) - [Firefox expands fingerprint protections, Mozilla](https://blog.mozilla.org/en/firefox/fingerprinting-protections/) ## When the Google tag never arrives - [When the Google tag never arrives](https://www.blackboxpersonalization.com/blog/when-the-google-tag-never-arrives): A content blocker can drop the analytics script while the page's own beacon still lands. Here's how to tell the two apart, what the gap does and doesn't measure, and why first-party isn't a loophole. Count page views in your analytics tool, then count the same pages in your server logs. The numbers almost never match. There are plenty of reasons, and one of the most common is simple: some browsers never ran the analytics script at all. The [blocking lab](https://www.blackboxpersonalization.com/lab/blocking) on this site measures that gap for its own visitors. This note explains how it decides whether the Google tag loaded, what it stores, and what the resulting percentage can and can't tell you. ## Two requests, two fates A consented visit to the lab produces two kinds of requests. The Google tag is third-party: its script comes from `googletagmanager.com`, and its hits go to `google-analytics.com`. The lab's own beacon is first-party: a `POST` to `/api/blocking` on this site's origin. A content blocker judges each request on its own. Blockers work from filter lists such as [EasyList and EasyPrivacy](https://easylist.to/), and EasyPrivacy describes its job as removing "web bugs, tracking scripts and information collectors." A typical rule looks like `||example.com^`, which in [Adblock Plus filter syntax](https://help.adblockplus.org/hc/en-us/articles/360062733293-How-to-write-filters) matches a domain and all of its subdomains. A well-known analytics host is an easy, stable target for a rule like that. A path on your own domain is harder to target, because it's different on every site. So the script can be dropped while the page's own beacon still arrives. That difference is what the lab counts. ## Why "is gtag defined?" is the wrong test The obvious check is to ask whether `window.gtag` exists. It almost always does, even when the script is blocked. Google's own [consent mode setup guide](https://developers.google.com/tag-platform/security/guides/consent) has you define the function inline, before the library loads: ```js window.dataLayer = window.dataLayer || []; function gtag(){dataLayer.push(arguments);} ``` That function is a stub. It pushes commands onto an array and waits for `gtag.js` to arrive and process them. If the script never arrives, the stub keeps working, the queue keeps growing, and nothing is sent. The lab points this out in its own words: the inline `gtag` is this site's stub, and it exists whether or not Google's script shows up. ## How the lab actually checks The check has two steps. **First, look at what already happened.** The lab reads Resource Timing entries for the page with `performance.getEntriesByType("resource")` and looks for a `gtag/js` request. If one has a body (`decodedBodySize` above zero) or an HTTP status between 200 and 399, the tag loaded. That evidence is often missing even when nothing was blocked. For cross-origin requests, the browser reports size and timing details as zero unless the other server sends a [`Timing-Allow-Origin`](https://developer.mozilla.org/en-US/docs/Web/HTTP/Reference/Headers/Timing-Allow-Origin) header. [`responseStatus`](https://developer.mozilla.org/en-US/docs/Web/API/PerformanceResourceTiming/responseStatus) is 0 when the CORS check fails, and MDN marks it as not yet available in some widely used browsers. A missing number doesn't prove a block. **Second, ask the network directly.** If the timing data is inconclusive, the lab calls `fetch()` on the tag's URL with `mode: "no-cors"`, `cache: "no-store"`, and a four-second timeout. According to [MDN's `RequestInit` reference](https://developer.mozilla.org/en-US/docs/Web/API/RequestInit), a `no-cors` response is opaque: the page can't read its headers or body, and its status is always 0. The lab doesn't need any of that. It only needs to know whether the promise resolved (some server answered) or rejected before a response (the request never got out). A rejection is the usual sign of a content blocker. It's not the only cause, though. A request that's still waiting after four seconds is aborted and counts as blocked too, and so does a network that filters the domain at the DNS level. That `fetch` does not wait for consent. `probeTag` runs when the lab mounts, and the consent check happens only after it returns. The check skips the first-party beacon and the analytics event when measurement is off. It does not skip the probe. This site also does not inject `gtag.js` until measurement consent is on, so before Accept the Resource Timing list is usually empty and the request to `googletagmanager.com` still goes out. ## Counting the gap without collecting people When measurement consent is on, the page sends one small first-party request with two fields: the consent state and whether the tag loaded. The server then: 1. Refuses the request if it carries `Sec-GPC: 1`, the [Global Privacy Control](https://w3c.github.io/gpc/) header. 2. Refuses it if consent isn't `granted`. 3. Checks for a 12-hour, `httpOnly` cookie. If it's there, the browser was already counted, so it returns the totals without adding to them. 4. Otherwise it increments a `beacon` counter, and a `ga_loaded` counter too if the tag loaded. That's the whole record: two integers. There's no row per visitor, no URL, and no IP address. The blocked share is `(beacon − ga_loaded) / beacon`. The lab also sends a `blocking_view` event to Google Analytics with an `analytics_blocked` flag. Think about which browsers can deliver that event: only the ones where the tag loaded. Google Analytics can't report its own absence, which is exactly why the first-party counter exists. ## What the percentage does and doesn't mean The number on the lab is honest, but narrow: - **It only covers people who said yes.** Visitors who decline measurement, or whose browser sends GPC, are in neither count. If people who run blockers also decline more often, the real share among all visitors could be different, and this page has no way to know. - **It only covers this page.** The lab says so directly: it describes browsers that opened the blocking lab, not every visit to the site. - **It counts browsers per 12 hours.** Not people, and not visits. - **The sample is small.** On a personal site, a few visits can swing the percentage a lot. - **"Blocked" lumps causes together.** A browser extension, a privacy-focused browser, a DNS filter, and a slow connection all land in the same bucket. Read it as "among consenting visitors to this page, this share didn't load the Google tag," and stop there. ## First-party isn't a loophole The tempting conclusion is to route everything through your own domain so blockers can't see it. There are two problems with that. The first is technical. Filter lists can target first-party requests. Adblock Plus syntax has a `$~third-party` option that restricts a rule to first-party requests, and a `ping` type that covers `navigator.sendBeacon()`. Path patterns match on any host. A first-party endpoint survives because no list targets it today, not because no list can. The second matters more. A blocker is often a person's choice about tracking. Consent and GPC still apply no matter whose domain the request goes to. The GPC spec itself says it isn't designed to cover data collection within the same context. This site honors it on that endpoint anyway, because that's what the visitor asked for. A first-party beacon is a better way to measure the people who said yes. It isn't a way around the people who said no. ## Try it 1. Open the [blocking lab](https://www.blackboxpersonalization.com/lab/blocking) with measurement consent on and no blocker. It should say the Google tag is reachable. 2. Turn on a content blocker with a privacy list such as EasyPrivacy, and reload. The status should flip to blocked. In your browser's developer tools, the Network panel shows the `gtag/js` request failing, while the request to `/api/blocking` goes through. 3. Reload again. The counters won't move, because this browser was already counted in the last 12 hours. 4. Then open the [Collection Lab](https://www.blackboxpersonalization.com/lab) and look at its network panel, which groups the same requests into first-party and third-party. *More on the request list and the beacon: [What a beacon actually sends](https://www.blackboxpersonalization.com/blog/what-a-beacon-actually-sends).