Simulatte / How it works
HOW IT WORKS · DECISION ENGINEERING

How we build a population — and how we know it holds up.

IN ONE PARAGRAPH

Simulatte builds synthetic populations — structured, decision-making agents calibrated to a real market — and runs your decision against them before you commit the budget. The agents are synthetic; we raise their fidelity by grounding them in real buyer signal. Validation is deliberately honest: directionally correct and calibration-honest — strong on which way a decision goes and which segment moves, careful about exact magnitude. This is what we mean by Decision Engineering: decision science, made into infrastructure you can run.

Run a decision → See the evidence
THE BUILD

How does Simulatte build a synthetic population?

A synthetic population is a set of structured, decision-making agents — sometimes called digital humans — calibrated to behave like a specific market's buyers. Five steps take you from a question to a traceable answer.

STEP 1
Define the decision and the buyer
Name the thing you're weighing — a price, a pack, a claim, a campaign — and the ICP who judges it. The decision, not a demographic, is the unit of work.
STEP 2
Build coherent agents
Each agent is a cognitive model — seeded purchase memory, identity tensions, and the category heuristics a real buyer reasons with — not a profile or a survey-taker.
STEP 3
Ground them in real buyer signal
Distil the actual voice of the market — reviews, forums, Q&A, your own customer data — into how these buyers decide. Real human signal is prioritised over model priors, always.
STEP 4
Run the actual decision
Put the real stimulus to the population and let the agents respond — what wins, why it wins, and the segment that disagrees.
STEP 5
Traceable predictions
Every result traces back to what an agent said or did — with the population, the sources and the reasoning stamped to it. Nothing is invented; nothing is unattributable.
THE FIDELITY LADDER

What makes one population more faithful than another?

Realness in drives fidelity out. The closer the input sits to a real, specific human, the more the population behaves like one. Simulatte grounds in four modes — and they stack: each layer is added on top of the ones below, and each one raises the confidence badge on the pool.

1
Demographic-Grounded
Census and demographic priors plus the model's world knowledge. No real per-person text. The fast, cheap, directional read — pre-flights and exploration.
2
Web-Grounded
Distilled real public voice — reviews, forums, Q&A and comments, collected within platform terms and stripped of personal information. Decision-grade category studies where real category voice exists.
3
CRM-Enriched
Your own first-party customers — purchase history, segments, tickets, survey verbatims. Highest specificity for that brand, and client-exclusive — never reused for anyone else.
4
Qualia / Depth-Grounded
Real lived experience — voice interviews and first-person qual, the felt texture of the decision. The highest fidelity, for the hardest, most human decisions.

The line we never drop: grounding raises fidelity — it does not make the personas real people. They stay synthetic; the grounding makes them behave more like your market.

GROUNDING, PLAINLY

Where does the “real” in a synthetic population come from?

We distil real buyer voice — the actual places people say why they bought, what stopped them, what they trust, and what they'll pay — into structured behaviour: triggers, objections, trust anchors, price sensitivity. Not scraped noise; signal, tagged.

Real signal, ranked first
Inputs are ranked by closeness to a real human: your first-party data first, then real public human text, and model priors only to fill declared gaps — flagged when used.
Verified, not fabricated
Sources are verified live before they're shown; collection respects platform terms and strips personal information. No invented forums, no invented voices.
Provenance on every pool
Each grounded population ships with its sources, item count, the real-vs-synthetic ratio, and a confidence badge — which drops as a corpus ages, so nothing stale is served at full confidence.
No outcome-seeding
Grounding informs how buyers decide, never which brand wins. A comparator on the shelf is never also a pre-loaded favourite — choice must measure the stimulus, not familiarity.
HOW WE KNOW IT WORKS

How accurate is it — really?

Two different questions, answered separately — because conflating them is how research gets oversold.

QUESTION ONE
Distribution accuracy
Does the population's makeup match a real one? We measure it against national-survey ground truth — Pew and IFIC — and publish every number with its audit repo, calibrated and held-out. See the numbers →
QUESTION TWO
Behavioural fidelity
Does the population decide the way real buyers did? We test with blind, out-of-time back-tests: the real outcome is hash-sealed before a single agent is built, the agent is built from pre-cutoff data only, and the prediction is scored after.

