A useful perspective.
An honest boundary.
How synthetic audience research works, and where its limits matter.
What we mean by synthetic
The homepage audience image is AI-generated artwork, not a photograph of research participants.
A synthetic reviewer is a simulated perspective, not a recruited person. Generated responses are labeled synthetic throughout the report and exports, and each study is saved privately to your account.
Geography describes context
The location directory uses the Census Bureau’s 2024 Gazetteer. A place name does not establish what its residents believe. Demographic information alone does not determine preferences, abilities or intentions.
How a study works
- Keep your material, audience definition and research question separate.
- Construct simulated perspectives with documented population inputs and explicit assumptions.
- Combine structured responses and retain the links from summaries to underlying responses.
- Show missing responses, source limitations and uncertainty alongside findings.
We use benchmark-informed methods under active evaluation. Transfer to new topics and populations is experimental. Historical performance does not establish accuracy for your study. We keep implementation details private while explaining the evidence and limits that matter for interpretation.
The homepage benchmark
The homepage compares weighted human and simulated support in a matched subset of the 2025 Cooperative Election Study: 997 distinct respondents, 9,380 paired answers and ten policy questions. Each question is excluded while fitting the combination. Respondents recur across questions, so this evaluates questions in an existing panel rather than new people.
The displayed method was selected after exploratory comparisons. Mean absolute share error is 4.94 percentage points across all ten questions; the largest error is 12.99 points. The renewable-electricity chart is selected from that original evaluation, with all ten results expandable. These are observed sample-level errors, not margins of error or evidence of county, product, advertising or individual predictive accuracy. Public historical data, related topics and exploratory selection limit generalization.
The homepage also shows four selected examples from 28 additional binary questions, including near-even splits and everyday work, energy, housing and health-care topics. The method was frozen before evaluating these new questions; the same people recur across evaluations. Mean absolute share error across those 28 questions is 5.00 percentage points. All 28 results are available in the homepage study details. Missing responses and simulation refusals affect which respondents enter each matched comparison. These figures do not establish individual-level accuracy.
Data citation: Schaffner, Brian; Pope, Jeremy; Shih, Marissa (2026), CES Common Content, 2025, Harvard Dataverse, V1. Published under CC0 1.0, which permits commercial reuse. Charts and simulated comparisons are our own analysis; the data creators and repository do not endorse this product.
What this cannot establish
Simulation does not measure public opinion, establish county-level demand or predict purchases, votes or campaign outcomes. It cannot replace representative surveys, interviews or real-world tests. Larger simulated panels do not remove model bias.
Your next real-world test
Use a finding to form a testable question: which phrase caused confusion, what assumption needs checking, and what would you ask a real person next?