Predict how real people react to your policy, product, or message — with thousands of AI personas.
Every agent has a job, an income, beliefs, and a memory of what you showed them last time. Show them a policy and read how a country reacts. Show them a product at a price and read what a market will pay. You get a decision memo or a market research report — not a guess.
They don’t just vote. They tell you why.
Every agent answers a decision in its own words, and the reasoning is the point — it’s what a percentage can never give you. A sample of what the population said about one decision:
One population. Two questions you can ask it.
The engine underneath is the same: people with lives, reacting to something specific and saying why. What changes is what you hand them, and what comes back.
Put a decision in front of a country.
A policy, a memo, a press release — anything with consequences. Every agent reads it and weighs it against their own life before answering.
Put a product at a price in front of a market.
The same population becomes a consumer panel. It reads what you are selling and what you want to charge, then answers whether it would buy.
The same people, in a different room.
Whatever you put in front of the population, it answers as itself and says why. These are the rooms teams walk it into — a cabinet, a war room, a pricing meeting, a launch review — each ending in something you can hand to someone.
Policy & public affairs
Pre-test a reform, subsidy, or regulation on a country before it is announced. Read the coalitions that form and the objection you will have to answer first.
Crisis comms & the PR war room
Rehearse a hostile news cycle. Attack angles, a coalition-risk map, and a language-distortability scan show how a statement gets reframed — before it goes out.
Product & pricing
Find the number a market will actually pay. Van Westendorp, a demand curve, elasticity, and a revenue-optimal price — computed from the panel, not guessed.
Brand & go-to-market
Concept and ad testing, segmentation, positioning, ranked channels, and a tiered influencer plan — packaged as a launch-ready market research report.
Workforce & org change
Model how a restructuring or a return-to-office mandate lands across departments, and where the resistance concentrates, before you send the all-hands.
Message & campaign testing
Run competing framings of the same decision past group representatives and watch opinion move round by round. Keep the wording that actually shifts people.
Built for the people who can’t afford to guess.
A cabinet weighing a reform, a founder before a build, a fund in diligence, an agency before a shoot — each walks up with a different question and the same fear of finding out too late. The proof line under each card is the published research, not our own scoreboard.
Pre-test the reform before you announce it.
Drop a policy, subsidy, or regulation on a simulated country and read the coalitions that form and the objection you will have to answer first — while it can still be reworded.
Fine-tuned agents predicted individual policy preferences at up to 77% vs ~51% random — Royal Society, 2024
Find out if a market wants it before you build it.
Screen concepts, probe pricing, and stress a product-market-fit claim against 10,000 panelists with incomes and current alternatives. The report is allowed to come back no-go.
Synthetic interviews run at ~$2–60 each vs $80–120 per agency-recruited human respondent
Kill the weak variants in an afternoon.
Test messages and creative across segments and markets, screen five concepts down to two, and keep only the framing that moves opinion — before a media budget is committed.
Replaces $25k–$100k, 2–4-week creative pretests — so you test 10 variants, not 2
A second opinion on the demand slide.
Re-run a founder’s demand and pricing claims against the segment they are targeting. You get the questions to ask in diligence — not a verdict, and never a substitute for the round.
EY reproduced a six-month global study in one day at 90% median correlation across 53 questions
Table-read the script before the shoot.
Pre-test storyboards and scripts against the audience they are meant for, then rehearse the social reaction and the news cycle in the war room to see how a line gets reframed.
One PR firm tested six candidate narratives with 189,756 synthetic responses before going public
Read the room before the premiere.
Run a script, trailer cut, or episode past audience segments and compare versions on reaction and reasoning. Built for choosing between cuts — not for predicting box office.
A synthetic focus group matched live family answers >95% of the time in one published study
Two shared phases, then the run forks.
Building the country and generating the people happens once, whichever question you are asking. After that the pipeline runs on its own. Where a phase relies on a published model, it’s named — the coefficients are inspectable, not hidden behind a prompt.
