Simulation
Agentic Journey Lab
Agentic journey personalization with dynamic pathing. Five panels run the loop on signals this site already has.
Understanding the labHow to use it and why you'd want to
Why you'd want to
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.
Simulation. Rules in this browser plus an epsilon-greedy bandit. No LLM call. No new paid service. A missing signal stays missing.
1 · Sense
State
consent: deniedsegments: missinglabs: missingstored history: missinglocal time: missingdevice: missingreferrer: missingUTM: missingscroll: missingvisits: missingsigned in: missing
Why am I seeing this? Each field is read from this browser or marked missing. Replay personas substitute a labeled preset. Stored visit history is missing while consent is off. The signed-in field waits for the session request.
2 · Infer intent
Candidate goals
- Learn basics0%
No contributing signal
- Go deep on GA4 and measurement0%
No contributing signal
- Explore personalization0%
No contributing signal
- Get help or a consultation0%
No contributing signal
Why am I seeing this? Confidence is this goal’s points divided by the top goal’s points. A bar at zero had no contributing signal. Segment names stay the English rule names.
3 · Decide path
Next step
Exploring, not the top rule. Next step: Personas.
Why am I seeing this? The highlighted node is one of the three or four real pages for the chosen goal. Ignore, override, or a returning visitor changes the highlight. Motion is skipped when the browser asks for reduced motion.
4 · Act
Applied on this page
Open the labs in the published order.
Short version: consent is off, so the page is not personalized.
Option 1
Personas
Option 2
Segments
Option 3
Weather
Option 4
Why labels
- No personalization beyond consent.
- No sensitive inferences.
- A user override always wins.
Why am I seeing this? Cards reorder and the hero grows only when consent is granted. Otherwise the published order and the short line stay. Override replaces the goal before this step. No health, finance, religion, politics, precise location, or identity is inferred.
5 · Learn
Bandit and decision log
Not stored. Measurement consent is off.
Primary is explore. The dimmer share is exploit.
- Learn basicsno clicks
- Go deep on GA4 and measurementno clicks
- Explore personalizationno clicks
- Get help or a consultationno clicks
Decision log
Inputs: consent:denied, stored-history:missing, segments:missing, labs:missing, time:missing, device:missing, referrer:missing, utm:missing, scroll:missing, visits:missing, signed-in:missing
Action: Explore personalization · Personas · explore
Outcome: suggested
Why am I seeing this? Epsilon is 0.2. A draw under that line explores a random goal. Otherwise the highest reward rate wins, or the rules do if nothing has been clicked. The log stores category tokens, not raw referrer or campaign values. localStorage is written only for this browser, only with consent, and Reset my data clears it.
WalkthroughHow we built itShowHide
Example vendor setup is marked below. eVar and event numbers are examples. This lab does not call Adobe.
The journey lab scores four goals from the signals it is allowed to read, then draws an explore or exploit arm and picks the next page on that path. After Accept it queues lab_interaction, and the lab does not call Adobe.
Score four goals
rankGoals adds points for a new visitor, measurement labs, personalization labs, a signed-in visitor, deep scroll, and the other listed signals. Confidence is each score divided by the highest score.
Draw explore or exploit
pickArm uses epsilon 0.2. Below that it picks a goal at random and marks explore. Otherwise it picks the goal with the highest reward rate, or rules when no arm has been pulled. The same seed repeats the same draw.
Choose the next page
decidePath lets an override win, then an explore draw, then an exploit draw, then the top rule. The chosen href is the first step not yet viewed and not ignored.
Shorten the path when consent is off
applied is true only when consent is granted. A desktop visitor who scrolled at least halfway gets the long copy. writeJourneySnapshot skips localStorage while consent is off.
Queue the outcome, not the scores
useLabView queues action view. Follow, ignore, persona, and override call trackLab with labels such as followed-basics, replay-engineer, and override-ga4. Points and confidence stay on the page.
