Does it hold up

The category’s first question is whether simulated people track real ones. The peer-reviewed evidence both ways, and how to check a claim yourself.

Pillar

Are synthetic respondents accurate? What the evidence says, both ways

Vendors quote 85 to 95%. The peer-reviewed record is more specific: aggregates often match, variation does not, and it rests on what the persona was built from.

23 Sept 2026 · 9 min
Evidence

The 85 to 95% accuracy claim, unpacked: seven numbers, seven different measurements

Every vendor quotes a number in the same band. Each one measures something different on a different task. Where each figure comes from, and what it can carry.

02 Oct 2026 · 9 min
Evidence

Why AI personas agree with everything, and how to build ones that push back

Sycophancy is how assistant models are trained, not one product’s bug. What the research says, why it flattens findings, and three mechanisms that counter it.

23 Sept 2026 · 7 min
Method

How to validate synthetic research: a concordance method anyone can run

Do not take a vendor’s number. Re-run a study whose human answer you hold, compare distributions not averages, and read the trace where they diverge.

23 Sept 2026 · 8 min

Private beta · by application

Bring one real research question.

Access opens by application. Tell us the question you are trying to answer and who has to be convinced by the result.

Apply for accessAll research notes

A rehearsal, not a replacement, for talking to real customers. Stated in the product and in the terms. Every exported report ends with the line that personas are AI simulations, not real customers.