Synthetic market research is market research in which AI-generated respondents answer the questions instead of recruited people. A language model is conditioned on a description of the audience you want to study (age, region, role, habits, attitudes, sometimes real survey or interview data), and each conditioned instance answers survey items, reacts to a concept or holds an interview. The answers are then aggregated like survey results or read like qualitative transcripts.
The method is useful because it is fast and cheap enough to run before any fieldwork: you can screen many concepts, messages or questionnaire drafts in a day. It is limited because the respondents are simulations. Their answers are estimates of how a group might respond, shaped by the model and the data it was given, and they need to be checked against real people before they carry a decision that is expensive to reverse.
This page covers the market research use of the method. For the wider practice, including UX research and policy work, see synthetic research.
Three elements make a study synthetic market research rather than a chat with an AI tool:
- A defined audience. The study names who is being simulated, for example "UK grocery shoppers aged 25 to 40 who buy plant-based milk at least weekly", and each synthetic respondent is built from that definition.
- A research instrument. The respondents answer the same structured questions, rate the same stimuli or go through the same interview guide, so answers can be compared and aggregated.
- An analysis and validation step. Results are summarised as distributions, rankings or themes, and someone checks how far they can be trusted, ideally against real survey data for the same audience.
Many vendors use different names for the same idea: synthetic respondents, synthetic panels, AI personas, digital twins of customers, silicon samples. The academic term "silicon sampling" comes from political science work that conditioned a language model on the backstories of real survey participants. The terms overlap; what matters is the three elements above. For a closer look at the respondent unit itself, read synthetic respondents and synthetic panels.
People often search for "synthetic data market research", and the phrase blurs two different things.
Synthetic data is any artificially generated dataset. In analytics it usually means a statistical copy of a real table, such as transactions or customer records, created so that analysts can work without exposing individuals. Its job is to preserve the patterns of data you already have.
Synthetic market research generates new answers to questions that have not yet been asked of anyone: how would this audience react to a price change, which of these three claims is most believable, what would stop them switching brands. Its job is to estimate opinions and reactions you do not yet have.
The answers a synthetic study produces are a kind of synthetic data, but the hard part of the method is not generating rows. It is defining the audience well, asking neutral questions and knowing how far the answers can be trusted.
A typical study runs in five steps.
- Frame the decision. Write down what will change depending on the result, for example "we launch claim A or claim B", and what level of confidence you need.
- Build the audience. Describe the segment, then generate respondents that vary within it. Better systems ground each respondent in evidence: published statistics, public sources, your own customer research or survey files. A respondent built only from a demographic label tends to answer like a stereotype.
- Run the instrument. Ask closed questions (single choice, multiselect, scales, rankings, forced-choice designs such as MaxDiff) and open questions, or show stimuli such as concepts, ads, packaging images, landing pages or prototypes.
- Analyse. Aggregate closed answers into distributions and compare segments; code open answers into themes; look for contradictions and for answers that sound generic.
- Validate and decide. Compare a sample of results with real data where it exists, label the output as directional, and decide which findings go forward to a human study.
How answers are collected matters more than it looks. In Minds' published survey replications, asking synthetic respondents for probabilities across answer options instead of forcing one choice each reduced distribution error by 12.47 to 19.72 percentage points across four public datasets (Minds research). Forcing one answer per respondent throws away uncertainty that real populations have.
There is no single accuracy figure for synthetic market research, and a vendor quoting one without naming the dataset, the questions and the comparison method is not giving you usable information. The published evidence points in both directions.
Evidence that it can work:
- Argyle and colleagues conditioned GPT-3 on thousands of socio-demographic backstories from real US survey participants and found it could emulate response distributions of many subgroups. They called this property "algorithmic fidelity" and the resulting samples "silicon samples" (Political Analysis, 2023).
- A Stanford-led team built agents for 1,052 Americans from two-hour interviews, surveys or both. On held-out General Social Survey items, the agents reached 83%, 82% and 86% of the participants' own two-week test-retest consistency, against 74% for agents given demographics only (arXiv, 2024). Grounding in real self-reports beat demographic labels.
Evidence of the limits:
- Bisbee and colleagues asked ChatGPT, prompted with personas, for feeling-thermometer scores that they compared with the American National Election Study. Averages matched closely, but the synthetic answers had less variation than real ones, regression coefficients often differed, results shifted with small wording changes and the same prompt gave different results three months apart (Political Analysis, 2024).
- Santurkar and colleagues found substantial misalignment between the opinions language models express and those of 60 US demographic groups, persisting even when the model was steered towards a group; some groups, such as people over 65, were reflected poorly (ICML 2023).
- In consumer research specifically, Brand, Israeli and Ngwe found that willingness-to-pay estimates from language-model responses were sometimes comparable to human studies but often inaccurate and in some cases had the wrong sign. Fine-tuning on earlier survey data from the same category improved alignment for new features within that category, but not for new product categories (HBS working paper, SSRN).
The pattern across these studies: synthetic answers are most reliable for well-documented attitudes in populations that are well represented in data, and when respondents are grounded in real evidence about the people they stand for. They are least reliable for variance, subgroup differences, new categories, price sensitivity and anything that depends on lived physical experience.
Minds treats validation as part of the product rather than a claim on the website. Two pieces are relevant.