From the code
Epsilon explore, then the arm with the best reward rate.
t = Math.imul(t ^ (t >>> 15), t | 1)
t ^= t + Math.imul(t ^ (t >>> 7), t | 61)
return ((t ^ (t >>> 14)) >>> 0) / 4294967296
}
}
/** Same seed always draws the same arm. The seed is categorical, not a raw signal value. */
export function drawFor(stats: ArmStats, seed: string) {
return pickArm(stats, mulberry32(hashSeed(seed)))
}
export function pickArm(stats: ArmStats, random: () => number): { goal: GoalId | null; mode: DrawMode } {
if (random() < EPSILON) {
const index = Math.min(GOAL_IDS.length - 1, Math.floor(random() * GOAL_IDS.length))
return { goal: GOAL_IDS[index] ?? "basics", mode: "explore" }
}
const pulls = GOAL_IDS.reduce((sum, id) => sum + stats[id].pulls, 0)
if (pulls === 0) return { goal: null, mode: "rules" }Override, explore, and exploit replace the top rule. The path choice continues below this block.
best = id
}
}
return { goal: best, mode: "exploit" }
}
export function pullArm(stats: ArmStats, goal: GoalId, reward: boolean): ArmStats {
const arm = stats[goal]
return { ...stats, [goal]: { pulls: arm.pulls + 1, rewards: arm.rewards + (reward ? 1 : 0) } }
}
export function decidePath(
input: JourneyInput,
options: { override: GoalId | null; ignored: readonly string[]; draw: { goal: GoalId | null; mode: DrawMode } | null },
): Plan {
const ranked = rankGoals(input)
const top = ranked[0] ?? { id: "basics" as const, confidence: 0, points: 0, signals: [] }
let goal: GoalId = top.points > 0 ? top.id : "basics"Google and Adobe
Google queues lab_interaction after Accept. Adobe is an example of the same fields.
Queued only after Accept. The page queues lab_interaction with action view and label page. Follow, ignore, a persona, and an override queue action run_demo. The label is the outcome and goal, a replay name, or override- plus the goal. Scores are not parameters. This site does not call gtag('set') or Adobe.
Simplified shape of the object queued after Accept. Register any new parameter as an event-scoped custom dimension. This site does not call gtag('set').
dataLayer.push({
event: "lab_interaction",
lab_id: "agentic-journey",
action: "run_demo",
label: "followed-basics",
})Example implementation. This site does not call Adobe. eVar42, prop42, and event42 are example variable numbers. In Launch, a data element reads the value and a rule sends it when the value changes. In the report suite, enable eVar42 with Most Recent allocation and expire it on the hit. A processing rule is unnecessary if the data element maps straight to that eVar.
Example alloy payload. _example is a placeholder schema, not a real tenant.
alloy("sendEvent", {
xdm: {
eventType: "web.webinteraction.linkClicks",
_example: {
journeyGoal: "basics",
},
},
})Example hit. eVar42 is hit-scoped in this example. prop42 dies at the same hit.
s.eVar42 = "basics"
s.prop42 = "basics"
s.events = "event42"
s.linkTrackVars = "eVar42,prop42,events"
s.linkTrackEvents = "event42"
s.tl(true, "o", "lab_interaction")Translation
Each lab field, the GA4 name, and the example Adobe variable. Scope is the part that usually gets lost in the translation.
Goal
GA4. Inside label on run_demo, as in followed-basics or override-ga4. The goal alone is not its own parameter.
Adobe. Example eVar42, expire on the hit.
Scope. Event. A user-scoped goal would freeze one arm onto the next visit.
Draw mode
GA4. Not sent. explore, exploit, rules, and override stay on the log in the page.
Adobe. Example prop42 on the hit.
Scope. Hit. The mode is how this suggestion was chosen, not a visitor trait.
Chosen page
GA4. Not sent. plan.chosen is a path on this site.
Adobe. Example eVar42 if you sent the path. The queued label is the outcome, not the href.
Scope. Hit. Opening the link later is a page view, a different hit.
Confidence
GA4. Not sent.
Adobe. Example: do not write the percentage into prop42.
Scope. A calculated share for this plan. It is not an event metric and not a user property.
Recommended next
Why these picksReading path transitions…