Synthetic market research is the practice of using AI respondents (language models conditioned to react like a specific slice of your market) to simulate how real customers would respond to a product concept, message, or price. It produces a fast, low-cost directional read that you then validate with real people when the decision matters.
Traditional market research forces a trade-off: fast and cheap, or rigorous and slow. Recruiting a panel, fielding a survey, and analysing the results takes weeks and real budget, so most concepts never get tested at all. Synthetic market research collapses the screening step to minutes and pennies, letting you interrogate an idea long before you commit to fieldwork.
The key word is screening. Synthetic responses are best understood as an instrument for finding risks, ranking options, and killing weak ideas early, not as a replacement for ground-truth validation with verified humans.
How synthetic market research works
A useful synthetic study is not just "ask ChatGPT what it thinks." The naive approach fails in a predictable way: prompt a language model to rate a concept 1–5 and it hugs the middle, almost never commits, and produces ratings no real market would ever give. Credible synthetic research works differently:
- Brief the respondent. Each synthetic respondent is conditioned on a real segment of your market: defined by category usage and buying behaviour, the attributes that actually shape a purchase decision.
- Elicit a genuine reaction. Instead of asking for a number, you collect what the respondent would actually say: an in-their-words reaction, plus what draws them in and what holds them back.
- Map words to a distribution. Each reaction is translated into a purchase-intent distribution by measuring how closely it resembles a set of calibrated reference statements: the meaning of the words, not a number the model invented.
- Aggregate across the panel. Across dozens of respondents you get a realistic spread of intent, not a flat "everyone says 3."
Is synthetic market research accurate?
It can be surprisingly accurate for the right questions. In peer-reviewed work by researchers at PyMC Labs and Colgate-Palmolive, an elicitation-and-similarity method recovered roughly 90% of human test–retest reliability on purchase intent, validated against 9,300 real responses across 57 consumer surveys, and the response distributions tracked the real human ones, not just the averages. (Maier et al., 2025, arXiv:2510.08338.)
That is strong evidence for a screening tool, though still not a licence to treat synthetic numbers as ground truth. Accuracy depends heavily on the category, the market, and how the study is run.
Where synthetic research is strong (and where it isn't)
Surfacing risks and objections fast, ranking variants and price points against each other, and producing realistic intent distributions in categories the model knows well, all at near-zero cost, so you can run counterfactuals freely.
Absolute numbers you'll bet the launch on, unfamiliar categories and non-Western markets (where the model is weakly calibrated), and demographic subgroup cuts by gender or ethnicity: those are persona colour, not statistical signal.
The honest way to use synthetic research is as the first layer of a funnel: screen broadly and cheaply with AI respondents, then validate the survivors with real, verified humans. The synthetic layer gets sharper over time as each human study calibrates it against reality.
Synthetic vs. traditional research
Synthetic research doesn't replace surveys or focus groups; it changes when you reach for them. Instead of committing budget to test one or two concepts, you screen twenty with synthetic respondents, then spend your fieldwork budget only on the ideas that survived. See synthetic research vs. traditional surveys and vs. focus groups for the side-by-side.