Build the country
Paste articles, reports, or a handful of facts. Country Studio returns a structured dossier: demographics, sectors, media, fault lines.
Generate the population
50 to 10,000 agents — each with a name, profession, income, beliefs, and a memory that carries between runs.
Drop the decision
A policy, a memo, a press release. Every agent reads it and weighs it against their life.
Individual reactions
Each agent weighs personal impact, exposure, identity, and institutional trust, then lands on support, oppose, neutral, or conditional.
Groups form
Agents with shared interests cluster into stakeholder groups. Business owners, religious leaders, and local influencers surface on their own.
Opinions shift
Representatives persuade; neighbours argue. Echo chambers and consensus emerge from the models rather than being scripted.
The memo
Coalitions, fault lines, the sectors that absorb the shock, and the objection you will hear first.
Research the market
A live web sweep compiles the dossier: price bands, named competitors, perception themes, and a category map. Every claim carries the URL it came from.
Pre-register the hypotheses
A strategy brief states the product-market-fit claim, the draft ICPs, the price hypotheses, the positioning territories, and the risks — before anyone reacts. Written afterwards, they could only be rationalised.
Run the panel
Up to 10,000 consumers answer at a given price: would they buy, would they switch, what is their ceiling, and what do they use today.
Price and segment
Van Westendorp, demand, elasticity, and the revenue-optimal price are computed, not written — no model does the arithmetic. k-means then clusters the panel on its own answers.
The report
Twelve sections, a verdict on every hypothesis the brief pre-registered, and a self-contained HTML file with the charts inlined.
Teach it a country before you ask it anything.
Paste articles, write a handful of facts in your own words, or drop in an image. The studio synthesizes a structured dossier, then injects it into every agent-generation, media, and propaganda prompt downstream. A richer dossier is a more lifelike simulation — and the same tool builds a real country or one you invent.
Demographics & sectors
The population blueprint every agent is drawn from — age, income, profession, and the economic sectors that absorb a shock.
Media outlets
The real publications and their leanings, so simulated coverage reads like the actual press rather than a generic newsroom.
Fault lines & key facts
The tensions, taboos, and specifics of the place — the context an agent needs to react like a local, not an average.
A knowledge graph
Who opposes whom and who is allied with whom — the entity graph that seeds lobbies and powers the risk scans.
Twelve sections, and a verdict on every claim you made first.
This is the deliverable a launch decision is actually made from. Sizing figures are marked estimated rather than measured, and their assumptions are printed next to them — that is the arguable part, which makes it the part worth reading.
Executive summary
Go, conditional-go, pivot, or no-go — with the confidence level stated and the reason it is only that high.
Market sizing
TAM, SAM, and SOM, labelled estimated rather than measured, with the assumptions boxed so you can argue with them.
Competitive landscape
Named competitors with prices, share estimates, strengths, weaknesses, and a threat level each.
Positioning map
Price against perceived premium. A quadrant with nobody in it is either a gap or a warning; the white-space note says which.
Price sensitivity
Van Westendorp. Where the four curves cross bounds the range the panel finds credible.
Demand and revenue
Share of the panel whose ceiling clears each price, with the prices you actually ran marked as stronger evidence than the interpolated line.
Segments
k-means over the panel’s own answers, drawn as a force-directed map. A model names the clusters it did not choose.
Brand and positioning
A recommended territory, an archetype, a promise, reasons to believe, messaging pillars, and what to avoid.
Channels
Ranked lead, support, and test channels, each with a rationale and the segments it reaches.
Influencer strategy
Nano, micro, and mid tiers with follower ranges, creator archetypes, content angles, platform mix, budget split, and guardrails.
Drivers and barriers
Ranked reasons to buy and reasons not to, each carrying verbatims from named panelists.
Hypothesis verdicts
Validated, rejected, or inconclusive on every claim the brief pre-registered, with the evidence and the so-what. Then a phased go-to-market and the questions this study could not settle.
The whole report exports as one self-contained HTML file with the charts inlined — it opens anywhere and prints to PDF. The panel responses, the segments, and the competitor set come out as CSVs.
