Synthetic panels: methodology questions and answers

How Pervasive Insights™, by City Research Solutions, creates, calibrates and validates synthetic research panels, and what the published evidence through 2026 says about doing it honestly. Every two weeks we review new papers, industry guidance and published objections to synthetic data, and add answers here for the questions they raise.

How do you stop synthetic respondents exaggerating differences between segments?

Published benchmarks in 2026 show that uncalibrated AI respondents inflate the gaps between customer segments by two to four times, which can point a team at the wrong target segment entirely. Our second pass gates each synthetic response through a probability model fitted to the client's own primary data, pulling segment gaps toward what that client's real respondents actually showed. We report a segment-gap ratio (the synthetic gap divided by the human gap, measured on held-out human data) so the client can see how close to reality the panel runs.

How much of our own data do you need?

Less than most buyers expect. Published research shows that a human pilot of roughly five percent of a full sample is enough to align a simulator on the structure of the data, meaning the relationships between measures. We calibrate on primary research the client already has, such as an existing tracker or usage and attitude study, and we set a minimum human cell size per subgroup before we will report that subgroup synthetically.

What fidelity do you report, and at what level?

Three levels. Marginal fidelity asks whether the synthetic response distributions match human ones. Structural fidelity asks whether the relationships between measures match. Individual fidelity, where a matched human sample exists, asks whether each synthetic respondent tracks a real counterpart. We report all three overall and by key subgroup, because published work shows calibration can look strong overall while failing badly for a specific group.

Do you train a shared model on our data?

No. Calibration is per client and per study. Your primary research never feeds a model that serves any other client, and nothing you share with us improves anyone else's panel. This is a deliberate design choice: the alternative, a single model trained on a pooled bank of past studies, cannot be tuned to your market, your categories, or your respondents.

How stable are results between runs?

Synthetic estimates that change every time you run them are not research. We fix the generation settings, run replicates, and publish a tolerance band alongside every synthetic estimate, so you know both the number and how much it moves. Run-to-run instability is one of the main published criticisms of synthetic sample, and it is answerable with engineering discipline rather than promises.

How do you avoid the flat, over-agreeable answers AI is known for?

Two mechanisms. First, responses are generated as text and then mapped to scale points, a technique shown in 2026 research to restore realistic response spread that direct numeric prompting collapses. Second, the probability gate corrects directional bias, which the published evidence shows text mapping alone does not fix. The combination addresses both known failure modes: too little variance and too much agreement.

How is synthetic content labelled in your deliverables?

Every table, chart and export carries a tag identifying it as synthetic or human, consistent with the transparency requirements of the ICC/ESOMAR International Code. The method note on every deliverable states the calibration source and its date. A client should never have to ask whether a number came from real respondents.