How it works

You're not asking one AI. You're polling a crowd.

Ask ChatGPT "what would young women think of this ad?" and you get one model's averaged guess. lewn does something different: it polls a panel of hundreds-to-thousands of distinct, demographically-grounded respondents — each modeled on real census and survey data — and shows you how they actually split. Agreement, disagreement, and all.

Grounded, not generic

Each respondent reflects a real slice of the population — age, gender, politics, income, location — so a poll of “liberals in Columbus” is actually liberals in Columbus.

Checked against real polls

Our underlying model is benchmarked against Pew, Gallup, and academic panels (PPIC, UT/Texas) — ≈7.5% average error across 460 benchmark questions (10.7% on a fully held-out set) vs. real human surveys.

Honest about confidence

Every result tells you whether it's a calibrated panel or a directional read — and when we top up a thin segment with synthetic respondents, we say so.

Why not just ask ChatGPT?

Because one model gives you one answer. Ask it to "pretend to be 100 people" and it still collapses to a single best guess (usually a crude average). lewn samples a real distribution of distinct respondents, so you see the spread — the 18% who'd hate your tagline matter as much as the 60% who'd love it. That spread is the whole point of research, and it's the thing a single chatbot can't give you.

What lewn is — and isn't

lewn is a fast, calibrated gut-check — a way to read a crowd in seconds before you spend on media, fielding, or a build. It's not a replacement for a full human study when the stakes are high. Synthetic respondents are estimates, not guarantees, and we show you the confidence on every result so you always know how much to lean on it.

Want the deep technical proof?

The full methodology — cross-validation tables, the 460-question benchmark across four independent panels, the pre-registered held-out set, and calibration details — lives on Lewsearch, our enterprise research platform.

Read the full methodology on Lewsearch →