A real study of a meetup app came back no-go, with four of its five pre-registered hypotheses rejected. It also found that price was never the problem. Because the claims were written before the panel ran, a rejection is a finding rather than a mistake — the study was set up to be able to lose.
Ten thousand people, each with a face and a stake.
Every agent carries a name, a profession, an income, and a memory that carries between runs. Facing a policy they hold a stance. Facing a product the same person is a panelist with a budget, a routine, and something they already use instead. A sample from the current run:
Every agent is a different person.
Each node is one agent with its own name, profession, beliefs, and memory. In a policy run the lines are who reacts to whom when a decision lands; in a market study the same people are the panel, and a second map groups them by how they answered. Four thousand are shown here; the engine scales well past that. Drag to rotate.
The models are published, not proprietary.
Every number in either report traces back to one of these. The first four decide how a country moves; the last four decide what a price is worth. All eight are arithmetic — a language model names things it did not compute. If you disagree with a coefficient, you can change it and run it again.
Multinomial logit
P(k) = exp(Uₖ) / Σⱼ exp(Uⱼ)
How each agent picks support, oppose, neutral, or conditional from calibrated utilities.
Friedkin–Johnsen
xᵢ(t+1) = λᵢ·xᵢ(0) + (1−λᵢ)·Σ wᵢⱼ·xⱼ(t)
Each agent blends stubbornness with neighbour influence until the population settles.
Hegselmann–Krause
xᵢ(t+1) = mean{ xⱼ : |xᵢ−xⱼ| < εᵢ }Agents only listen within their confidence bound — echo chambers emerge on their own.
Leontief input-output
Δx = (I − A)⁻¹ · Δd
A shock in one sector cascades through the ten-sector economy via the inverse matrix.
Van Westendorp
OPP = { p : F_cheap(p) = F_expensive(p) }Four cumulative curves over the panel’s price ceilings. Where they intersect bounds the range the market finds credible.
Price elasticity
ε = (ΔQ / Q) ÷ (ΔP / P)
Read off the demand curve at the tested price. Below −1 the market is price-sensitive; above it, the price is not what is stopping them.
Revenue-optimal price
p* = argmax₍ₚ₎ p · D(p)
Demand at each price times that price. Deterministic arithmetic over the panel — no language model computes it.
k-means segmentation
argmin₍S₎ Σₖ Σ₍x∈Sₖ₎ ‖x − μₖ‖²
Clusters the panel on what it actually answered rather than on demographics you picked in advance.
Why a simulated population tracks a real one.
The idea that people can be simulated well enough to decide from is a research finding, not a pitch. These are the peer-reviewed results the method rests on — figures from the field, cited so you can check them, not our own marketing numbers.
Generative agents of 1,000 people
Agents built from two-hour interviews with a stratified sample of 1,052 Americans answered held-out survey questions at 83–86% of each person’s own two-week test–retest consistency — the practical ceiling. Thin demographic personas managed only 74%.
Silicon sampling & algorithmic fidelity
Conditioning a model on real socio-demographic backstories reproduces the response distributions of distinct human subgroups — the founding result that a population, not just an average, can be simulated.
Agents as economic subjects
Given endowments and preferences, LLM agents qualitatively replicate five classic behavioral-economics experiments — fairness, status-quo bias, social preferences — the basis for reading demand-side behavior off simulated people.
Purchase intent that matches surveys
Eliciting free-text reactions and mapping them to a rating scale — rather than asking for a number outright — reaches about 90% of human test–retest reliability, validated across 57 real product surveys and 9,300 respondents.
Simulation as a poll booster
Augmenting a small 5% human sample with LLM-simulated preferences lifted aggregate estimation from about R²≈30% to ≈75% — the hybrid that makes simulation a supplement to real fieldwork rather than a replacement for it.
Find out before it is real.
Put a policy in front of a country and read the memo. Put a product at a price in front of a market and read the report. Either way you end up holding a document you can hand to someone.