Synthetic respondents: what they replace, and what they cannot

Outline. Each heading below carries one line saying what the section has to establish. Expand each into two to four paragraphs, put at least one number or worked example in the piece that nobody else could publish, then delete this note.

Synthetic sample arrived as a cost argument and is settling into something narrower and more defensible. The line is not real against fake — it is which decisions hold up afterwards.

What a generated respondent actually is

A distribution conditioned on a profile, not a person. Say it plainly; most coverage does not.

Where it holds: pretesting, piping logic, instrument length

Uses where being wrong is cheap and the alternative is testing on real people for nothing.

Where it fails: anything with a tail

Generated answers regress to the mean, so the minority position — the one you commissioned the study to find — disappears.

Incidence cannot be estimated from it, and that is the expensive mistake

Explain why a model trained on population data cannot tell you who is reachable in your panel this fortnight.

Disclosure, and what a buyer is entitled to know

If any part of a deliverable is generated, the buyer has to be told which part. State the position.

The question to ask a vendor

Not how accurate — ask what it got wrong on a study where the real answer was already known.