Audience Validation. An Audience in Minds can be checked against real published surveys, against survey files you upload (PDF, Excel, CSV or Word), or both. Minds keeps only the questions whose published answers are a fair target for that Audience, asks its Minds all of them in one run, and lists every question it left out with the reason. Each survey gets a score out of 100 with a 95% range, its source and who it asked. New Audiences are validated automatically once trained, and any Audience can be validated again later. This gives you a check on the specific audience you are about to use, which is more useful than a general accuracy claim.
Published replications. Minds has published replications against five public datasets: GSS 2024, ANES 2024, the 2024 Canadian Election Study, PISA 2022 and the UK Food Standards Agency's Food and You 2 survey. Mean distribution error ranged from 4.72 to 8.33 percentage points (full results). The studies use different instruments and one is weighted, so do not pool them into one accuracy score. The same report notes that on the Food and You questions an equal-probability baseline already reached 7.00 points of error, so a low error figure on its own is not proof of a large advantage. Read how Minds research panels are built for the method.
What this does not show: matching the published percentages of a survey is not the same as simulating each individual correctly, and good results on one survey do not transfer automatically to a different audience or a different kind of question. That is why the check should be run on your audience, close to your question.
| Method | Typical speed | Relative cost | Representativeness | Best for | Main risk |
|---|
| Synthetic market research | Minutes to hours | Low per study | Not representative by itself; depends on grounding and validation | Screening concepts and messages, pre-testing questionnaires, exploring segments that are hard to recruit | Plausible but wrong answers, too little variance, model bias |
| Online panel survey | Days to weeks | Medium | Can approximate a population with good sampling and weighting | Sizing, tracking, representative estimates | Low-quality or fraudulent respondents, panel fatigue |
| Focus groups and interviews | Weeks | High per participant | Not representative; small samples | Depth, language, emotional reactions, unexpected reasons | Moderator and group effects, small n |
| A/B or in-market test | Weeks, needs live traffic | Medium to high | Real behaviour of your actual users | Final choice between live variants | Only tests what you can ship; slow for many options |
These methods work best in sequence. Synthetic research narrows the options, a survey or qualitative study confirms the shortlist, and a live test settles what can only be settled by behaviour.
It earns its place when speed or reach matters more than precision:
- Early concept and message screening, when there are more options than the budget allows to field.
- Questionnaire development: finding confusing questions, missing answer options and leading wording before paying for fieldwork.
- Exploring hard-to-recruit or expensive audiences, such as specialist B2B buyers, as a first read before targeted interviews.
- Iterating on positioning, packaging copy or onboarding text where each round would otherwise take weeks.
- Preparing for human research: forming hypotheses, choosing which segments to sample and drafting the discussion guide.
For concept work in particular, see AI concept testing.
Use recruited people, observed behaviour or established measurement instead when:
- You need a representative estimate: market size, penetration, share, or the exact percentage who would buy.
- You need price elasticity or willingness to pay at specific price points to set a price.
- The claim will be published, regulated or audited, including advertising claims that need substantiation, clinical or health claims and election forecasting.
- The product must be touched, tasted, smelled or used physically.
- The audience, category or behaviour is new, so little reliable data describes it, or the audience is poorly represented in public text (the misalignment studies above point to older people and other under-represented groups).
- The decision is expensive to reverse and you have not validated the synthetic Audience against real data for this kind of question.
Synthetic results also carry correlated error: every respondent comes from the same underlying model, so they can be wrong in the same direction at once. Real respondents make independent mistakes; simulated ones may not.
Ask vendors for evidence, not adjectives:
- Which public datasets have you replicated, with which questions, and what was the error? Is a baseline reported?
- Can I validate my own Audience against real survey data, and do I see which questions were excluded and why?
- How are respondents grounded: demographic labels only, or sources, statistics and my own research files?
- Which question types are supported (open, single choice, multiselect, scales, MaxDiff) and how are closed answers collected?
- What do you say the tool should not be used for?
A vendor comparison is in our synthetic market research tools guide, and a broader method comparison is in synthetic research vs traditional market research.
Minds is a platform for running synthetic market research end to end: build an Audience of Minds grounded in sources and your own files, run Studies with open and closed questions or MaxDiff, test stimuli such as concepts, images, websites and copy, analyse and export the results, and validate the Audience against real surveys. Minds is built for directional decisions. It does not make a synthetic study statistically representative, and for representative estimates, regulated evidence or physical product tests you will still need recruited people.
- Argyle, L. P., Busby, E. C., Fulda, N., Gubler, J. R., Rytting, C. and Wingate, D. (2023). Out of One, Many: Using Language Models to Simulate Human Samples. Political Analysis, 31(3), 337-351.
- Bisbee, J., Clinton, J. D., Dorff, C., Kenkel, B. and Larson, J. M. (2024). Synthetic Replacements for Human Survey Data? The Perils of Large Language Models. Political Analysis, 32(4), 401-416.
- Brand, J., Israeli, A. and Ngwe, D. (2023). Using LLMs for Market Research. Harvard Business School working paper, SSRN 4395751.
- Santurkar, S., Durmus, E., Ladhak, F., Lee, C., Liang, P. and Hashimoto, T. (2023). Whose Opinions Do Language Models Reflect? Proceedings of the 40th International Conference on Machine Learning.
- Park, J. S. and colleagues (2024). LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals, first released as "Generative Agent Simulations of 1,000 People". arXiv 2411.10109.
- Minds (2026). We Tested Synthetic Audiences Against Reality.
