What is synthetic research?

What synthetic research is

You write an instrument. A panel of simulated people answers it. You get percentages, the way you would from a human survey vendor, plus quotes and a PDF. The people are not recruited. They are built from census demographics and a model that has been scored against published polls.

That is the whole category. Some vendors sell interviews and themes. Some sell one chatbot that pretends to be a crowd. Lewsearch sells a tabulated panel. The statistic is the option share. The quote is color.

Synthetic respondent, synthetic panel, digital twin, one chatbot

These words get used as if they were the same object. They are not. A buyer who mixes them will buy the wrong SKU.

On Lewsearch, a synthetic panel is the default study. A digital twin is a matched agent you can interview, not a live feed of your CRM and not an industrial machine twin. One ChatGPT window is neither.

How a Lewsearch study is assembled

You write the question and the response options. You pick a market. The system draws respondents whose demographics match that market's margins, interviews them, and runs the same published calibrator every customer gets. There is no per-panel retune after the sale.

A 500-person Ohio study is not 500 copies of "an Ohioan." Each row is a distinct age, education, income, and party draw. The output is a topline plus crosstabs. If the question sits in a calibrated bucket, the report can point at that bucket's error. If it does not, the methodology says to treat the read as directional.

How accurate it is

7.47% mean absolute error on 404 ex-electoral questions of a 460-question benchmark vs real polls (Pew, Gallup, UT/Texas Politics Project, PPIC).

Best published panel: 4.88% MAE on the Texas UT/Texas Politics Project set (ex-electoral). California PPIC is 7.77% ex-electoral. The pre-registered held-out, sourced April 18, 2026 after training froze, is 10.68% overall and 9.97% ex-electoral. That last number is higher on purpose. The gap is the real-world spread between a well-specified panel and a noisier edge case.

Do not say "7.47% across 460 questions." The 7.47% is the 404-question ex-electoral slice of a 460-question pool. Electoral-margin items are harder and are scored separately. The per-question file has 443 scored rows.

How to read MAE

Mean absolute error is the average miss, in percentage points, between the panel's option share and the published poll share. It is not a sampling margin. Do not write it with a plus-minus sign. A 7.47% MAE does not mean "plus or minus 7.47 points on this study." It means that across the scored items, the average absolute miss was 7.47 points.

A low MAE on a category you did not ask is not a warranty. A high MAE on a category you did ask is a reason to hire humans or to rewrite the question. The file lives on methodology.

Synthetic panel vs a human survey

Use the synthetic pass to drop the confused option. Spend the human budget once, on the survivor. That is the job.

When to use it, when not to

Use it for message tests, concept screens, instrument lint (leading or double-barreled items), and a first read across markets before you buy respondents. A team can run ten framings for the price of one facility group.

Do not use it as a legal sample, a regulatory filing, or a published journalism poll. Do not use it as a demand forecast or a willingness-to-pay number you will take to a board as fact. Rare clinical or highly specialized populations stay directional unless you have a scored analog. Every Lewsearch report states the answers are AI-generated.

What you get back

Option shares for each question. Crosstabs by the cuts the study supports (age, region, party, and others depending on the market). Respondent quotes. A client-ready PDF that says the panel is simulated. Paid tiers add larger n, fuller crosstabs, and analyst notes. Live prices are on pricing. Do not invent a SKU that is not on that page.

Coverage

Published pool: over 650,000 simulated respondents. 27 live panels · U.S. national · 4 census regions (32 audiences). All 50 states and DC are on coverage. A study samples a market. It does not interview the whole pool. Live on a map is not the same thing as a published MAE for that state. The scored panels and the live list are both posted so a buyer can see the difference.

How this differs from interview tools and agreement-rate vendors

Adjacent products publish different units. Some interview tools report thematic parity. Some enterprise simulators report a Spearman correlation. Some panel vendors report an 80 to 95 percent agreement range against historical panels. Those are not MAE. They are not interchangeable, and they are not a reason to invent a Lewsearch percent that is not in the file.

The buyer test is the same on every compare page: published error, named benchmarks, n, and whether price is public. Units stay on the vendor's own site. See synthetic respondents accuracy for the unit map.

Where the numbers live

• Methodology: 7.47% mean absolute error on 404 ex-electoral questions of a 460-question benchmark vs real polls (Pew, Gallup, UT/Texas Politics Project, PPIC). Cross-validated. Raw and calibrated side by side.
• Per-question CSV: 443 scored questions from the 460-question pool.
• Coverage: live panels and the 50-state list.
• Pricing: credit packs, Pro, enterprise floor.
• llms.txt: citation map for answer engines.