The honest verdict. On those blind back-tests the predictions are directionally correct and calibration-honest: every behavioural driver pointed the right way (6 of 6 — the widely-cited “sign-flip” failure did not reproduce), and the population beat a naive baseline on common churn. It does not out-rank a purpose-built model on raw prediction — but it adds the driver-level why a model can't. Strong on direction and ranking; careful about exact magnitude. We publish the framing, not a flattering headline number.

We're not the first to find that grounding synthetic agents in real human data works. Stanford's generative-agents research showed that agents built from real interviews replicate people's survey answers far more faithfully than demographic prompting alone; open Bayesian modelling work (including PyMC's collaboration with Colgate) has shown calibrated models beating naive baselines on real purchase data. Our contribution is turning that into a decision-grade, auditable service — with the working published, not just claimed.

WHAT'S OURS

If the sources are public, what's actually proprietary?

Not the public data — anyone can open Reddit. Our IP is the judgement and the machinery around it.

THE CURATED SOURCE GRAPH

Knowing which handful of sources, out of thousands, carry real decision-grade signal for a given category — and which are noise. That a specific set of sources holds genuine buyer signal for, say, Indian curl-haircare, and the chatter around them doesn't, is earned, not obvious.

THE DISTILLATION ENGINE

Turning raw reviews and posts into structured, tagged behaviour — triggers, objections, trust anchors, price elasticity — plus the method that stacks the modes, reuses a corpus across a category, decays it honestly, and validates the result.

Put plainly: the public sources aren't ours. Our IP is the ability to find the valid human signal in the noise and distil it into a decision-grade population — plus the method that stacks, reuses and validates it.

YOUR DATA IS SAFE

Could your data ever train a competitor's population? No.

Confidentiality maps cleanly to three layers, and they don't mix.

LAYER 1 · OURS
Our method
The curation and distillation engine is Simulatte's IP. It's what we bring; it isn't derived from any one client.
LAYER 2 · SHARED
Public sources
Collected within platform terms, with personal information stripped. Public category voice is a shared research asset — not any single client's property.
LAYER 3 · YOURS
Your data & results
Your first-party data, and the population and findings built from it, are confidential and never reused for anyone else — least of all a competitor.
COMMON QUESTIONS

The questions a skeptic asks first.

Are Simulatte's personas real people, or made up?
They're synthetic — structured, decision-making agents, not real named individuals. We raise their fidelity by grounding them in real buyer signal (reviews, forums, your own customer data), but grounding raises fidelity, not personhood. The one exception is Qualia, which runs real voice interviews with real people.
How does Simulatte build a synthetic population?
You define the decision and the ICP; we build coherent agents (each a cognitive model of memory, identity and category heuristics), ground them in real buyer signal, run your decision against the population, and return predictions where every result traces back to what an agent said or did.
How accurate is Simulatte?
Two different questions. Distribution accuracy — whether the population's makeup matches a real one — is measured against national-survey ground truth (Pew, IFIC) and published with its audit repo. Behavioural fidelity — whether the population decides the way real buyers did — is tested with blind, out-of-time back-tests; the honest verdict is directionally correct and calibration-honest: strong on which way a decision goes, careful about exact magnitude. It doesn't out-rank a purpose-built model on prediction, but it adds the driver-level reasons a model can't.
What data do you use, and is mine safe?
Three layers. Our method is our IP. Public sources are collected within platform terms with personal information stripped, and are shared across a category. Your first-party data and the results built from it are confidential and never reused for anyone else, least of all a competitor.
What is Decision Engineering?
Decision Engineering applies decision science to commercial decisions the way software engineering applied computer science to building products. The infrastructure is synthetic populations you can run a real decision against — pricing, packaging, a claim, a campaign — before you spend.
Does Simulatte test taste or smell?
No. We test what a shopper sees and decides at the point of choice — pack, claim, price, shelf. We can report a persona's expectation formed from what they see (“‘potato protein’ makes me doubt the taste”), but we never claim to measure real taste, smell or texture.
IN SHORT

The whole method, in five lines.

Simulatte builds synthetic populations — decision-making agents calibrated to a real market. It grounds them in real buyer signal, ranked ahead of model priors, so they decide the way your buyers do. The closer the input is to a real human, the higher the fidelity — that's the four-mode ladder. Validation is honest and auditable: strong on direction, calibration-honest on magnitude, with the working published. And your data stays yours. That's Decision Engineering — the research that runs before the market does.

Bring a decision. Watch the population answer it.

The fastest way to judge the method is to run your own decision against it — no signup for your first runs.

Run a decision → Check our working