Survey 10,000 simulated Americans in minutes. Now in all 50 states.
Over 650,000 living agents with persistent memory. Published error rate: 7.47% MAE on 404 ex-electoral questions of a 460-question public-poll benchmark.
- 7.47%
- average miss
- 404
- questions scored
- 51
- state panels
Live panel
Which subject line wins?
Standing by
Pick your state
10,000 respondents in every state and DC. Click one.
All 51 panels, with calibration dates →Plus 16 occupation panels
Published proof
What you can run
Pricing
$0
Free. One study on a work email, no card. Once per account.
$595/mo
Starter. 20 studies a month. Surveys and message tests.
$2,900/mo
Pro. 175 studies a month, focus groups, the research API, 5 seats.
From $90,000/yr
Enterprise. A dedicated GPU deployment, a cohort built to your population, uncapped under contract.
A one-time pack of 3 studies is $149. One pack purchase a year, then a subscription.
All pricing →Direct answers
Every number traces to methodology or pricing.
The Lewsearch Report
Join our email list
New issues, scorecards, and methodology updates from Lewsearch. Free, no spam, unsubscribe anytime.
Prefer a walkthrough? Book a 30-min demo
Questions
How accurate is this?
Average miss is ~7.5 points. That is 7.47% mean absolute error across 404 questions real pollsters had already asked, scored against their published results. On a stricter test, 22 questions picked after the model was frozen, the miss is 9.97%. Every question, our answer and the real answer are in one public CSV.
See the error, question by question →What is this?
A survey tool that asks living agents instead of recruiting people. Each respondent is a census-built record with persistent memory, not a one-shot persona. You write questions, the panel answers, and you get percentages, crosstabs, quotes and a PDF in about 15 minutes.
What can I use it for?
Testing which message lands before you spend on media. Screening concepts, names and pricing while you still have ten of them. Running a poll before you field it with humans, to catch bad question wording. Exploring how an audience splits by age, income, region or politics.
Are these real people?
No, and every report says so on the cover. They are AI respondents built from census demographics. The published number is how closely the aggregate tracks real polls.
When should I not use this?
When the sample itself is the deliverable: litigation, regulatory filings, or a number a journalist will print.
What does it cost?
One study is free on a work email, no card. Packs are a one-time buy, 3 studies for $149 or 10 studies for $399, one pack purchase a year. Starter is $595 a month or $5,950 a year for 20 studies a month, surveys and message tests. Pro is $2,900 a month or $29,000 a year for 175 studies a month, and adds focus groups, the research API, and five seats. Enterprise starts at $90,000 a year.
Pricing →