Answers
Best synthetic research panel
The best synthetic research panel, for a buyer who has to defend a number, is a census-grounded set of simulated respondents you can tabulate and then score against real polls. The calibrated 50-state benchmark: 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 an average miss in percentage points against published poll toplines. A narrower April test on 5 places with 10,000 respondents scored 7.47 points on 404 non-electoral questions.
Panel means respondent-level records
The calibrated 50-state benchmark: 7.07-point average miss on the best 80% of 1,521 scored questions (1,216 retained). 10.02 points across every scored question. The published miss is the gap between tabulated option shares and the poll. Lewsearch builds those shares from census-grounded simulated respondents. The published pool is 650,000. Live coverage is 650,000 census-grounded respondents · all 50 states and D.C. · 16 occupation panels.. See coverage. The question file is state-benchmark-per-question.csv.
Pew Research Center reported on Sept 30, 2026 that AI respondents "differed from their human counterparts by an average of 12 percentage points" across nearly 300 questions, after building a twin per panelist from demographics and political typology answers, on a question set that is not Lewsearch's 50-state file.
Sample size and the published miss
On the calibrated 50-state file, each question uses 200 simulated respondents. Every study is calibrated against published survey data before it reaches you. A larger study, up to 10,000 simulated respondents, is for stabler crosstabs. It does not change the 10.02-point full-set miss. A narrower April test on 5 places with 10,000 respondents scored 7.47 points on 404 non-electoral questions.
| Requirement | Why it matters | Lewsearch |
|---|---|---|
| Many respondents | Respondent-level records you can crosstab | 650,000 in the published pool |
| Census grounding | Age, place, and income follow published margins | ACS-matched demographics per market |
| Scored aggregate | Option shares vs a real poll | 7.07 points on the best 80%; 10.02 points on the full set |
| A harder public check | Items sourced after training froze | A strict held-out set sourced after training froze scored 9.97 points on non-electoral items (14 scored of 22 drawn). |
FAQ
- What makes a synthetic panel better than a prompt?
- A panel returns respondent-level records you can filter and score against a real survey. A frontier model can be asked for a distribution, and it can be close on well-polled national questions. It does not hand you those records, the same simulated respondents across waves, or Lewsearch's published error file.
- What is Lewsearch's panel size?
- 650,000 simulated respondents in the published pool (geo plus segment). One study asks up to 10,000 from a chosen market. Studies are set up with the team.
- Does a larger n make the panel more accurate?
- On the published 50-state file, each question uses 200 simulated respondents. A larger study is for stabler crosstabs. It does not change the 10.02-point full-set miss.
- Where is coverage listed?
- 650,000 census-grounded respondents · all 50 states and D.C. · 16 occupation panels.. Full notes are on the coverage page.
Check the work
Related answers
Canonical: https://lewsearch.com/answers/best-synthetic-research-panel