* ## Sources - [EasyList overview (EasyList and EasyPrivacy filter lists)](https://easylist.to/) - [How to write filters, Adblock Plus Help Center](https://help.adblockplus.org/hc/en-us/articles/360062733293-How-to-write-filters) - [Set up consent mode on websites, Google for Developers](https://developers.google.com/tag-platform/security/guides/consent) - [PerformanceResourceTiming: responseStatus, MDN](https://developer.mozilla.org/en-US/docs/Web/API/PerformanceResourceTiming/responseStatus) - [Timing-Allow-Origin header, MDN](https://developer.mozilla.org/en-US/docs/Web/HTTP/Reference/Headers/Timing-Allow-Origin) - [RequestInit (fetch options), MDN](https://developer.mozilla.org/en-US/docs/Web/API/RequestInit) - [Global Privacy Control specification, W3C](https://w3c.github.io/gpc/) ## Check the cart before the holiday rush - [Check the cart before the holiday rush](https://www.blackboxpersonalization.com/blog/check-the-cart-before-the-rush): Four GA4 ecommerce events on a demo notebook: view_item, add_to_cart, begin_checkout, and purchase. What the payload has to carry, how a transaction id dedupes a replay, and why consent still comes before the holiday rush. The holiday rush multiplies the measurement problems a store already has. The Friday after Thanksgiving is 27 November this year. The four events under a cart are still `view_item`, `add_to_cart`, `begin_checkout`, and `purchase`. Miss a transaction id, a currency, or the item array, and the revenue number is a guess. A rule that says someone bought the notebook is only as trustworthy as that payload. The [ecommerce lab](https://www.blackboxpersonalization.com/lab/ecommerce) is a demo store with one product, a field notebook that is not for sale. Place demo order does not take a payment. ## What each event has to carry Google's [Measure ecommerce](https://developers.google.com/analytics/devguides/collection/ga4/ecommerce) guide puts products on an `items` array for viewing item details, adding an item to a cart, starting checkout, and completing a purchase. The [recommended events reference](https://developers.google.com/analytics/devguides/collection/ga4/reference/events) is the parameter list. `items` is required on all four events. `transaction_id` is required on `purchase`. `value` is marked required for meaningful reporting. `currency` is required when `value` is set, as a three-letter ISO 4217 code, so revenue metrics can be computed accurately. Set `value` to the sum of `price * quantity`, and leave shipping and tax out of that sum. On `purchase` they are separate parameters. Each item needs `item_id` or `item_name`; the guide recommends both. Omitted `quantity` defaults to 1. The array can hold up to 200 items, plus up to 27 custom parameters on an item. On the lab, `value` is that price times quantity. There is no shipping and no tax. ```json { "item_id": "notebook-field", "item_name": "Field notebook", "price": 24, "quantity": 1 } ``` Every step carries `currency` set to `USD` and `value` set to 24. ![The ecommerce lab on the live site, captured 9 November 2026, with measurement declined. view_item for the field notebook shows currency USD, value 24, and item id notebook-field.](https://www.blackboxpersonalization.com/blog-image/check-the-cart-before-the-rush/ecommerce-view.png) ## The transaction id [Minimize duplicate key events with transaction IDs](https://support.google.com/analytics/answer/12313109) calls a transaction ID a unique identifier for each transaction, such as an order confirmation number. Google Analytics then deduplicates purchases registered with the same transaction ID. The page says that is how you deduplicate transactions from the same user, and that the same ID should not be used across different users. Letters, numbers, dashes, and spaces are allowed. Anything that could identify a customer is not. An empty string is a trap. Analytics deduplicates every purchase with `transaction_id=""`, so a blank id collapses unrelated orders. The value also has to change per order. A tag that sends one ID for every transaction undercounts. The events reference says the same thing in a line: `transaction_id` is required on `purchase`, and it "helps you avoid getting duplicate events for a purchase." The lab adds the field only on `purchase`. **Place demo order** sets it to `demo-` plus the current time in base 36, so a second press mints a new id. A real order should keep its order id. A second view of the confirmation page should send that same id or skip the push. Fire `purchase` once and let the documented dedupe cover a replay from the same user. The capture below was taken with measurement declined, so the JSON stayed on the page. ![The ecommerce lab after Place demo order, captured 9 November 2026 with measurement declined. The purchase payload has currency USD, value 24, item notebook-field, and a transaction id starting with demo-.](https://www.blackboxpersonalization.com/blog-image/check-the-cart-before-the-rush/ecommerce-purchase.png) ## Consent decides whether the push happens The JSON is drawn either way. The push waits for consent. `view_item` runs once, after measurement consent is granted. The buttons still change the on-screen step when consent is off. **Begin checkout** stays disabled until the cart holds the notebook. **Place demo order** stays disabled until checkout has started. Consent on pushes the same object to the dataLayer. Accepting later sends that one `view_item`. It does not replay cart or checkout clicks made while consent was off. Those pushes happen only when the click already finds consent granted. [From a push to a Google hit](https://www.blackboxpersonalization.com/blog/from-a-push-to-a-google-hit) follows the array to a request. The [consent lab](https://www.blackboxpersonalization.com/lab/consent) is the gate: the Google tag is not requested until Accept, and advertising storage stays denied after it. ## Check the paste before the weekend The [event quality lab](https://www.blackboxpersonalization.com/lab/event-quality) lints a payload in the browser. The paste is not forwarded. The starter `purchase` puts `currency`, `value`, and `page_location` of `https://shop.example/thanks?email=ada@example.com` under `params`. **Check** warns that `transaction_id` and `items` are missing, and it errors twice on the email: the string contains an address, and the query parameter is named `email`. A `currency` that is not three uppercase letters is an error, and so is an `items` value that is not an array, or a parameter named like an email, a phone, or a card number. Google's [best practices on personally identifiable information](https://support.google.com/analytics/answer/6366371) say no data should be passed to Google that Google could use or recognize as personally identifiable information, including an email in a URL path or query parameter. The starter includes one on purpose. The same check applies Google's [event collection limits](https://support.google.com/analytics/answer/9267744): 40 characters for an event name, 40 for a parameter name, 100 for a parameter value (longer for `page_title`, `page_referrer`, and `page_location`), and 25 parameters on one event. The lab's copy also says a property holds only 500 distinct event names. The help page scopes that 500 to app data streams, per app user. Web streams have no such cap. On a website, 40, 100, and 25 are the limits a paste can trip. The ecommerce lab prints `currency`, `value`, and `items` beside `event`. The checker reads `params`, `event_params`, or a gtag-style array. Paste the lab JSON unchanged and it warns that those fields are missing, because it never looks beside `event`. Move them under `params` before you trust the warnings. ![The event quality lab on the live site, captured 9 November 2026. Check flags the starter purchase for an email in page_location, a missing transaction id, and a missing items array.](https://www.blackboxpersonalization.com/blog-image/check-the-cart-before-the-rush/event-quality.png) ## Why the payload is the personalization A slot that changes because someone bought a notebook is quoting `item_id`. Two transaction ids for one order invent a second buyer. A missing `items` array is revenue with no product. A `value` that includes shipping no longer matches the documented sum. A hit sent before Accept was not granted. The ["Why am I seeing this?" panel](https://www.blackboxpersonalization.com/) on the home page names the rule behind a slot. The [Why lab](https://www.blackboxpersonalization.com/lab/why) documents that control. The label is only as honest as the events under it. ## Try it 1. Open the [ecommerce lab](https://www.blackboxpersonalization.com/lab/ecommerce), decline measurement, and read `view_item`: `notebook-field`, `USD`, and 24, with the payload still on the page. 2. Add the notebook, begin checkout, and place the demo order. `purchase` should add a `transaction_id` that starts with `demo-`. 3. On the [event quality lab](https://www.blackboxpersonalization.com/lab/event-quality), press **Check**, then paste that purchase under `params` with the email removed. 