Silicon Sample

Research groups showed that language models can mirror the response patterns of certain population groups remarkably well when the prompt conditions the persona sufficiently. In practice teams use this for pre-testing: claim variants, price reactions, concept directions. The limits are well documented: models smooth extremes, under-represent minority positions and reflect training-data bias. Serious applications therefore calibrate against at least one real reference survey, report deviations openly and use synthetic results purely for shortlisting, never as the sole basis for a decision. Robust market figures still require real fieldwork.

For marketing teams a silicon sample shortens the loop between idea and first signal from weeks to hours. Used correctly it reduces expensive wrong turns because weak concepts are eliminated earlier.

Before a campaign launch a team tests twelve claim variants against a synthetic panel of four personas. The three best-scoring ones then go into a real survey with 400 participants — cutting test cost and duration considerably.

The first academic work on language models as substitutes for human respondents appeared in 2022 and 2023 and triggered an intense methodological debate. From 2024 market research vendors offered synthetic panels commercially. By 2026 a pragmatic practice has settled: synthetic pre-studies to form hypotheses, real surveys to confirm them — accompanied by industry guidance on labelling synthetic data.