AI market research, scored against real polls
Survey 10,000 simulated Americans in minutes.
Lewsearch answers your survey questions with AI respondents built from U.S. census data. No humans are polled. To keep us honest, we run questions real pollsters already asked and publish our error. Average miss: about 7.5 points.
No account. No card. About a minute.
Flash A/B · finished in 00:00:28
Subject A
…“Quick question about your Q4 stack”
Subject B
“We cut Acme's CAC 23%. Worth 10 min?”
50 simulated buyers · benchmarked to real polls · quick gut-check, not a live survey
As featured in
Are these real people? No. Here is why that still works.
What they are
Simulated survey respondents. Each one is built from census data: age, income, region, education, politics. Together they mirror a real population, like a wind tunnel mirrors real weather.
Why use them
A human panel costs thousands of dollars and takes weeks. A simulated panel answers in minutes and a full study costs $19. You can test ten ideas for the price of lunch, then spend real money on the winner.
The catch, stated plainly
Simulated answers are estimates. Every report labels how much to trust each result. If you need polling for court, regulators, or the front page, hire a human panel. We will tell you the same thing on a call.
How do you know it's accurate? Check us.
We publish 7.47% mean absolute error on 418 ex-electoral questions of a 460-question benchmark vs real polls (Pew, Gallup, UT/Texas Politics Project, PPIC). In plain terms: when a real poll says 48%, we typically land within about 7 points.
418 questions, scored in public
We answer questions real pollsters already asked, then compare. Every number is out-of-sample and published.
Read the methodology →Forecasts frozen before the answer exists
We publish predictions of upcoming Pew surveys, timestamped, before Pew fields them. When the real numbers land, we grade ourselves on this site.
See the open forecasts →We called a real vote
July 2026: we polled Prince George's County on its data-center moratorium and published the read before the council vote. The vote matched the read.
Read the report →Don't take our word for it. Download the published benchmark set, 443 scored questions with our prediction and the real poll result in one CSV. Or read a sample report first.
Three steps, about five minutes
Step 1
Pick an audience
178,600 simulated respondents across 11 U.S. markets. Plus 15 professional panels: execs, VCs, doctors, nurses, teachers, engineers, marketers, and more.
Step 2
Ask your questions
A poll, an A/B message test, a focus group, or one-on-one interviews. Paste your copy or upload a PDF.
Step 3
Get the report
Percentages, respondent quotes, and results broken out by age, politics, and income. A PDF you can hand to a client.
What you can run
Message Testing
Test two subject lines or ads head to head. A winner in about a minute.
Test a message →Digital Twins
Describe your customer. We build matched simulated twins. You interview them.
Interview your twins →Lewn
Free chat. Ask one question, see how a crowd of simulated respondents splits, in seconds.
Try Lewn free →Concierge Study
Tell us what you need. We design the study with you and deliver a client-grade PDF in 48 hours. $99 flat.
Start your study, $99 →Who runs this
I'm Jeffrey Mead, founder, Columbus, Ohio. I built the model, the benchmark pipeline, and this site. If you want to poke holes in the accuracy claims before you pay, email me and I'll send you the raw numbers. jeffrey@lewsearch.com
Start at $19. No subscription required.
One-off study
$19
500 respondents, PDF report, no plan needed.
Starter
$149/mo
15 studies a month. For consultants and small teams.
Pro
$699/mo
100 studies a month. For agencies running research weekly.
Solo plan at $39/mo. Enterprise from $5,000/mo. See all pricing →
178,600 simulated respondents.
11 U.S. markets: Columbus, Cleveland, Ashtabula, Ohio, Atlanta, Georgia, Dallas, Texas, California, New York, Virginia. 15 professional panels: from C-level execs, VCs, and doctors to nurses, teachers, engineers, marketers, realtors, and skilled trades.
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
FAQ
Frequently asked questions
- What is Lewsearch?
- A survey platform that uses simulated respondents instead of human panels. You ask questions, a census-grounded AI panel answers, and you get percentages, quotes, and a PDF in minutes.
- Are the answers from real people?
- No. Every answer comes from an AI respondent built from census data, and every report says so on the cover. We prove the panel tracks real opinion by scoring it against published polls and posting the error.
- How accurate is it?
- 7.47% mean absolute error on 418 ex-electoral questions of a 460-question benchmark vs real polls (Pew, Gallup, UT/Texas Politics Project, PPIC). Our strictest test, 22 questions sourced after training froze, scores 9.97% ex-electoral. Both numbers and the per-question data are published on our methodology page.
- When should I not use this?
- When you need a defensible human sample: legal, regulatory, or published journalism polling. Lewsearch is for testing messages, concepts, and survey questions fast, before you spend real money.
- What does it cost?
- A free preview every day with no account. One-off studies are $19. Starter is $149 a month and Pro is $699 a month.
Run one test. Judge it yourself.
Free preview, no account. If the read looks right, your first full study is $19.
No account. No card. About a minute.