4. On a store you run, repeat the four events against a real order id before 27 November, and reload the confirmation page once. ## Sources - [Measure ecommerce, Google for Developers](https://developers.google.com/analytics/devguides/collection/ga4/ecommerce) - [Recommended events, Google for Developers](https://developers.google.com/analytics/devguides/collection/ga4/reference/events) - [Minimize duplicate key events with transaction IDs, Analytics Help](https://support.google.com/analytics/answer/12313109) - [Event collection limits, Analytics Help](https://support.google.com/analytics/answer/9267744) - [Best practices to avoid sending Personally Identifiable Information, Analytics Help](https://support.google.com/analytics/answer/6366371) ## Measuring LinkedIn Ads Without a Martech Stack - [Measuring LinkedIn Ads Without a Martech Stack](https://www.blackboxpersonalization.com/blog/measuring-linkedin-ads): What the LinkedIn Insight Tag sends from the browser, what the Conversions API adds from your server, and how click IDs, UTMs, attribution windows, and consent fit together using first-party data you already have. Paid media measurement gets sold as a tooling problem. Buy a platform, connect it to every ad network, and the numbers will line up. Most of it is simpler than that. A LinkedIn conversion reaches LinkedIn one of two ways: from the visitor's browser or from your server. Each path carries a few fields you can read yourself. Once you know what those fields are, the browser and your own first-party data cover most of what you need. ![Diagram: browser Insight Tag through a consent gate to LinkedIn, and server Conversions API to LinkedIn](https://www.blackboxpersonalization.com/blog-image/measuring-linkedin-ads/flow.png) ## What the Insight Tag collects in the browser LinkedIn describes the Insight Tag as "a piece of lightweight JavaScript code" that supports conversion tracking, website retargeting audiences, and audience insights ([LinkedIn Insight Tag, Marketing Solutions Help](https://www.linkedin.com/help/lms/answer/a489169)). According to LinkedIn's FAQ, the tag collects "the URL, referrer, IP address, device and browser characteristics (User Agent), and timestamp" of visits. LinkedIn says IP addresses are truncated or hashed, direct identifiers are removed within seven days, and the remaining pseudonymized data is deleted within 180 days ([LinkedIn Insight Tag FAQs](https://www.linkedin.com/help/linkedin/answer/a427660)). The tag also sets cookies and, for traffic from outside the EU, EEA, UK, and Quebec, a first-party pseudonymous identifier called the LinkedIn Ads ID (`li_adsid`). Every tag has a partner ID that ties incoming data to your ad account ([Insight Tag help](https://www.linkedin.com/help/lms/answer/a489169)). If you turn on enhanced matching, the page can also send an email address that is hashed in the browser before it goes to LinkedIn ([FAQs](https://www.linkedin.com/help/linkedin/answer/a427660)). You can check all of this yourself. Open DevTools, go to the Network tab, and filter for `linkedin`. The script loads from `snap.licdn.com/li.lms-analytics/insight.min.js`. In the copy we fetched on October 30, 2026, the collection endpoint was `px.ads.linkedin.com/collect`. That's the beacon: a plain HTTP request with your page URL and the partner ID in it. Our post [What a beacon actually sends](https://www.blackboxpersonalization.com/blog/what-a-beacon-actually-sends) covers how to read requests like this one. ## The Conversions API: the same event, from your server The [LinkedIn Conversions API](https://learn.microsoft.com/en-us/linkedin/marketing/conversions/conversions-faq) lets your server post conversion events to `api.linkedin.com/rest/conversionEvents`. Those can be online events, such as a form submit your backend already logs, or offline ones, such as a lead your CRM later marks qualified. LinkedIn positions CAPI as working "in tandem with your Insight Tag" and deduplicates events that arrive from both sources ([Conversions FAQ](https://learn.microsoft.com/en-us/linkedin/marketing/conversions/conversions-faq); [Deduplication guide](https://learn.microsoft.com/en-us/linkedin/marketing/conversions/deduplication)). CAPI matches events to members through user identifiers. The [Conversions API schema](https://learn.microsoft.com/en-us/linkedin/marketing/integrations/ads-reporting/conversions-api-schema) lists `SHA256_EMAIL`, which must be lowercased, stripped of whitespace, and SHA-256 hashed, and `LINKEDIN_FIRST_PARTY_ADS_TRACKING_UUID`, which is the click ID. LinkedIn's FAQ says reports can take up to 72 hours to fill in fully: up to 24 hours for ingestion and another 48 for reporting. The [server-side lab](https://www.blackboxpersonalization.com/lab/server-side) sends the same test event from the browser and from a first-party endpoint, so you can see what a blocker drops. The [CAPI lab](https://www.blackboxpersonalization.com/lab/capi) walks through the same pattern for Meta's Conversions API: a hashed email, a shared event ID for dedupe, and a dry run. The mechanics carry over to LinkedIn. ## Click IDs and UTMs: two different jobs ![Diagram: anatomy of a landing URL showing UTM parameters versus li_fat_id](https://www.blackboxpersonalization.com/blog-image/measuring-linkedin-ads/anatomy.png) When you enable enhanced conversion tracking in Campaign Manager, LinkedIn adds a click ID, `li_fat_id`, to the end of your landing page URLs. The tag can then use first-party cookies ([Enable first-party cookies](https://www.linkedin.com/help/lms/answer/a423304)). LinkedIn's developer docs call capturing `li_fat_id` from the URL or cookie "critical" for CAPI match rates ([Enabling Click IDs](https://learn.microsoft.com/en-us/linkedin/marketing/conversions/enabling-first-party-cookies)). The practical move is to read the parameter on landing, keep it in a hidden form field or your own session, and store it with the lead. Then your server can send it later. UTM parameters do a different job. You write them, every click on the ad carries the same values, and your own analytics reads them. Google's help center covers how GA4 picks up `utm_source`, `utm_medium`, and `utm_campaign` from custom URLs ([Collect campaign data with custom URLs](https://support.google.com/analytics/answer/10917952)). In short, UTMs tell you which campaign sent a session, and the click ID tells LinkedIn which click it was. Use both. The [UTM builder lab](https://www.blackboxpersonalization.com/lab/utm-builder) checks naming rules before you copy a link, and the [traffic source lab](https://www.blackboxpersonalization.com/lab/source) shows how a visit is classified, with campaign parameters checked first and the referrer second. ## View-through vs click-through ![Diagram: post-click and view-through windows on a timeline](https://www.blackboxpersonalization.com/blog-image/measuring-linkedin-ads/windows.png) LinkedIn records two kinds of conversions. A post-click conversion follows an ad click. A view-through conversion follows a member seeing an ad without clicking it. The Campaign Manager default is a 90-day click and 90-day view window, and you can set either to 1, 7, 30, or 90 days. For some conversion types sent through CAPI or CSV upload, you can also choose 180 or 365 days. For most conversion types LinkedIn uses last-touch attribution and credits the most recent click or view ([LinkedIn conversion window](https://www.linkedin.com/help/lms/answer/a426359)). View-through numbers are LinkedIn's own view, and your analytics can't reproduce them, because no click ever reached your site. Report the two separately. Compare post-click conversions against your UTM-tagged sessions, and treat view-through as a signal to test, not proof. The [attribution lab](https://www.blackboxpersonalization.com/lab/attribution) runs one journey through last click, first click, linear, and other models so you can see how much the choice of model moves the credit. ## Consent comes first None of this should fire before the visitor agrees to it. With a consent management platform (CMP), the clean pattern is to keep the Insight Tag from loading until marketing consent is granted, and to send CAPI events only for leads whose consent you recorded. If you also run Google tags, [Consent Mode](https://developers.google.com/tag-platform/security/guides/consent) tells them which consent states apply, and they adjust their behavior to match. LinkedIn also says the Insight Tag should not be installed on pages that collect or contain sensitive data ([Enable first-party cookies](https://www.linkedin.com/help/lms/answer/a423304)). ![Screenshot of the live Paid and social measurement lab with its consent banner, captured October 30, 2026](https://www.blackboxpersonalization.com/blog-image/measuring-linkedin-ads/lab-ad-measurement-screenshot.png) The [consent lab](https://www.blackboxpersonalization.com/lab/consent) shows what this site sends before Accept, what changes after, and how a Global Privacy Control signal keeps measurement off. ## You already own most of this Look at what the pipeline actually needs. You need a landing URL with UTMs you wrote and a click ID LinkedIn appended. You need a form that stores both. You need a consent record, and a server that can hash an email and send a POST request. All of that lives in the browser and in your own database. A big martech vendor can schedule those pieces for you, but it doesn't add a signal you couldn't capture yourself. The [Paid and social measurement lab](https://www.blackboxpersonalization.com/lab/ad-measurement) puts it together: one taxonomy, UTM rules, aligned tags, and a dry run that sends nothing and stores nothing. Start there, open DevTools, and check what your own site sends. ## How a visit becomes a segment - [How a visit becomes a segment](https://www.blackboxpersonalization.com/blog/how-a-visit-becomes-a-segment): How a visit becomes a segment. Eleven rules run in this browser from the viewport, the clock, stored visits, and the URL. Two of them change the home page call to action when consent is on. The [Segments lab](https://www.blackboxpersonalization.com/lab/segments) runs eleven on/off rules in this browser and shows the values each one just read. **Your segments** updates as you resize, scroll, and open pages. **Consented visitors** counts profiles that hold a snapshot and does not show a visitor key. **[Why am I seeing this?](https://www.blackboxpersonalization.com/lab/segments#segment-why)**, under the cards, names the two segments that can change the home page. The same panel sits beside the hero. The lab-menu entry for the rules on this page opens it from anywhere. Mobile is on when [`window.innerWidth`](https://developer.mozilla.org/en-US/docs/Web/API/Window/innerWidth) is