A model of the visit.
Some people scan a screen and leave. Some read the proof. Some arrive from an ad already checking the promise. Those differences are the model, not a name swapped into a prompt.
before you spend real traffic
RunPivot Simulations: a new persona simulation compares an original homepage with a variant, using four simulated visitor types and a free trial goal. The simulation runs, then shows a predicted winner, with each visitor type's preference and reasoning, and a button to create a real experiment.
Run simulations on the page you already have.
A live URL, a variant, or a mockup nobody has shipped. Set the goal (a trial, a demo, the click) and let it go. Original, variant, and mockup can share one run. You do not need a live experiment first.
Setup
What are you testing?
Existing pageExperimentVariantURLMockup
Candidates
AOriginal
acme.com/
BVariant 1
Outcome-led headline
CMockup
pricing-first.png
Goal
Start a free trialRequest a demoClick the CTAComplete signup
Metric
Trial signups
Watch different visitors go through it.
Not one average person. A paid click, a skimmer, a researcher, someone coming back. They do not want the same page. Patience, reading depth, and intent change what each of them reaches.
See who stayed, who left, and why.
Every result is marked predicted, so it cannot be mistaken for the test. Skimmers never reached the button on A. Researchers still wanted A. That split is the experiment.
Results
Homepage hero test
40 simulated visits · Goal: Start a free trial
Predicted winner: BPredicted, not measured
Candidate B · journeys
Desktop
Simulated trial starts
A · Original7/20
B · Variant 112/20
By visitor type
- Paid-ad visitorPrefers B · 4 of 5
B repeats the ad's promise, so it stayed past the hero.
Engaged: Hero headline · On A, 3 of 5 left at the hero
- Mobile skimmerPrefers B · 5 of 5
B's trial button is on the first screen. On A it's two scrolls down.
Engaged: Primary button · On A, stopped scrolling before the CTA
- Skeptical researcherPrefers A · 3 of 5
Wanted pricing and proof before a CTA. B asked too early.
Engaged: Pricing and logos · On B, left from the pricing table
- Returning visitorPrefers Tie · 5 of 5
Found the trial button on both. B was one click faster.
Engaged: Primary button · No drop-off on either
Hand the strongest idea to a real experiment.
Simulation isn't the result. It drafts the test, attaches the notes, and leaves the launch with you. The draft keeps the variants, the goal, and what each visitor did. You review it. Real visitors decide.
Simulation to experiment
Simulation
Homepage hero test
- Predicted winnerB
- Best withAds, mobile
- Still unsureResearchers
Homepage hero test
DraftA/B test · acme.com/
VariantsMetricsTargetingTraffic
- AOriginalControl50%
- BVariant 1Outcome-led headline50%
Goal: Start a free trial · Metric: Trial signupsSimulation notes attached
Review & launchGive each simulation a job.
One page, one goal, a review when it's done. Run a simulation when you have a specific question, not a pile of ideas.
Hero test
Does the first screen hold a cold click?
The homepage you have, next to a new first screen. See whether a paid click stays through the fold.
Offer check
Trial, demo, or waitlist: which ask survives?
Same page, different goal. See which ask each visitor type actually reaches.
Ad match
Does the page repeat what the ad promised?
A paid-ad visitor is checking that promise. If the hero says something else, they leave at the top.
Pricing pass
Who leaves when the number shows up?
Some visitors look for the price. Others never scroll that far. You see which type stops, and on which version.
What comes back is a trace.
Which version each type preferred, why, and where they stopped. When they disagree, that is the brief for the experiment.
Where the research fits.
The useful finding in the public work is that a grounded agent beats a demographic sketch, and that simulation is an intervention you can replay, not a story about what is likely. That work is Stanford's generative-agent papers, and Simile, the company built from them.
We follow the distinction.
We do not claim their numbers. The 85% result is agreement with a person's own survey retest, on agents built from long interviews. A page rehearsal earns its own, against live experiments.
Common questions
Questions before you run one
How is a simulation different from an A/B test?
An A/B test measures what real visitors did. A simulation walks the page before that traffic is spent, and the result is marked predicted. Use it to choose what is worth testing. Use the experiment to decide what ships.
Is this a chatbot pretending to be a customer?
No. A language model can read the page. The visit comes from differences in how people browse: some scan and leave, some read the proof, some arrive from an ad already checking the promise. Those differences are the model, not a name swapped into a prompt.
What can I run a simulation on?
A live URL, a variant, or a mockup nobody has shipped. Set one goal, such as a trial, a demo, or a click. You do not need a live experiment first.
Do I need traffic to run a simulation?
No. Simulations do not use your real visitors. That is why they help when traffic is expensive, or when a site does not have enough of it for every idea.
Can I treat the result as a winner?
No. It shows who preferred which version, why, and where they stopped. It is predicted, not measured. When visitor types disagree, that split is the brief for a real experiment.
How does this relate to the Stanford and Simile research?
The public work shows that a grounded agent beats a demographic sketch, and that a simulation is something you can replay, not a story about what is likely. We follow that distinction. We do not claim their numbers. The 85% result is agreement with a person's own survey retest, on agents built from long interviews. A page rehearsal earns its own, against live experiments.
Ready to see the visit?
Start with the page you already have. One goal. Who stayed and who left, then a draft experiment you still launch yourself.