The AI panel market has become a fight measured in hundreds of millions, and every vendor publishes its own ranking. Here is ours, with one rule we apply to ourselves first: every claim carries its source, self-declared numbers are labeled as such, and we say where the others are better. You leave with a clear map, and four questions to ask any vendor, us included.
The prediction tier. Simile (founded by the authors of the Stanford generative agents papers) simulates entire populations and raised over $300 million in five months, at a $2 billion valuation. Aaru sells outcome prediction and has proven it at least once: the New York Democratic primary, called within 371 votes. Entry price: enterprise, on quote.
The synthetic research tier. Evidenza (founded by LinkedIn's former B2B Institute team) claims over 100 clients including Salesforce, and a 95% match with real surveys, self-declared. FishDog advertises 300,000 personas and an accuracy figure its own site reports. YouGov bought Yabble; Qualtrics and Toluna are wiring synthetic into their panels. The shared promise: market research, faster.
The message testing tier. Artificial Societies (Y Combinator) drops content into a simulated society and watches it spread, free then about $40 a month. Useful for a post or a campaign, light by design.
| Tool | Its promise | Its published proof | Known price |
|---|---|---|---|
| Simile | Simulate populations to anticipate behavior | Press-reported demos; method born at Stanford | Not published, enterprise |
| Aaru | Predict outcomes (elections, markets) | A primary called within 371 votes, verified by the press | Not published |
| Evidenza | B2B research without fieldwork | 95% match, self-declared | Market estimate: $50-100k/year |
| FishDog | Self-serve research on 300,000 personas | 92% "audited", reported by its own site | Self-serve tiers |
| Artificial Societies | Test a message in a simulated society | Claimed usage volume | Free, then about $40/month |
| ChatGPT / Claude | An instant opinion on anything | None; independent research documents its flattery | $0 to $20/month |
| Kapari | Put your decision on the test bench before you announce it: who leans in, who pushes back, on what blind spot, and what to adjust | A replayable public exam: 16 famous US decisions, every scorecard published, the worst included; every source opened and checked one by one; deterministic math | Beta on request, free |
Figures as of August 6, 2026, sources at the bottom of the page. "Self-declared" means: stated by the vendor, with no published independent verification.
A comparison that never says where its competitors win is worth nothing, so here it is. Simile's science is real: the second Stanford paper replicated the survey answers of 1,052 actually interviewed people with remarkable fidelity; it is the academic reference of the field. Aaru's proof is real: calling an election within 371 votes is a verifiable feat, and if what you need is a forecast, look there. Evidenza and the panel giants know how to serve a large enterprise: compliance, teams, framework contracts, everything a founder-run beta does not claim to offer today. If your need sits there, this page has at least saved you time.
But notice what those three strengths share: none of them answers the question you ask yourself the night before an announcement. Replicating a population in a lab, calling an election, delivering a sector study: none of it tells you who among your stakeholders will push back on Monday, on which sentence, and what to adjust before then.
The question that was going to rattle you in the boardroom, you have already heard it, cold, in your office. That is the Kapari move: a panel of documented voices reacts to your decision, you see who leans in, who doubts, who pushes back and on what blind spot, while an adjustment still costs one rewrite. Then you walk into the room with the file: the graded verdict, the coalitions, the decision brief, the export ready to project.
Kapari takes the exam it gives your decisions: 16 famous US decisions, submitted without their ending, graded against 48 objections people actually raised at the time, each verified in a dated article. Every scorecard is published, and the worst one, 33% on Airbnb's 2020 layoffs, sits on the same page as the best.
The engine builds its panels on a knowledge base of 33 documented frames (Census Bureau, BLS, Gallup and Ipsos series), every source opened and checked against the primary record before it enters. The day your board asks "where does this come from?", you click, and the source appears.
The verdict and the spread come from a deterministic engine: same settings, same result, replayable in front of a witness. The AI writes the texture of the voices; the math recomputes identically. What recomputes can be defended.
Before you launch, Kapari writes down what this test answers, and what it does not. No other tool in the table bounds its own answer before selling it to you.
Demand the same proof from any tool that wants to touch your decisions: its sources open, its scorecards published, its math replayable.
Predicting an outcome, delivering a study, testing a post and preparing a decision are four different jobs. The right tool is the one whose promise matches the question you are actually asking.
A replayable benchmark, on cases never used for tuning, with the failures displayed. A vendor that only shows its good cases is hiding the variance that matters.
An accuracy percentage cannot be verified; a source can. Ask to click.
A tool that promises prediction carries the burden of proof, and the FTC has already sanctioned an unsubstantiated accuracy claim. Overpromise is a risk, not a guarantee.
Simile and Aaru simulate populations to predict behavior or outcomes. Kapari serves a different object: your decision. A panel of documented voices reacts to it, you read who leans in, who pushes back and on what blind spot, and you adjust before you announce. Kapari does not predict: it publishes its exam, its sources and its math.
No, and no simulated panel should. Kapari prepares a decision: the objections heard before anyone else, the fault lines visible while an adjustment still costs one rewrite. The real field stays the real field, and Kapari is how you avoid walking into it blind.
A chatbot returns one opinion, often flattering, with no sources and no structure. Kapari composes a contrasted panel, computes the verdict with a deterministic engine, opens its sources one by one and delivers a shareable file. The point-by-point detail: Kapari or a chatbot?
Reliable for exploring a range of reactions, not as a numeric stand-in for a poll: the peer-reviewed record is clear, and we summarized it, sources attached, on a dedicated page.
Kapari composes a panel of simulated voices and reads the structure of their reactions, to prepare a decision before it goes public. Not a poll, not a prediction: a test bench, with a method published and graded, even when the grade is bad. It is because you can see the worst scorecard that you can trust the others.
Simile: TechCrunch, July 30, 2026 · Stanford paper (1,052 people): arXiv 2411.10109 · Aaru: Research Live, Redpoint · Evidenza: official site, Mi3 · Artificial Societies: SiliconANGLE · Independent research on synthetic panels: User Evaluation, 2026, MeasuringU · The Kapari exam: the proof page.
Put a real decision on the test bench: the objections before anyone else, a verifiable file on the way out.
Kapari measures its audience with Google Analytics, to know how many people read these pages. Nothing is stored on your device until you accept.