under 760 pixels. Desktop is on at 1100 or wider. From 760 through 1099, both stay off. That property is the layout viewport, scrollbar included, so the rule follows the window rather than the User-Agent. Night owl is on at local hour 22, 23, or 0 through 4. [`Date.prototype.getHours()`](https://developer.mozilla.org/en-US/docs/Web/JavaScript/Reference/Global_Objects/Date/getHours) returns 0–23 in local time. The card's timezone is [`Intl.DateTimeFormat().resolvedOptions().timeZone`](https://developer.mozilla.org/en-US/docs/Web/JavaScript/Reference/Global_Objects/Intl/DateTimeFormat/resolvedOptions), the runtime's time zone. Neither rule waits for consent. Campaign visitor is on when the URL has `utm_source`, `utm_medium`, `utm_campaign`, `utm_term`, `utm_content`, or `utm_id`. [Google's campaign URL guide](https://support.google.com/analytics/answer/10917952?hl=en) documents those names, and others this site does not read. Values are cut at 120 characters. [`document.referrer`](https://developer.mozilla.org/en-US/docs/Web/API/Document/referrer) is classed as direct, search, social, internal, or other and shown on the card. An empty referrer, which is a direct visit such as a bookmark, does not flip the segment. Signed-in member follows the session. It can be on while consent is off. The snapshot is written only while consent is on. New visitor and Returning visitor read `site.visitor.v1` in [`localStorage`](https://developer.mozilla.org/en-US/docs/Web/API/Window/localStorage), storage for this origin that lasts until it is cleared. A flag in [`sessionStorage`](https://developer.mozilla.org/en-US/docs/Web/API/Window/sessionStorage) stops a refresh from counting as another visit. That store ends when the tab closes. The first consented session stores 1. Under 2, with consent on, is New visitor. Two or more is Returning visitor. Consent off does not read the count, so both stay off. Engaged reader turns on after two notes reach 75% scroll, or fit the window: `scrollHeight` within 8 pixels of `innerHeight`. Depth is `scrollY` divided by the scrollable height. Dwell is printed and is not part of the test. This tab counts at once. Older notes load only with consent on. Lab explorer turns on at two paths that are `/lab` or start with `/lab/`, including earlier sessions when consent is on. The pathing count on the card is this tab's sequence in `site.path.v1`. It does not decide the segment. West Coast and Other region call `GET /api/edge-place` only after consent. [Vercel](https://vercel.com/docs/headers/request-headers) documents `x-vercel-ip-country` as a two-character ISO 3166-1 code and `x-vercel-ip-country-region` as the region portion of the ISO 3166-2 code for the public IP. West Coast matches `CA`, `OR`, or `WA`, or those full state names, when the country is blank, `US`, or `USA`. If the region header is missing, the zones `America/Los_Angeles`, `America/Vancouver`, `America/Tijuana`, and `America/Whitehorse` count instead. A region that is present, and is not one of those states, is not replaced by the timezone. Other region is on when some region or a non-UTC timezone is known and West Coast is off. Both stay off while the request is in flight. The headers describe the IP, which is the subject of [Where the site thinks you are](https://www.blackboxpersonalization.com/blog/where-the-site-thinks-you-are). With consent on, a changed set goes to `POST /api/segments` after 400 milliseconds. The body is `consent` of `granted` and `{ id, on }` for every id, including the ones that are off. The request sets `keepalive`, and a `pagehide` listener flushes a change still waiting. [MDN](https://developer.mozilla.org/en-US/docs/Web/API/Window/pagehide_event) says that event is not reliable on mobile, so the timer is the send that runs while the page is open. The route refuses `Sec-GPC: 1`. The [Global Privacy Control specification](https://w3c.github.io/gpc/) requires that header value when the signal is on for the top-level page, exposed in script as `navigator.globalPrivacyControl`. Any other consent value is refused, unknown ids are dropped, and a body over 8,000 characters is rejected. The snapshot is stored on the `bbp.visitor` profile. A missing cookie is set on the response: `HttpOnly`, `SameSite=Lax`, `Path=/`, `Max-Age` 400 days. Script cannot read it. How that differs from a script cookie is [Why a first-party cookie forgets you](https://www.blackboxpersonalization.com/blog/why-a-first-party-cookie-forgets-you). A signed-in session stores the account id beside the flags. The flags are not the history, the URL, or an email. `GET /api/segments` returns the latest snapshot for each of the 2,000 most recently updated profiles, reused for about 30 seconds. Each bar is the share of those profiles with the segment on, rounded to the nearest percent. Those live figures are not repeated here. When the on-set changes, the page also pushes `segment_update`, a comma-separated list of the ids that are on. That list is the GA4 user property in [From a push to a Google hit](https://www.blackboxpersonalization.com/blog/from-a-push-to-a-google-hit). The snapshot stays on this origin. Consent off stores nothing and leaves the page on the default. Two segments change the call to action, and only while consent is on and the comparison control is Personalized. A consented reload starts there. Engaged reader sets the primary action to "Engaged reader · see the rule" and wins ahead of the scroll, returning-visitor, and campaign actions. Night owl sets the second action, otherwise "Read the notes," to "Night owl · see the segment," on top of whichever primary action won. The home page panel lists the segments that are on and the winning rule for the hero, the call to action, the notes, and the layout. **Default** shows the baseline page without clearing storage. **Personalized** puts the rules back. With consent off, the panel says the rules were evaluated and were not applied. The header **Why** button is a separate control: labels beside each slot, kept across pages, stored as `site.why.v1` when consent is on. The [Why page](https://www.blackboxpersonalization.com/lab/why) explains it. The other nine segments are labels. Related rules read the same signals directly, which [Personalization from the same signals](https://www.blackboxpersonalization.com/blog/personalization-from-the-same-signals) walks through, and the [personas lab](https://www.blackboxpersonalization.com/lab/personas) runs them for five fixed visitors and for you. ## Try it 1. Open the [Segments lab](https://www.blackboxpersonalization.com/lab/segments) with consent off. The visit cards stay off. Resize across 760 pixels and across 1100. Mobile and Desktop trade places, and the band between them turns both off. 2. Accept measurement and reload. New visitor or Returning visitor matches the stored count, and this browser joins the snapshot. 3. Open a second path under `/lab`. Lab explorer turns on. The pathing count can move without that. 4. Read two different notes past 75%, or two short notes that fit the window. Reload the home page. The primary action should name Engaged reader. **Default** restores the baseline label, and **Personalized** brings it back. During hours 22, 23, or 0 through 4, the second action names Night owl. 5. Reload the lab with `utm_source` set to `notes`. Campaign visitor turns on. The same parameter can change the home page hero, and the panel names the segment and the rule. ## Sources - [Window.innerWidth, MDN](https://developer.mozilla.org/en-US/docs/Web/API/Window/innerWidth) - [Date.prototype.getHours(), MDN](https://developer.mozilla.org/en-US/docs/Web/JavaScript/Reference/Global_Objects/Date/getHours) - [Intl.DateTimeFormat.prototype.resolvedOptions(), MDN](https://developer.mozilla.org/en-US/docs/Web/JavaScript/Reference/Global_Objects/Intl/DateTimeFormat/resolvedOptions) - [Document.referrer, MDN](https://developer.mozilla.org/en-US/docs/Web/API/Document/referrer) - [Window.localStorage, MDN](https://developer.mozilla.org/en-US/docs/Web/API/Window/localStorage) - [Window.sessionStorage, MDN](https://developer.mozilla.org/en-US/docs/Web/API/Window/sessionStorage) - [Window: pagehide event, MDN](https://developer.mozilla.org/en-US/docs/Web/API/Window/pagehide_event) - [URL builders: Collect campaign data with custom URLs, Google Analytics Help](https://support.google.com/analytics/answer/10917952?hl=en) - [Request headers, Vercel](https://vercel.com/docs/headers/request-headers) - [Global Privacy Control specification, W3C](https://w3c.github.io/gpc/) ## Where the site thinks you are - [Where the site thinks you are](https://www.blackboxpersonalization.com/blog/where-the-site-thinks-you-are): Two ways a place shows up on a request: an approximate city from the network, and a precise point only you can hand over. Here's where each one comes from and how far to trust it. Every request your browser makes carries a rough answer to "where is this person?" You never typed it, and no prompt asked you for it. It comes from the network. This note looks at that answer: how it ends up on a request, what this site keeps of it, why it's so often wrong, and how it differs from the precise point a browser shares only when you say so. The [IP geolocation lab](https://www.blackboxpersonalization.com/lab/location) shows both. Open it in another tab and follow along. ## The city you never typed Your browser doesn't send a location. It sends