What does census-grounded mean?

Each simulated respondent is built from census-style demographic margins (age, education, income, race, party, place). The answers are generated by a model. The published check is how close the aggregate comes to real polls.

When is a synthetic screen the right tool?

When you need to drop a weak message, catch a leading question, or take a first read across markets before you spend on humans. A traditional focus group takes weeks to recruit. A queued synthetic study returns in about 15 minutes. Book a demo. The team sets up the study with you.

When do I still need a human sample?

When the sample itself is the deliverable: litigation, a regulatory filing, a published journalism poll, or any number that has to be defended as a probability design. Every Lewsearch report states the answers are simulated.

Answers

Is synthetic research a probability sample?

Last updated: October 2, 2026

Synthetic research is not a probability sample. Lewsearch interviews census-grounded simulated respondents. 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. Use it to screen messages and instruments. Hire humans when the number has to stand up in court, at a regulator, or in a newspaper.

What a probability sample is

A probability sample starts with a defined frame of real people. Every unit in that frame has a known, non-zero chance of selection. The statistician can write the design, the weights, and the sampling error. A court, a regulator, or a newspaper can ask who was eligible, who was reached, and how non-response was handled. The number stands or falls on that design.

That is a human research product. It takes time and money because the people have to be found.

Why a synthetic panel is not one

A Lewsearch study draws simulated respondents whose demographics match a market's published margins, interviews them with a model, and tabulates the answers. There is no inclusion probability. There is no non-response. There is no person who could have been selected and was not. The output looks like a survey (option shares, crosstabs, quotes, a PDF). The design is a simulation.

The published check is the miss against real polls, in percentage points. It is an average absolute gap. It is a sampling margin only when a human design produces one. Do not write the miss with a plus-minus sign. 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. A narrower April test on 5 places with 10,000 respondents scored 7.47 points on 404 non-electoral questions. Do not say 7.47 points across 460 questions.

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.

Probability sample versus a synthetic panel
	Probability sample	Lewsearch synthetic panel
Who answers	Recruited people from a defined frame	Census-grounded simulated respondents
What you can defend	The design, if it holds	An AI-labeled directional read
Error you can quote	Sampling error from the design	7.07 points on the best 80% of 1,521 scored questions; 10.02 points on the full set
Where it can be filed	Court, regulator, newspaper, if the design holds	A screen, ahead of a legal sample

What census-grounded means here

Census-grounded means the demographic row is built from census-style margins. Age, education, income, race, party, and place are drawn so a 500-respondent Ohio study is 500 distinct rows. The published pool is over 650,000 simulated respondents. A study samples a market. It does not interview the whole pool. A state on the coverage map has a panel. A published miss for that state is a separate file. Live coverage is on coverage (650,000 census-grounded respondents · all 50 states and D.C. · 16 occupation panels.).

The answers come from a model. Every report says so on the cover. The per-question file for the 50-state miss is on methodology and at state-benchmark-per-question.csv.

When a screen is the right tool

Use the synthetic pass to drop a confused option, lint a leading or double-barreled item, rank messages before media spend, or take a first read across markets. A team can screen several framings before it books a facility group. Queued studies return in about 15 minutes. Every study is calibrated against published survey data before it reaches you. Pricing is scoped on a short call. Book a demo. New studies are set up with the team.

A low miss on a category this study did not ask does not warranty the question in front of you. A high miss on a category you did ask is a reason to hire humans or rewrite the question. The per-question file is public.

When the number has to be human

Hire a probability sample when the result has to stand up in court, at a regulator, or in a newspaper. The same rule covers clinical or highly specialized populations you cannot score, demand forecasts you will take to a board as fact, and any filing that treats the sample itself as the deliverable. Lewsearch's terms say the same thing: results are directional research in those contexts.

The product is a screen.

FAQ

Is synthetic research a probability sample?: Synthetic research is not a probability sample. Lewsearch interviews census-grounded simulated respondents. 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. Use it to screen messages and instruments. Hire humans when the number has to stand up in court, at a regulator, or in a newspaper.
What does census-grounded mean?: Each simulated respondent is built from census-style demographic margins (age, education, income, race, party, place). The answers are generated by a model. The published check is how close the aggregate comes to real polls.
When is a synthetic screen the right tool?: When you need to drop a weak message, catch a leading question, or take a first read across markets before you spend on humans. A traditional focus group takes weeks to recruit. A queued synthetic study returns in about 15 minutes. Book a demo. The team sets up the study with you.
When do I still need a human sample?: When the sample itself is the deliverable: litigation, a regulatory filing, a published journalism poll, or any number that has to be defended as a probability design. Every Lewsearch report states the answers are simulated.

Check the work

Methodology and coverage Pricing Compare vendors Book a demo or take the free Blind Challenge.