Answers
What is a synthetic panel?
A synthetic panel is a set of simulated respondents built from census demographics, asked the same survey instrument, and aggregated into option shares the way a survey is. Each simulated respondent keeps a persistent profile and a memory of prior answers, so the same records can be cut into crosstabs. 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.
How a study is assembled
Book a demo. The team sets up the study with you. You bring the question, the response options, and the market. The draw matches that market's census margins. Every study is calibrated against published survey data before it reaches you. 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 file uses 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.
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 best 80% keeps the 1,216 lowest-error questions. The other 305 questions stay in the 10.02-point full-set average. Geographic notes and live coverage (650,000 census-grounded respondents · all 50 states and D.C. · 16 occupation panels.) are on coverage. The question file is state-benchmark-per-question.csv.
What the records let you do
A frontier model asked for a national split can be close on well-polled questions. Lewsearch adds respondent-level records, so a cut by age, place, or party adds up to the total. The same simulated respondents can be re-asked across waves. Audiences can be fit to a client's own customer data and scored on the client's own past studies. Forecasts in the Lewsearch Report are frozen before the official number and graded in public. Enterprise workloads run with no external model API calls.
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.
Worked example: 500 Ohio adults, one question
A 500-respondent Ohio study is 500 distinct rows: age, education, income, and party. The output is a topline, crosstabs, and quotes. If the question matches a scored category on the methodology page, use that category's miss as context. If it does not, treat the read as directional. The report states that the answers are simulated.
For the category definition, the error file, and when to hire humans, see what is synthetic research.
| Object | What you get | What you can score |
|---|---|---|
| Synthetic panel | Respondent-level records, option shares, crosstabs, quotes | Aggregate vs a real poll, in points of MAE |
| Frontier-model prompt | A national split or a distribution, if you ask for one | Score it only if the vendor publishes a file. No persistent record to recontact. |
| Synthetic interviews | Transcripts and themes | Thematic parity, if the vendor publishes the definition |
| Human probability sample | A sample with a written design | Sampling error, when the design supports it |
FAQ
- What is a synthetic panel?
- A synthetic panel is a set of simulated respondents built from census demographics, asked the same survey instrument, and aggregated into option shares the way a survey is. Each simulated respondent keeps a persistent profile and a memory of prior answers, so the same records can be cut into crosstabs. 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.
- How is that different from ChatGPT?
- A frontier model asked for a national split can be close on well-polled questions, and it can be asked for a distribution. A Lewsearch panel returns respondent-level records from a census-grounded pool, so results can be cut into crosstabs and the same simulated respondents can be re-asked across waves. Every scored question is in a public file.
- Are the people real?
- No. The respondents are simulated. Demographics are drawn from census-style margins. The answers are generated by a model. Accuracy is checked by scoring the aggregate against real polls.
- How big is Lewsearch's panel?
- 650,000 simulated respondents in the published pool, all 50 states and D.C. A study samples a market and can ask up to 10,000 simulated respondents. Coverage: 650,000 census-grounded respondents · all 50 states and D.C. · 16 occupation panels..
Check the work
Related answers
Canonical: https://lewsearch.com/answers/what-is-a-synthetic-panel