a request from an IP address. Before that request reaches this site's code, Vercel's edge network looks the address up in a geolocation database and attaches the result as request headers. Vercel documents headers such as `x-vercel-ip-country`, `x-vercel-ip-country-region`, `x-vercel-ip-city`, `x-vercel-ip-latitude`, and `x-vercel-ip-longitude`. These are the edge headers on the enriched record in [What a beacon actually sends](https://www.blackboxpersonalization.com/blog/what-a-beacon-actually-sends). Nothing runs in your browser to produce them, and there's no permission prompt, because the browser isn't involved. That's why, on this site, the only control over them is measurement consent. Decline, and the lab doesn't read a stored point, doesn't ask for GPS, and doesn't load the map. ## What gets kept The lab stores latitude and longitude from those headers to two decimal places. One hundredth of a degree of latitude is about 1.1 km. East to west the distance shrinks as you move away from the equator, so the stored point is a square of roughly a kilometer or less. Keeping more digits wouldn't make it more accurate. The header is already an estimate for a whole block of addresses, so extra decimals would only add false precision. The map stays unloaded until you accept for a related reason. Map tiles are images fetched from a tile server, and each tile request tells that server your IP address and which patch of the world you're looking at. A map is a third-party request in its own right, so it waits behind the same consent switch as everything else. ## Why the city is often wrong A geolocation database maps blocks of IP addresses to where the network registers or routes them, not to where you're sitting. That gap shows up in ordinary ways: - **Mobile data** often leaves the carrier's network through a gateway in another city. - **VPNs and company networks** exit wherever their servers are, which can be another state or country. - **Privacy relays**, such as iCloud Private Relay, deliberately report only a coarse region. - **Fallbacks.** When a database only knows the country, it may return a point near the middle of it. One database's default point for the United States famously landed on a farm in Kansas, which then got blamed for millions of addresses it had nothing to do with. So treat the IP city as a fact about the network, not a fact about the person. The lab says the same thing: it is approximate, it is not a street address, and it can be wrong when the network exits far from the visitor. It's also one reason the home page's personalization rules use time, device, connection, referrer, and reading history rather than the IP city. ## The point only you can hand over The second kind of location is different in every way that matters. It comes from the browser's Geolocation API, `navigator.geolocation.getCurrentPosition()`, and it only works on a secure (HTTPS) page. Calling it makes the browser show its own permission prompt. If you allow it, the page gets a latitude, a longitude, and an accuracy radius in meters, worked out from GPS, Wi-Fi, or cell towers depending on the device. On this site that only starts when you press **Share precise location**. That means two separate locks: measurement consent on the site, then the browser's prompt. Both have to say yes. Even then, latitude and longitude are rounded to three decimal places, and accuracy to the nearest 100 meters, before the point is stored. It stays out of the anonymous city clusters, so a precise point never ends up mixed in with everyone's approximate ones. ## The map doesn't need the database One detail in the lab is easy to miss. The map draws your city even when the beacon can't be stored. That works because the city isn't read back from the database. It comes from the headers on the request that loaded the page. Storage is a separate step that can fail without taking the map down. The reverse holds too. On a local machine there's no edge in front of the site, so those headers are empty, and the record says so instead of inventing a place. Empty is a result. ## Try it Open the [IP geolocation lab](https://www.blackboxpersonalization.com/lab/location) and turn measurement consent on. The map loads and shows the city your request arrived from. Is it where you are? Then change the network. Switch your phone from Wi-Fi to mobile data, or turn a VPN on, and reload. The city often moves even though you didn't. Last, press **Share precise location** and accept the browser prompt. Now there are two points on the screen, answering two different questions. The IP city answers "where is this network?" The precise point answers "where is this device?" Only the second one is really about you, and only you can give it. *More on the request that carries these headers: [What a beacon actually sends](https://www.blackboxpersonalization.com/blog/what-a-beacon-actually-sends).* ## MCP server - [Server card](https://www.blackboxpersonalization.com/server-card): Read-only identity for POST /api/mcp. Protocol revision 2026-07-28. - [MCP manifest](https://www.blackboxpersonalization.com/.well-known/mcp.json): Points at the server card and the catalog. - [MCP catalog](https://www.blackboxpersonalization.com/.well-known/mcp/catalog.json): The site-wide list of that one server. - [Agent document](https://www.blackboxpersonalization.com/.well-known/agents.json): Journey tools for the advanced labs. No external accounts. - [What is MCP?](https://www.blackboxpersonalization.com/lab/mcp): Host, client, and server, and what the protocol is not. - [MCP playground](https://www.blackboxpersonalization.com/lab/mcp-playground): Pick a tool, fill the schema, read the JSON-RPC. - [How a site audit works](https://www.blackboxpersonalization.com/lab/site-audit): The sandbox and address checks. The scanner itself is on Black Box Site Design. Tools: list_labs, get_lab, search_blog, example_report, inspect_datalayer, journey_catalog, journey_next. Writes are refused. ## Español - [Caja negra y Google Analytics](https://www.blackboxpersonalization.com/es): Abriendo la caja negra de Google Analytics. Un sitio gratis que muestra cómo se mide una visita y cómo esas mismas señales pueden cambiar la página. - [Usuarios activos de GA4 en tiempo real](https://www.blackboxpersonalization.com/es/live): GA4 en tiempo real, últimos 30 minutos: usuarios, eventos clave, páginas, ciudades, países y dispositivos. La vista de terceros de las mismas visitas. - [Libro público de pruebas](https://www.blackboxpersonalization.com/es/experiments): Libro público de los experimentos del sitio: hipótesis, reparto, lectura bayesiana, regla de parada y una raíz de Merkle que cualquiera puede comprobar. - [Notas de medición y personalización](https://www.blackboxpersonalization.com/es/blog): Notas de medición y personalización: qué envía un beacon, qué revela un navegador y cómo esas señales cambian una página. Gratis, con el laboratorio al lado. - [Caja negra de Google Analytics](https://www.blackboxpersonalization.com/es/about): La caja negra de Google Analytics, abierta en público. Un recorrido práctico y gratuito de la recolección y la personalización, y un lugar para pedir el próximo laboratorio. - [Escribe sobre los laboratorios](https://www.blackboxpersonalization.com/es/contact): Escribe sobre los laboratorios, una página o una idea para el siguiente. El formulario se guarda en este sitio y no depende del consentimiento de medición. - [Política de privacidad](https://www.blackboxpersonalization.com/es/privacy): Cookies de analítica solo tras aceptar, el texto del chat fuera de la analítica, el historial de rutas en este navegador y el acceso de solo lectura a Google. - [Qué cambió en este sitio](https://www.blackboxpersonalization.com/es/changelog): Qué cambió en este sitio. Notas con fecha de las páginas publicadas. Una sugerencia sigue en borrador hasta que una persona aprueba la petición. - [Estado de producción en vivo](https://www.blackboxpersonalization.com/es/status): Estado de producción en vivo de este sitio: la aplicación, la base de datos y las API internas, cada una marcada como operativa, degradada o caída, con su tiempo de respuesta. - [Lo que revela tu dispositivo](https://www.blackboxpersonalization.com/es/showcase): Lo que revela tu dispositivo sin pedirlo: hardware, ajustes, red y lo que hace esta pestaña, y cómo limitar cada señal. - [Paquetes de medición web](https://www.blackboxpersonalization.com/es/packages): Un destino de la prueba de siguiente paso. Aquí no se vende nada. La prueba sigue apagada hasta su aprobación, y una visita sin consentimiento no se cuenta. - [Laboratorios gratis de Google Analytics](https://www.blackboxpersonalization.com/es/lab): Analítica y personalización gratis en este navegador: la página, las peticiones y el registro que deja un beacon de primera parte en el servidor. Abre cualquier laboratorio. - [Usar el panel Sources de Chrome](https://www.blackboxpersonalization.com/es/lab/sources): Usar el panel Sources de Chrome para pausar una puntuación de práctica, entrar en las funciones auxiliares y ver que la oferta sigue apagada. Gratis en el Chrome que tienes abierto. - [Llamar a una API pública desde la página](https://www.blackboxpersonalization.com/es/lab/calls): Llama a una API pública desde la página que ya tienes abierta. Los scripts de práctica usan endpoints de solo lectura de este sitio y no escriben datos. Gratis para editar. - [Qué hace dataLayer.push](https://www.blackboxpersonalization.com/es/lab/datalayer): Cómo