Simulate human decision at scale.
AI market research that simulates how, and why, real people respond to change.
There’s prediction, and there’s simulation.
Prediction reads what already happened, last year's sales and last quarter's search volume, then forecasts what comes next.
But predicting a market is a different thing from moving one.
Moving one means testing what has not happened yet: the launch with no history, the price nobody has seen, the campaign that either lands or becomes a public apology.
True simulation lets you change the future instead of forecasting it.
Craze is the simulation company for the people whose choices decide what happens next.
// The model
A population is a complex system.
We resolve it into something you can query.
01Population
A whole market, moving at once. Every post, every purchase, every switch of allegiance. All of it real, and none of it legible.
02Grounding
We interview the people inside it. Every respondent verified as real, recruited across the population rather than off one panel, paid for quality and screened for junk. Those interviews are what the twins are trained on.
03Simulation
What resolves is a population you can run. Ask once, and every segment answers at the same time, each result tagged with how far to trust it.
// Validation comes first
A twin is only worth as much as the human behind it.
Held-out response distribution ·
Alignment 41%
- Craze population
- Human respondents
- Base model
Animated chart comparing response distributions across several questions. A base model sits well off the human distribution; the Craze-trained population converges onto it, then the question changes and it converges again.
How the population is built
Every population starts with real people. We run verified interviews every week, across ages, incomes and regions, then train on that alongside proprietary behavioral data.
How it is validated
Weekly, against real people. Thousands of evaluations across subpopulations and live brand use cases, and those runs train a confidence model that tags every result with a predicted accuracy.
How it is refreshed
Agents consume the same feeds as the people they represent, so a sound that broke on Tuesday is in the model by the time you ask. New populations, including hard-to-reach ones, on demand.
You can bring your own data
Loyalty records, purchase histories, app telemetry, all opt-in and customer-governed. Your data trains your private model and closes the loop, proving customers behaved as simulated.
// One question, the whole population
Ask once. See how far to trust it.
⏎ run
Simulating…
- Would switch
- ··
- Loyal today, still switchable
- ··
- Loyal and likely to stay
- ··
Read
// Comparable scenarios
Run the same population against every option.
A question becomes a set of scenarios, and the same twins answer all of them, so you see how one population moves as the concept, the price, or the wording changes.
Simulation · 4,200 twins
Which launch will most grow first-time trial?
- Study Hours82%
late-night focus occasion
- Gym Bag38%
pre-workout occasion
- Late Set60%
music and nightlife tie-in
Published in the open
Our homework, in public.
The simulation partner for brands that have to move real people
Built by researchers from
// Start here
Bring your own what-if.
Pick a population, type the question, get the answer.