Market Research Started Interviewing People Who Do Not Exist

The pitch is easy to see the appeal of. Instead of recruiting a panel, waiting a fortnight and paying per completed response, a research team defines a segment — a small-business owner in a mid-sized city, price-sensitive, sceptical of subscriptions — and asks a model to answer as a thousand of them. Concept tests, message variants and pricing questions come back in minutes. Some agencies now offer this as a product, and some clients have quietly substituted it for early-stage qualitative work. The results are seductive because they are fluent and internally consistent, and that is precisely the problem. A simulated respondent answers from the aggregate of everything written about people like the one described, which means it reproduces the conventional wisdom about that segment with great confidence and misses whatever the segment actually thinks that has not yet been written down. It is a very good mirror of existing assumptions and a very poor instrument for discovering that they are wrong — which is the only reason to do research at all. Comparisons against real panels show the simulated ones cluster toward the middle, underrepresent strong minority views, and drift with the wording of the persona prompt in ways a real population does not. Where the technique earns its place is upstream of the real work: stress-testing a questionnaire before it goes to a panel, generating the range of objections a message might meet so the human study can probe them, and rehearsing interviews so that junior researchers arrive at the real ones prepared. Used that way it makes the expensive part better rather than pretending to replace it. The line to hold is disclosure. A deck that presents synthetic responses as if they were fieldwork is not a shortcut; it is a fabrication that happens to have been produced by a machine. Label the method, keep the synthetic and the human findings in separate columns, and let the reader see which is which.