el dataLayer se vuelve un hit: la cola, la etiqueta, Consent Mode y la petición collect. Pruébalo gratis con el consentimiento encendido o apagado y lee el hit decodificado. - [Navegador y lado servidor](https://www.blackboxpersonalization.com/es/lab/server-side): Recolección en el navegador y en el servidor del mismo evento de prueba, desde gtag y desde un endpoint de primera parte, y qué descartan los bloqueadores. Gratis para enviar ambos. - [Cookies, almacenamiento y consentimiento](https://www.blackboxpersonalization.com/es/lab/cookies): Cookies, almacenamiento y consentimiento en este origen: quién posee cada clave, cuánto dura y qué servicios no son cookies. Gratis, y Speed Insights no escribe ninguna. - [Consent Mode y Global Privacy Control](https://www.blackboxpersonalization.com/es/lab/consent): Consent Mode y Global Privacy Control en este sitio: qué se envía antes de Aceptar, qué cambia después y cómo una señal GPC mantiene la medición apagada. - [Lo que podría ver una huella](https://www.blackboxpersonalization.com/es/lab/fingerprint): Lo que podría ver una huella sin cookies: rasgos calculados en esta pestaña y no almacenados. Una lectura gratuita del navegador, que desaparece al cerrar la página. - [Cuando la etiqueta de Google se bloquea](https://www.blackboxpersonalization.com/es/lab/blocking): Cuando la etiqueta de Google se bloquea, esta página lo dice para este navegador y muestra la proporción de visitas al laboratorio cuya etiqueta no cargó. La comprobación se queda aquí. - [Constructor de enlaces de campaña](https://www.blackboxpersonalization.com/es/lab/utm-builder): Construye un enlace de campaña UTM, comprueba las reglas de nombres y mira cómo este sitio y GA4 leen los parámetros. Gratis para copiar la URL cuando los campos pasan. - [Analítica de formularios sin las respuestas](https://www.blackboxpersonalization.com/es/lab/forms): Analítica de formularios sin las respuestas: foco, tiempo, saltos, abandono y correcciones. Los caracteres que escribes no se guardan. Gratis para probar la demostración. - [Core Web Vitals, medidos](https://www.blackboxpersonalization.com/es/lab/performance): Core Web Vitals medidos en esta visita, junto al percentil 75 de los eventos de web vitals almacenados. LCP, INP, CLS y TTFB, gratis para leerlos en la página. - [Activadores, etiquetas y variables](https://www.blackboxpersonalization.com/es/lab/tag-manager): Activadores, etiquetas y variables en un contenedor simulado. Un clic muestra qué se dispararía. No se carga ningún script de Google Tag Manager. Gratis para recorrerlo. - [Lo que GA4 recoge por su cuenta](https://www.blackboxpersonalization.com/es/lab/enhanced-measurement): Lo que GA4 recoge por su cuenta: scroll, clics salientes, descargas, búsqueda del sitio y vídeo. Este sitio envía el scroll por su cuenta. Gratis para comparar las dos vías. - [Eventos de ecommerce de GA4](https://www.blackboxpersonalization.com/es/lab/ecommerce): Eventos de ecommerce de GA4 en una tienda de demostración. Ver, añadir al carrito, empezar el pago y comprar muestran los payloads. No se cobra nada. Gratis, clic a clic. - [Mapa de calor de scroll y clics](https://www.blackboxpersonalization.com/es/lab/heatmap): Mapa de calor de scroll y clics a partir de beacons con consentimiento: profundidad, conteos de clics y una muestra etiquetada cuando el conjunto almacenado es escaso. - [Enlaces de seguimiento entre dominios](https://www.blackboxpersonalization.com/es/lab/cross-domain): Enlaces de seguimiento entre dominios: qué ocurre con un id de visitante, un UTM, un referrer y un parámetro _gl al cruzar una redirección u otro dominio. Gratis, paso a paso. - [Seguimiento SPA de cambios de ruta](https://www.blackboxpersonalization.com/es/lab/spa-tracking): Seguimiento SPA de cambios de ruta: una ruta de cliente puede perder el page_view, y un listener de history envía el siguiente. Gratis para comparar ambos caminos lado a lado. - [Reproducción de sesión con enmascarado](https://www.blackboxpersonalization.com/es/lab/session-replay): Reproducción de sesión con enmascarado: qué graba una reproducción, cómo se ocultan los valores de los campos y un grabador que se queda en esta pestaña. Gratis, y no se sube nada. - [Revisor de calidad de eventos GA4](https://www.blackboxpersonalization.com/es/lab/event-quality): Revisor de calidad de eventos de GA4. Pega un push del dataLayer y verás nombres incorrectos, parámetros ausentes, datos personales y límites de tamaño. El pegado no se reenvía. - [Eventos de app y un webview](https://www.blackboxpersonalization.com/es/lab/app-tracking): Eventos de app y un webview: en qué se diferencia una vista de pantalla de un hit web, y cómo un app instance id llega a la página solo si la app lo pone en la URL. - [Conversiones offline, explicadas](https://www.blackboxpersonalization.com/es/lab/offline-conversions): Conversiones offline, explicadas: une un lead de demostración con un cierre y envíalo con el Measurement Protocol cuando el secreto está configurado. Gratis para recorrerlo. - [Para qué sirven los webhooks](https://www.blackboxpersonalization.com/es/lab/webhooks): Para qué sirven los webhooks: un POST firmado cuando ocurre algo, cómo verificar la firma y un registro en vivo de entregas de eventos de este sitio. Gratis para probar. - [La API de conversiones de Meta](https://www.blackboxpersonalization.com/es/lab/capi): La API de conversiones de Meta: un evento de servidor junto al Pixel, correo y teléfono en SHA-256, un event id compartido y una prueba gratuita que no envía nada a Meta. - [Medición de paid y social](https://www.blackboxpersonalization.com/es/lab/ad-measurement): Medición de paid y social para cada plataforma de anuncios: una taxonomía, reglas de UTM, etiquetas alineadas y una prueba gratuita que no envía ni guarda nada. - [Personalización con almacenamiento local](https://www.blackboxpersonalization.com/es/lab/storage): Personalización con almacenamiento local: las claves que escribe este sitio, qué cambia cada una en la página y cómo editarlas o borrarlas. Gratis para inspeccionar en este navegador. - [El recorrido web de una visita](https://www.blackboxpersonalization.com/es/lab/pathing): El recorrido web de una visita: una línea de tiempo de primera parte de esta sesión y un flujo anónimo de saltos entre páginas. Gratis para verlo cuando el consentimiento está activo. - [Precisión de la geolocalización por IP](https://www.blackboxpersonalization.com/es/lab/location): Ubicación por IP a nivel de ciudad, un punto preciso opcional y un seguimiento con Estándar o Preferir alta precisión. El seguimiento no guarda un recorrido. - [Segmentos de audiencia en vivo](https://www.blackboxpersonalization.com/es/lab/segments): Segmentos de audiencia en vivo a partir de señales que ya están en este navegador, más un conteo anónimo de visitantes con consentimiento. Reglas legibles, sin perfil de terceros. - [Cómo ve el sitio un navegador headless](https://www.blackboxpersonalization.com/es/lab/puppeteer): Cómo ve este sitio un navegador headless: qué es Puppeteer, cómo un script de Node dirige Chrome y los hits de GA4 que capturó una ejecución real. Gratis para leer el rastro. - [Modelos de atribución, explicados](https://www.blackboxpersonalization.com/es/lab/attribution): Modelos de atribución explicados en un solo recorrido: último clic, primer clic, lineal, decaimiento, posición y un efecto de eliminación de Markov. Gratis, lado a lado. - [Análisis de caída en el embudo](https://www.blackboxpersonalization.com/es/lab/funnels): Análisis de caída en el embudo entre visita, laboratorio, informe y lead de los últimos 28 días, con filtros de dispositivo y fuente. Los conteos salen de eventos almacenados. - [Análisis de retención por cohorte](https://www.blackboxpersonalization.com/es/lab/cohorts): Análisis de retención por cohorte, con semanas que empiezan en lunes. Cada celda es la parte de ese grupo con una vista posterior. Una tabla gratuita a partir de las primeras visitas. - [Picos y caídas de tráfico](https://www.blackboxpersonalization.com/es/lab/anomalies): Picos y caídas de tráfico frente a una media móvil y una banda de dos desviaciones estándar. Un día se marca cuando el conteo queda fuera de los siete días previos. - [Calculadora de significancia A/B](https://www.blackboxpersonalization.com/es/lab/significance): Calculadora de significancia para un test A/B: valor p, intervalo de confianza, probabilidad de superar y tamaño de muestra, a partir de los conteos del laboratorio de experimentos. - [Filtrar el tráfico de bots](https://www.blackboxpersonalization.com/es/lab/bots): Filtrar el tráfico de bots: los beacons de rastreadores, headless y alta frecuencia se guardan, y los informes de Entender los omiten. Gratis para revisar las reglas. - [Controles de calidad de la analítica](https://www.blackboxpersonalization.com/es/lab/data-quality): Controles de calidad de la analítica sobre vistas de página, web vitals y envíos de formulario almacenados. Cada control es un campo que este sitio espera. Los ids de visitante no salen. - [Preguntar analítica en castellano](https://www.blackboxpersonalization.com/es/lab/ask): Preguntar analítica en castellano. Las preguntas preparadas se vuelven sentencias SELECT sobre el registro. El emparejado con IA es opcional y el SQL permanece en la página. - [Cómo ven esta página los buscadores](https://www.blackboxpersonalization.com/es/lab/seo): Cómo ven esta página los buscadores: