Can synthetic audiences answer questions your survey never asked?

Decide what is actually new

A new question might concern a different message in a familiar category, an unmeasured attitude, or a market the source research never covered. Those are different claims about what a model can predict.

Kantar cautions that digital twins need questions relevant to their underlying knowledge. Success within one product category does not establish performance on unrelated decisions. Ask the vendor to name the boundary of its supporting evidence. Kantar on synthetic-data limits.

This also differs from boosting an existing survey, where the original questionnaire and observed response patterns are already available.

Match the provider to the research output

“Synthetic audience” covers several purchases. The following examples describe vendors’ published offerings, not verified accuracy or a ranking.

• Customer-data audiences. Fairgen describes turning research respondents into twins for subsequent studies. Panoplai combines data ingestion, synthetic enrichment, digital twins and reporting. Ask what happens when the new question concerns something absent from the source records. Fairgen, Panoplai.
• Survey estimates. Qualtrics documents a U.S. general-population synthetic panel in English. It favors forward-looking and attitudinal questions over detailed recall and advises broad screening. Confirm audience and instrument compatibility before buying. Qualtrics Synthetic Panels documentation.
• Specialized buyers or scenarios. Evidenza markets surveys of AI-generated customer copies, including hard-to-reach B2B buyers. Aaru describes simulated populations for product, price, message and strategy decisions. Ask for evidence on your particular audience and decision, beyond the category description. Evidenza, Aaru.
• Exploration and product feedback. Synthetic Users describes a discovery co-pilot using simulated research. Versive offers AI-persona tests alongside separate studies with real participants. Confirm whether each deliverable contains generated feedback or observed human responses. Synthetic Users, Versive’s workflow comparison.

Ask for a test of unseen questions

A useful buyer test is to withhold entire target questions or new stimuli from audience construction and tuning, then compare the predictions with human answers kept by the evaluator. If prior questions are permitted inputs, record exactly which ones.

Holding out respondents while letting the model learn from answers to every target question tests a different task. It may assess prediction for additional cases on familiar variables. It does not, by itself, demonstrate performance on questions the source study never asked.

Illustrative example: a retailer has historical research about store visits and wants to screen three new delivery propositions. Give the provider the permitted historical inputs and the three propositions. Keep the new human comparison results hidden. Evaluate the proposition ranking and key objections using rules agreed before seeing either result. This example is a proposed test, not a Lewsearch case study.

Buy the scope the evidence supports

Useful limitations should appear in the buying discussion. Ipsos positions InnoTest with Synthetic Respondents for early idea and concept screening. Its documentation limits current stimuli to text, notes limited handling of pictures, pricing and sizing, and says the synthetic version is not recommended for concept validation. It also assumes price is accepted rather than measuring expensiveness. Those boundaries belong to that product; do not generalize them to every provider. Ipsos InnoTest documentation.

In your proposal, record the supported market, category, question types, stimuli and output. Decide which findings will advance to human research. A compelling interview transcript does not establish accurate population shares, and a useful screening result does not establish a sales forecast.

Start with a question you can check

For a Lewsearch pilot, agree on the audience and questions, keep the human evaluation results hidden, and compare only after predictions are locked. Use that result to decide the next research step. Demo-led self-serve starts at $595/month; Enterprise deployments are custom scoped.

Book a demo and review the blind-validation guide.