Dot map of the United States. Hover or tap a respondent to open a short preview.
Test messages, screen concepts, and explore customer perspectives with a panel of 650,000 census-grounded simulated Americans. Compare our results with human research through public benchmarks, graded forecasts, or your own Blind Challenge.
Have a completed study? Your first qualified Blind Challenge is free.
From the public 50-state file: a hit, a typical result, and a miss.
How concerned, if at all, are you personally about the price of food and consumer goods?
UT/Texas Politics Project · Texas · June 2026 · n=1,200 registered voters
| Option | OursLewsearch | PollHuman poll | Miss |
|---|---|---|---|
| Very concerned | 63.2 | 65.0 | 1.8 |
| Somewhat concerned | 26.6 | 27.0 | 0.4 |
| Not too concerned | 4.5 | 6.0 | 1.5 |
| Not at all concerned | 2.8 | 1.0 | 1.8 |
| Don't know | 2.8 | 1.0 | 1.8 |
650,000
Simulated Americans
Built from census microdata, all 50 states and D.C.
1,521
Questions scored against real polls
Every one in a public file
10.02pp MAE
Average miss across the full benchmark
1,521 scored questions · 50 states · 200 simulated respondents per question
7
Forecasts on the public record
3 wins, 1 miss, 3 awaiting the official number
MAE is the average miss in percentage points, not percent accuracy. This retrospective benchmark compares questions with published poll toplines, not independent poll fieldings. It does not validate open-ended interviews or predict accuracy on your next study. The lowest-error 80% (1,216 questions, selected after scoring) averaged 7.07 points →
The Blind Challenge
Bring a human study you already fielded. We run the agreed questions before we see your results, and you compare the two on your own terms.
Your first qualified U.S. consumer Blind Challenge is free. No purchase required.
Trackers and multi-wave studies are scoped as Concierge or Enterprise on the call. How sample boosting works →
Scored after the freeze
NRF back-to-school, 2026
Mean K-12 spend within 1%
Read the scorecard →Pew Research Center · September 30, 2026
“AI polling is not a replacement for rigorously surveying real humans.”
What you can run
Every study comes back with a receipt: what was asked, who answered, and how much weight the result can carry. One study asks up to 10,000 respondents.
PANEL STUDY · TEXAS · n=500 · Q1 OF 3
Would you switch insurers to save $30 a month?
Party D 47 · R 34 · I 43
Age 18-34 52 · 35-54 41 · 55+ 31
RECEIPT · 3 QUESTIONS · 500 RESPONDENTS · TWO MEMOS · PDF
How it works →Multiple-choice questions to up to 10,000 census-grounded respondents. Toplines, crosstabs, and quotes in about 15 minutes.
You send
Questions and an audience in plain English
You get
Toplines, crosstabs, quotes, two analyst memos, PDF
Where it fits
Human fieldwork is the standard of record. A frontier model can estimate a national split in seconds. Lewsearch sits between them: a standing panel you can field in minutes, cut any way, re-ask next month, and check against a public record.
Use human fieldwork when the sample itself is the deliverable: a tracker of record, a regulatory claim, a number a journalist will print. Use Lewsearch before and beside it, to screen options, pre-test a questionnaire, explore segments, and fill cells that are expensive to recruit.
The Lewsearch Report
Each Issue freezes a forecast before the official figure is published, then grades it in public against a rule set in advance. The misses stay up.
| Issue | What we forecast | Our call | Official | Result |
|---|---|---|---|---|
| Issue 07 | Deloitte holiday survey: average planned spend, 2026 Win if $1,509 to $1,591 | Our call$1,589 | OfficialDue mid-October | Awaiting grade |
| Issue 06 | NY Fed household spending growth, August Inside the band; April's 4.8 was closer | Our call4.9 | Official4.7 | Miss |
| Issue 05 | Michigan consumer sentiment, September preliminary Second guess; the first, 55.4, missed | Our call52.1 | Official47.8 | Win |
| Issue 04 | 225 public-opinion questions for 2026–27 | Our call225 frozen | OfficialResolves Jan 8, 2027 | Awaiting grade |
| Issue 03 | Five 2026 Senate races | Our call5 margins | OfficialResolves Nov 3, 2026 | Awaiting grade |
| Issue 02 | Prince George's County data-center pause | Our call94.9% support | OfficialPassed July 7 | Win |
| Issue 01 | NRF back-to-school spend per K-12 household Two frozen runs, both within 1%. Two timing questions were instrument mismatches | Our call$857 / $870 | Official$864 | Win |
For insights teams
Enterprise gives your team a standing synthetic panel of your own market: fit to your customers, calibrated to studies you already ran, run privately, and wired into the tools you already use.