título, canonical, datos estructurados y el HTML antes de JavaScript, para cualquier URL de este sitio. Gratis para obtener y comparar. - [El JSON-LD que publica este sitio](https://www.blackboxpersonalization.com/es/lab/json-ld): El JSON-LD que publica este sitio: Person, Organization y WebSite en cada página, un TechArticle en cada laboratorio y migas de pan en las páginas anidadas. Gratis para inspeccionar. - [Laboratorio gratuito de velocidad](https://www.blackboxpersonalization.com/es/lab/pagespeed): Laboratorio gratuito de velocidad para una página de este sitio: puntuaciones de Lighthouse, datos de campo cuando Chrome los tiene y un historial de puntuaciones guardadas. - [Datos en vivo de Search Console](https://www.blackboxpersonalization.com/es/lab/search-console): Datos en vivo de Google Search Console: clics, impresiones, CTR, posición media, consultas principales y páginas principales, cuando el acceso a la propiedad está configurado. - [Google Trends y sus límites](https://www.blackboxpersonalization.com/es/lab/trends): Google Trends y sus límites: un índice de 0 a 100, muestra, el alfa oficial, pytrends y SerpApi. Un gráfico en vivo cuando el endpoint responde. - [Trends y Search Console](https://www.blackboxpersonalization.com/es/lab/trends-search-console): Trends junto a Search Console: clics, impresiones, CTR y posición. Un hueco es interés al alza sin un rango fuerte en las consultas obtenidas. - [APIs de datos del sitio](https://www.blackboxpersonalization.com/es/lab/data-apis): APIs de datos que llama este sitio: Trends, Search Console, PageSpeed, Wikidata, Open-Meteo y el envío del sitemap. Cada tarjeta abre ese laboratorio. - [Cambios del sitemap y su envío](https://www.blackboxpersonalization.com/es/lab/sitemap): Cambios del sitemap y su envío: una línea de tiempo de ediciones de URL y lastmod, y cómo este sitio avisa a Search Console e IndexNow. Gratis para leer la diferencia. - [Haz que busquen tu sitio](https://www.blackboxpersonalization.com/es/lab/get-found): Haz que busquen tu sitio con indexación, descubrimiento y medición, usando este sitio como ejemplo trabajado. Una lista gratuita de pasos que ya están en producción. - [Insertar una tarjeta de prueba del navegador](https://www.blackboxpersonalization.com/es/lab/embed): Inserta una tarjeta de prueba del navegador en otro sitio. Lee este navegador y enlaza al informe completo del laboratorio. El marco en sí no se indexa. Gratis para copiar. - [Medir las referencias de IA](https://www.blackboxpersonalization.com/es/lab/ai-referrals): Medir las referencias de IA, los user agents de rastreadores y un registro de citas que completas a mano. Una mirada gratuita a cómo este sitio separa esas tres fuentes. - [Qué es un archivo llms.txt](https://www.blackboxpersonalization.com/es/lab/llms-txt): Qué es un archivo llms.txt: una convención emergente. Este sitio genera /llms.txt y /llms-full.txt al construir, desde páginas, laboratorios y notas. - [Cómo funciona una petición API](https://www.blackboxpersonalization.com/es/lab/api): Cómo funciona una petición API, de los métodos y códigos de estado a una búsqueda en vivo de centros de datos en Wikidata. Gratis, con la petición y la respuesta a la vista. - [Clima y hora del día](https://www.blackboxpersonalization.com/es/lab/weather): El clima y la hora del día eligen el titular, el tema y una lectura sugerida a partir de un pronóstico a nivel de ciudad y de la hora local. Gratis, y se muestra la petición a Open-Meteo. - [Quién está cerca, con aviso](https://www.blackboxpersonalization.com/es/lab/nearby): Quién está cerca ahora, solo si aceptó. La ubicación se ajusta a celdas de una milla, caduca en 30 minutos y una celda con menos de tres no se muestra. - [Página según el lugar](https://www.blackboxpersonalization.com/es/lab/geo-page): La misma página cambia con un lugar aproximado: ciudad, hora local, clima y un punto cercano. Un panel nombra cada señal usada. No se guarda nada. - [Dibuja una zona de prueba](https://www.blackboxpersonalization.com/es/lab/geofence): Dibuja una zona en el mapa y arrastra un pin. Entrada, salida y permanencia cambian el contenido solo en este navegador. No se envía ni se guarda nada. - [Radio o tiempo en coche](https://www.blackboxpersonalization.com/es/lab/isochrones): Compara un radio en línea recta con el tiempo en coche para el mismo punto. Esa diferencia explica por qué un círculo falla en el reparto y en los anuncios. - [Qué dice la IP y el GPS](https://www.blackboxpersonalization.com/es/lab/ip-vs-gps): Pon la ciudad de la red junto a un punto del navegador que decidas compartir. El mapa dibuja la diferencia. El punto se queda en esta pestaña y no se guarda. - [La misma página, cinco personas](https://www.blackboxpersonalization.com/es/lab/personas): La misma página, cinco personas, y tú. Cada tarjeta nombra las señales y la regla que cambió el titular, la oferta y las notas. Los ajustes se quedan en este navegador. - [Recomendaciones de la página siguiente](https://www.blackboxpersonalization.com/es/lab/recommendations): Recomendaciones de la página siguiente a partir de transiciones con consentimiento, con los conteos detrás de cada elección. Gratis para ver por qué una página se eligió o quedó por defecto. - [Un test A/B y un bandit](https://www.blackboxpersonalization.com/es/lab/experiments): Un test A/B y un bandit sobre dos titulares. Uno reparte el tráfico a la mitad. El otro dedica más impresiones al titular que convierte. Conteos agregados y gratuitos. - [Personalizar por hora y por día](https://www.blackboxpersonalization.com/es/lab/context): Personalizar por hora y por día: la hora local, el día de la semana y una tabla de festivos eligen el titular. Reglas gratuitas que no necesitan un perfil. - [Fuente de tráfico, referrer descifrado](https://www.blackboxpersonalization.com/es/lab/source): Fuente de tráfico, con el referrer descifrado: primero los parámetros de campaña, luego el host del referrer para búsqueda, social, otro sitio, este sitio o directo. Gratis para leer. - [Puntuación de intención del visitante](https://www.blackboxpersonalization.com/es/lab/intent): Puntuación de intención del visitante. Scroll, permanencia, visitas de regreso y profundidad de páginas se suman. La oferta del informe aparece a partir de 40. Gratis para ver los pesos. - [Avisos de intención de salida](https://www.blackboxpersonalization.com/es/lab/exit-intent): Avisos de intención de salida, bien hechos: una nota breve cuando el puntero sale por arriba, o cuando un teléfono sube el scroll muy rápido. Tope de 12 horas. Gratis para probar. - [Perfilado progresivo, explicado](https://www.blackboxpersonalization.com/es/lab/profiling): Perfilado progresivo, explicado: rol, luego objetivo, luego sector. Las respuestas son datos zero-party y se pueden reiniciar. Una mirada gratuita a cómo cambia el titular. - [Un titular a partir de un segmento](https://www.blackboxpersonalization.com/es/lab/ai-copy): Un titular a partir de un segmento y un rol. Una plantilla legible rellena las líneas. El texto asistido por IA es opcional. El laboratorio muestra las entradas y el resultado. - [Por qué veo esta página](https://www.blackboxpersonalization.com/es/lab/why): ¿Por qué veo esta página? El control de la cabecera etiqueta cada espacio personalizado con la regla y las señales que lo dispararon. Gratis para activarlo en todo el sitio. - [Laboratorio de recorrido agéntico](https://www.blackboxpersonalization.com/es/lab/agentic-journey): Simulación transparente de un recorrido agéntico: percibir, inferir, decidir, actuar y aprender. Reglas y una banda en el navegador. Sin modelo. - [Fricción que la página ve](https://www.blackboxpersonalization.com/es/lab/friction-detector): Clics de rabia, clics muertos, vaivén de scroll y un campo atascado en un widget. Las reglas intervienen con ayuda, una versión simple o un atajo. - [Un agente responde en JSON](https://www.blackboxpersonalization.com/es/lab/agent-handoff): Un asistente simulado pide una guía. Reglas de palabras devuelven JSON de schema.org y una página real. Sin modelo y sin guardar nada. - [Consentimiento del agente](https://www.blackboxpersonalization.com/es/lab/privacy-negotiator): Qué haría y no haría un agente en cada nivel de consentimiento. El trueque de ubicación es reversible. Las inferencias sensibles siguen apagadas. - [Ruta desde una frase](https://www.blackboxpersonalization.com/es/lab/concierge): Escribe un objetivo. Reglas de palabras arman una ruta de páginas reales, con un motivo en cada parada. Sin modelo y sin guardar la frase. - [Una acción y luego espera](https://www.blackboxpersonalization.com/es/lab/next-best-action): Una acción para esta sesión: una guía, un laboratorio, un empujón de consulta o nada. Un tope de frecuencia y una visita apurada la retienen. - [Bloques por visitante](https://www.blackboxpersonalization.com/es/lab/content-assembly): Los mismos cuatro bloques, en distinto orden para dos personas etiquetadas. El bloque de prueba enlaza un laboratorio y no cita un número. - [Tests adaptativos por segmento](https://www.blackboxpersonalization.com/es/lab/goal-seeking-test): Tráfico simulado para dos segmentos. Las tasas plantadas están en el gráfico. Asignación epsilon-greedy. No es una medición de este sitio. - [Memoria que un viaje guarda](https://www.blackboxpersonalization.com/es/lab/cross-channel-memory): Un recorrido escrito de un anuncio social de pago a este sitio, un correo y un regreso. Las categorías siguen. Los ids y