Sign inCreate account

How a study works.From start to finish.

Test with evidence-backed cohorts before the budget commits.

See what you define, how a study runs on behaviour we recorded, and how every result traces back to the question behind it, including the questions it will refuse.

You defineSteps 1–2

Set the conditions. Simulate what comes after.

Start with the decision and what success means. Choose the audiences who can affect it, then write the questions that test each option. Four question shapes cover most of what people bring.
  • Creative

    “Will this ad stop the scroll?”

    What comes back

    Stop rate · split by cohort · representative comments

  • Positioning

    “Which of these three lines lands, and with whom?”

    What comes back

    Per-variant distribution · pairwise gaps · ties called honestly

  • Risk

    “Who will read this the wrong way?”

    What comes back

    Divergence by cohort · the dissenting quotes

  • Audience

    “Who is engaging with this space?”

    What comes back

    Cohort profiles · observed topics · reach percentiles

Inside a runSteps 3–4

Reactions travel a graph that was recorded, not invented.

The step that separates this from prompting a model with a demographic description: reactions move along follow ties we observed, so a message entering a dense community behaves differently from one entering the periphery. Real graphs work that way.

Each persona loads its evidence pack of observed topics, behavioural distributions and verbatim voice, then answers in rounds. An agent answering in a later round has already seen how its neighbours reacted, the way a message travels through a real feed.

rounds 2…n · opinions carry forwardBrief
message + audience, in plain language
Match
cohorts drawn from the catalogue
Sample
independent draws per cohort
Propagate
opinion moves along recorded ties
Report
split, interval, divergence, quotes
Enforced at every stepNo cohort below 100 real accounts · no individual ever addressable · every share carries its interval
Figure 2 · One study, end to end. Steps three and four repeat: an agent answering in a later round has already seen how its neighbours reacted, the way a message travels through a real feed.

The comparison against what you already use. It decides whether the rest of this page is relevant to you, and nothing in the fast half of the chart replaces a panel.

slower ← turnaround → fasterindicative ← evidential weight → defensibleSurvey paneldays–weeksFocus groupweeksDesk researchdaysSynthetic personasminutesmimiminutes
Figure 1 · Where each method sits. Nothing in the fast half replaces a panel for a claim that has to survive scrutiny. The gap this fills is the weeks of guessing that happen before anyone can justify commissioning one.
MethodStrongest atTurnaroundReach for it when
Survey panelReal humans; defensible for regulated claimsDays–weeksThe number must survive legal or procurement
Focus groupDepth, follow-ups, genuine surpriseWeeksYou do not yet know what to ask
Synthetic personasInstant and cheapMinutesNothing is riding on the answer
mimiDirection and divergence, grounded in recorded behaviourMinutesYou need to choose before you can justify fielding
What returnsStep 5

A distribution with its intervals, never one confident sentence.

Every option with its share and a 95% interval; the split by cohort with its own sample sizes; a divergence score that says when the average is hiding the finding; quotes retrieved from what members wrote.
Would stop scrolling41%Would keep scrolling34%Would comment16%95% interval, not a pointshareBy cohortwhere the headline hides a disagreementUS · beauty commentary58%US · fitness & routine37%Global · food creators22%divergence0.31
Figure 3 · Anatomy of a result. The headline is an average of audiences that may disagree, so the per-cohort split and a divergence score sit beside it. Where two options’ intervals overlap, the report says they are tied rather than ranking noise.

  • Distribution, not a verdict

    Every option with its share and a 95% interval. Overlapping intervals are reported as tied.

  • The split by cohort

    The headline is usually an average of audiences that disagree. Each cohort is shown separately, with its own sample size.

  • Divergence score

    One number for how far apart the cohorts are. High means the average is hiding the finding.

  • Representative quotes

    Drawn from the cohort's indexed voice, so tone is retrieved from what members wrote rather than invented.

  • Stated limits

    Which cohorts were thin, what the evidence does not cover, and the run's own parameters.

Cohorts
387
each ≥100 accounts
Accounts
265K
resolved, not sampled
Reactions
311K
observed, not inferred
Voice samples
9.9K
indexed for tone
What it will not tell youStated plainly

A buyer surprised in month three is worse than one who declines today.

These are construction limits, not small print. Each one is enforced by the platform itself. A study that hits one says so on the result rather than guessing past it.

Absolute forecasts

Direction and relative gaps hold up far better than a predicted percentage. Treat the numbers as a ranking, not a projection.

Audiences outside the catalogue

If a brief matches thin cohorts, the platform says so rather than substituting neighbours.

Anything about an individual

Every cohort is an aggregate of at least 100 accounts, by construction. There is no individual to address.

Reactions to breaking events

Evidence carries a capture latency. A cohort does not know what happened yesterday.

Bring the question you were going to guess at.

Create account The platform underneath