We build the panel to match the people you actually survey: your customers, your segments, your cells, from your CRM, loyalty, or sample files. On one Vermont question, matching the study's age, sex, party, and education cut the miss from 20.4 points to 9.5.
The Vermont comparison →We scope calibration around the studies and trackers your team already trusts, then evaluate the agreed questions in a blind replay and report the miss question by question.
How the blind replay works →We build a self-serve workspace for your team: your audiences, your question library, your brand. Analysts launch studies themselves, and our researchers stay on call.
The platform →Toplines, crosstabs, respondent-level files, verbatims, and client-ready PDFs. A REST API and MCP access put the panel inside your own tools, dashboards, and AI assistants.
The research API →Each Enterprise deployment is a custom build that runs on Lewis, our own survey model, with no external model API calls. A dedicated, physically separate instance can be included.
Enterprise details →When a question needs a full instrument, our team designs it, runs the study, and writes the report, so your team gets an answer as well as a dataset.
Concierge research →How teams start
Blind Challenge
Bring a U.S. consumer study you already ran. We predict it blind and you grade the result.
Pilot
Your questions, your audience, calibrated to your past studies. It rolls straight into the annual engagement.
Enterprise
A standing, calibrated panel of your market, your own workspace, and studies whenever your team needs them.
Questions
Something else? Email jeffrey@lewsearch.com or read the methodology.
A synthetic research company. Lewsearch keeps a panel of 650,000 simulated Americans built from census microdata, each with its own beliefs and a memory of what it has been asked. Research teams use it for panel studies, message tests, focus groups, and digital-twin interviews, and can check every result against a public record of scored questions and graded forecasts.
No. They are simulated respondents built from census microdata, and every report says so on the cover. What we publish is how closely their aggregate answers track real polls, question by question, in a public file.
7.07-point average miss on the best 80% of 1,521 scored questions (1,216 retained). 10.02 points across every scored question. MAE is the average miss in percentage points against published poll toplines, on calibrated panels with 200 simulated respondents per question. A narrower April test on 5 places with 10,000 respondents scored 7.47 points on 404 non-electoral questions. A strict held-out set sourced after the model froze scored 9.97 points on non-electoral items. Every scored question in the 50-state benchmark is in the public file.
See the error, question by question →A frontier model asked for a national split can come close on well-polled questions. Lewsearch adds what a research team needs beyond a topline: respondent-level records you can cut into crosstabs that add up to the total, the same respondents re-asked across waves, audiences fit to your customers and scored on your own past studies, a public record of every scored question and graded forecast, and Enterprise workloads with no external model API calls.
Yes. We build a panel from your customer profile, your CRM or loyalty file, or the samples from your past studies. We calibrate it to studies you already fielded and report the miss question by question before you rely on it. Your team gets its own workspace for those audiences, with API and MCP access, and Enterprise workloads make no external model API calls.
Enterprise →When the sample itself is the deliverable: litigation, regulatory filings, a tracker of record, or a number a journalist will print. Use Lewsearch before and beside that fieldwork: to screen options, pre-test a questionnaire, explore segments, and fill cells that are expensive to recruit.
Pricing is scoped to your audiences and study volume on a 30-minute call with the team. Your first qualified U.S. consumer Blind Challenge is free. Enterprise teams start with an 8-week pilot on their own questions, which rolls into an annual engagement.
Book a call →Thirty minutes with the team. Bring a questionnaire, a study you already ran, or a decision due this quarter.
The Lewsearch Report
New issues, scorecards, methodology updates. Free, no spam.