la dirección se olvidan. - [El agente revisa su plan](https://www.blackboxpersonalization.com/es/lab/self-critique): El agente propone una acción. La equidad, las categorías sensibles, un tope de frecuencia y el consentimiento pueden vetarla. Un veto la retiene. - [Eventos de un reproductor](https://www.blackboxpersonalization.com/es/lab/video-tracking): Un clip original de ocho segundos y un embed opcional de YouTube tras el consentimiento. Los hitos y el tiempo visto se quedan en esta página. - [Apps y altavoces](https://www.blackboxpersonalization.com/es/lab/apps-and-voice): Un teléfono simulado, un medidor local de micrófono y enlaces oficiales de apps, altavoces y privacidad. No es asesoramiento legal. - [Tu GA4, solo lectura](https://www.blackboxpersonalization.com/es/lab/your-data): Tu GA4, solo lectura. El recorrido, la atribución y un gemelo usan tus datos si el acceso de Google está puesto. Si no, hay datos de ejemplo. - [Qué es MCP en la práctica](https://www.blackboxpersonalization.com/es/lab/mcp): Qué es MCP en la práctica. Un anfitrión abre un cliente por servidor. Herramientas y recursos viajan como JSON-RPC. Esta página cita la revisión 2026-07-28. - [Cómo un agente lee con MCP](https://www.blackboxpersonalization.com/es/lab/browser-mcp): Cómo un agente lee este sitio con Browser MCP. Las llamadas de ejemplo se quedan en este origen, la pregunta no se guarda y la respuesta enlaza a la página. - [Pide un informe por MCP](https://www.blackboxpersonalization.com/es/lab/analytics-mcp): Pregunta qué laboratorio se vio más con una herramienta de informe por MCP. La tabla es de ejemplo. No hay GA4 en vivo y la pregunta no se guarda. - [Cómo un agente audita etiquetas](https://www.blackboxpersonalization.com/es/lab/tag-audit-mcp): Cómo un agente audita etiquetas aquí. Lee los pushes antes y después de Aceptar, se queda en este origen y no envía un formulario. Es un ejemplo. - [Cómo funciona una auditoría](https://www.blackboxpersonalization.com/es/lab/site-audit): Cómo funciona una auditoría de página. Un navegador aislado, comprobaciones de dirección y un límite. El escáner está en Black Box Site Design. - [Borradores del próximo arreglo](https://www.blackboxpersonalization.com/es/lab/self-improving): Borradores para el próximo arreglo del sitio. La evidencia son las páginas de abandono y las respuestas no encontradas. Una persona aprueba. - [Servidor MCP de solo lectura](https://www.blackboxpersonalization.com/es/lab/mcp-server): Servidor MCP de solo lectura. inspect_datalayer lee la sesión de este navegador. La pregunta no se registra. - [Prueba una llamada MCP](https://www.blackboxpersonalization.com/es/lab/mcp-playground): Prueba una llamada MCP. inspect_datalayer lee la sesión de este navegador. El texto escrito no se registra. - [Cómo construir un chatbot](https://www.blackboxpersonalization.com/es/lab/chatbot): Cómo construir un chatbot del sitio que responde con tus laboratorios y notas, enlaza el origen, espera al consentimiento y no guarda la pregunta. - [Arquitectura avanzada](https://www.blackboxpersonalization.com/es/lab/advanced): Dieciséis laboratorios de arquitectura para probar en este navegador: etiquetado, consentimiento, modelos y pruebas. - [Etiquetado en servidor](https://www.blackboxpersonalization.com/es/lab/advanced/server-side-tagging): Un evento pasa del navegador a una ruta /collect de este origen. La ruta redacta y no reenvía a GA4. - [Máquina del consentimiento](https://www.blackboxpersonalization.com/es/lab/advanced/consent-time-machine): Una visita escrita bajo Aceptar, Rechazar y consentimiento parcial. Compara cookies, claves y hits. - [Personalización local](https://www.blackboxpersonalization.com/es/lab/advanced/on-device-personalization): Un modelo pequeño puntúa el paso siguiente desde el recorrido que pulsas. Las puntuaciones se quedan aquí. - [Contratos del dataLayer](https://www.blackboxpersonalization.com/es/lab/advanced/datalayer-contracts): Cada evento tiene un contrato JSON. El panel marca en rojo lo que falta, el tipo malo y la clave prohibida. - [Test A/B en el edge](https://www.blackboxpersonalization.com/es/lab/advanced/edge-ab-testing): Un reparto en el edge, conteos de ejemplo y una probabilidad bayesiana. Los conteos no son tráfico en vivo. - [API para agentes](https://www.blackboxpersonalization.com/es/lab/advanced/agent-api): agents.json y las herramientas de recorrido en /api/mcp, junto al recorrido de una persona. - [Vitales y conducta](https://www.blackboxpersonalization.com/es/lab/advanced/vitals-vs-behavior): Vitales de ejemplo unidos a una interacción de ejemplo. No son los datos de campo de este sitio. - [Grafo del recorrido](https://www.blackboxpersonalization.com/es/lab/advanced/live-event-graph): Un mapa D3 de esta visita. Las páginas son nodos y los eventos son aristas. - [Privacidad diferencial](https://www.blackboxpersonalization.com/es/lab/advanced/differential-privacy): Un control de epsilon cambia un conteo de ejemplo con ruido de Laplace. - [Aprendizaje federado](https://www.blackboxpersonalization.com/es/lab/advanced/federated-learning): Navegadores simulados dan un paso local. La página promedia los pesos y dibuja la pérdida. - [Atribución causal](https://www.blackboxpersonalization.com/es/lab/advanced/causal-attribution): Un modelo de Markov sobre recorridos de ejemplo. Quita una página y mira cómo se mueve el crédito. - [Intención prevista](https://www.blackboxpersonalization.com/es/lab/advanced/predictive-intent): El scroll, la permanencia y la profundidad se vuelven dos puntuaciones. La nota de la página cambia. - [Agente con agente](https://www.blackboxpersonalization.com/es/lab/advanced/agent-to-agent): Un agente visitante y el agente del sitio negocian el paso siguiente. No se llama a un modelo alojado. - [Analítica verificable](https://www.blackboxpersonalization.com/es/lab/advanced/verifiable-analytics): Después de Aceptar, los eventos se hashean en un árbol de Merkle y puedes comprobar la prueba. - [Gemelo digital](https://www.blackboxpersonalization.com/es/lab/advanced/digital-twin): Personas escritas recorren una regla. Compara el embudo previsto antes y después. No son visitantes. - [Viaje en el rastreo](https://www.blackboxpersonalization.com/es/lab/advanced/tracking-time-travel): Una línea de tiempo del dataLayer y del almacenamiento de esta pestaña, con un control para volver atrás. - [Brújula y un punto](https://www.blackboxpersonalization.com/es/lab/compass): Brújula y un punto: lecturas de orientación, un punto en memoria y una distancia en línea recta. No es un navegador de senderismo. - [Inclina el nivel](https://www.blackboxpersonalization.com/es/lab/tilt): Inclina el nivel con el teléfono después de Iniciar, o con el teclado, un arrastre o un ejemplo simulado si el sensor falta o se niega. No se guarda nada. - [Escanear un código QR](https://www.blackboxpersonalization.com/es/lab/qr-scanner): Escanear un código QR en el navegador. La cámara empieza solo al pulsar, no se sube nada y el código de muestra solo dice hola. Si se niega, queda la demo. - [Tráfico de referidos IA](https://www.blackboxpersonalization.com/es/lab/ai-referral-traffic): Tráfico de referidos de IA: hosts de referrer, utm_source y user agents de rastreadores, con gráficos de ejemplo. No es el tráfico de este sitio. - [Laboratorio D3](https://www.blackboxpersonalization.com/es/lab/d3): Laboratorio D3: un árbol de ejemplo, de la fuente al recorrido y a la página. El color es el LCP. No es el tráfico de este sitio. - [Banco de visualización](https://www.blackboxpersonalization.com/es/lab/visualization): Banco de visualización: una tabla de ejemplo, campos calculados y un desglose de canal, página y evento. No es tráfico real. - [Registros de CDN](https://www.blackboxpersonalization.com/es/lab/cdn-logs): Registros de CDN, un ejemplo: borde y origen, HIT, MISS y STALE, estado, TTFB y región. Las filas son inventadas. No es el tráfico de este sitio. ## Files - [llms.txt](https://www.blackboxpersonalization.com/llms.txt): The same public page list, without the guides and note bodies. - [Sitemap](https://www.blackboxpersonalization.com/sitemap.xml): Public English and Spanish URLs, lastmod, and hreflang alternates. Admin, account, preview, and noindex routes are omitted. - [robots.txt](https://www.blackboxpersonalization.com/robots.txt): Allows / for every crawler, and repeats that allow for Googlebot, Bingbot, GPTBot, OAI-SearchBot, ChatGPT-User, PerplexityBot, Perplexity-User, ClaudeBot, Claude-User, Google-Extended, Applebot-Extended, CCBot, Bytespider, Amazonbot, meta-externalagent. Also allows the server card, llms.txt, llms-full.txt, security.txt, the MCP manifest, the agent document, and /api/mcp and /api/status. Disallows /account, /admin, /api, /debug, /preview, /register, /sign-in, /your-data, /white-paper/download, /white-paper/thanks. - [Security contact](https://www.blackboxpersonalization.com/.well-known/security.txt): The mailbox and the expiry date. - [MCP server card](https://www.blackboxpersonalization.com/server-card): Public identity and the read-only JSON-RPC endpoint. Protocol revision 2026-07-28. - [MCP manifest](https://www.blackboxpersonalization.com/.well-known/mcp.json): Points at the server card and the catalog. - [MCP catalog](https://www.blackboxpersonalization.com/.well-known/mcp/catalog.json): The site-wide list of that one server. - [Agent document](https://www.blackboxpersonalization.com/.well-known/agents.json): Journey tools for the advanced labs.