Synthetic market research answers market research questions with AI-generated respondents or generated data instead of, or before, real people. It covers simulated survey panels, synthetic data that extends a real survey, and synthetic interview participants. It works best for screening concepts, prices and messages before a panel, and it doesn’t replace real respondents where you need a reportable number.
This page is for market research and insights teams. It maps the methods you already run (surveys, interviews, concept tests, pricing studies) to the synthetic options that can help with each, shows what published research says about accuracy, and explains where Candor fits. For the method as product teams use it, read What is synthetic user research?
The term covers three different things, and vendors often use it for only one of them.
Synthetic data starts from a real dataset, such as a survey you already ran, and generates more records with the same patterns. Teams use it to fill out subgroups too small to report on, or to share data without exposing real respondents.
Simulated survey respondents (often called synthetic respondents) are AI-generated people who answer a questionnaire. They estimate how a group would answer closed questions, and they are checked by comparing their answers with a real survey of the same group.
Synthetic interview participants are AI-generated people you question in depth, the way you would in a qualitative interview. They explain why they would choose, reject or pay for something. This is what Candor builds.
A fourth kind, digital twins of real people, copies one specific person rather than a type of customer. The synthetic user research explainer sets out all the kinds side by side, with what would show each one works, and the glossary entry on synthetic respondents covers the survey kind in more depth.
Synthetic research helps most with the questions that come before a panel: which ideas to pursue, why customers react the way they do, and which wording or price to test. It helps least with the questions a panel exists to answer: how many people, measured with confidence, over time.
| Method | What it answers | Synthetic option | How far it goes |
|---|---|---|---|
| Customer survey | What share of the market thinks or does something | Simulated survey respondents, or synthetic data that extends a real survey | Directional only. Report percentages from real respondents |
| Survey with a small subgroup | How a group too small to report on compares with the rest | Synthetic data generated from the real survey's patterns | Works only when the real data holds enough of that group to learn from |
| In-depth interview | Why people think, choose and object the way they do | Synthetic interview participants built from evidence about the audience | Strong for early rounds. Real interviews for nuance that depends on lived experience |
| Focus group | How people react to an idea and to each other | Synthetic interviews, run one participant at a time | Group dynamics are not simulated. Candor interviews each participant on their own |
| Concept test | Which ideas are worth developing, and why | Synthetic interviews that react to each concept | Strong for screening many concepts. Send finalists to a panel for numbers |
| Pricing study | Where price resistance starts, and what people would pay | Synthetic interviews that test price points and the reasons behind them | Directional. Don't set a launch price from synthetic results alone |
| Message and name testing | Which wording is clearest and makes people want more | Synthetic interviews that compare versions side by side | Strong for comparing versions. Confirm the winner with real people before a major launch |
| Segmentation | Which groups exist in the market, and how big each one is | Synthetic participants spread across groups found in the evidence | Can explore how groups differ. Can't measure how big each group is |
| Brand tracking | How awareness and perception change over months | None that holds up | Needs real people measured over time |
| Claim substantiation | Evidence that an advertising or label claim is true | None | Use a study with real respondents for any claim you will publish |
For how synthetic and panel research fit into one research cycle, see Traditional vs synthetic market research.
Accurate enough to rank options and explain reactions, when the simulated people are built from real data about the audience. Not accurate enough to report a percentage. Three published studies show where the line sits.
Argyle and colleagues (2023) built simulated survey respondents from the profiles of people in large American surveys. The simulated answers matched the real ones beyond surface patterns, reflecting how people’s background and attitudes combine.
Park and colleagues (2024) interviewed 1,052 Americans for two hours each and built a simulated version of each person. The simulations reproduced people’s survey answers about 83% as well as the people reproduced their own answers two weeks later. Simulations built from demographics alone did noticeably worse, at 74%.
Brand, Israeli and Ngwe asked language models what people would pay for products and features. The estimates sometimes matched studies with real people, but were often wrong, some in the wrong direction. Training the model on earlier survey data from the same product category helped. It did not help for new categories or for differences between customer segments.
The pattern across all three: accuracy comes from real data about the people being simulated, and it falls off for new categories and for exact numbers. A synthetic study is only as good as the evidence behind its respondents. Ask any vendor what theirs are built from.
Candor hasn’t yet run its own studies side by side with the same studies run on real people. Until it does, each Candor participant shows which of their fields rest on evidence you can check and which were filled in. The essay What synthetic research skeptics get right covers the strongest objections to the method.
Run it before the panel, not instead of it. A project that starts with 20 concepts can rarely afford a panel round for each, so most get cut on judgment. A synthetic round gives every concept a set of reactions and reasons first, and the panel budget goes to the few that earn it.
It also makes the panel round itself better. Interviewing synthetic participants first shows which questions confuse people, which answer options are missing, and which objections the questionnaire should probe. That costs hours, not a second fielding.
The essay Where synthetic research fits in your workflow walks through the hand-offs in detail, and Concept testing for CPG: the case for synthetic first applies it to consumer goods.
Candor builds synthetic interview participants (often called synthetic personas) from real evidence about your audience: web sources about the market, population data for the region, and your own research if you upload it. Business and consumer audiences are modeled separately, because a buying committee and a shopper decide in different ways.
You then run the study types insights teams already know: concept testing, price testing, message testing, name testing and problem discovery. Most studies finish within a working day, mostly in the background, and end in a report with the full transcripts behind it.
Candor doesn’t simulate survey panels or produce percentages with confidence intervals. It answers the “why” and “which one” questions that come before them. The how-it-works page shows each step, and Candor for consumer insights teams shows how insights teams use it.
Start with which kind of synthetic research you need, because a survey simulator and an interview platform solve different problems. Then ask four questions of any vendor:
Best synthetic user research tools (2026) compares 29 platforms, including the survey simulators, in nine categories with a seven-point framework for judging them.
Synthetic data in market research is data generated by a model rather than collected from real people. It comes in two main forms. One extends a real dataset, for example by filling out a survey subgroup that was too small to report on. The other simulates new respondents: AI-generated people who answer survey questions or sit for interviews. The first is checked against the real dataset it came from. The second is checked by running the same questions with real people and comparing the answers.
It depends on the question and on what the model knows about the audience. Published studies show simulated respondents can match real survey patterns closely when they are built from real data about the people they stand in for. Accuracy drops for new product categories, exact percentages and differences between segments. Treat synthetic results as directional: good for deciding which options deserve real research, not for reporting a number to a board.
No. Synthetic research can screen concepts, explore why customers decide the way they do and sharpen a questionnaire before it goes to a panel. A real panel is still the right tool for statistically confident percentages, brand tracking over time and any claim you have to substantiate. Most teams get the most from synthetic research by running it first, so panel budget goes to the questions that survive.
The pros: a study takes hours rather than weeks, it costs a small fraction of a panel round, it needs no recruiting, and it reaches audiences that are hard or slow to recruit. The cons: it gives directional answers rather than statistical estimates, it is weakest for brand-new categories and lived-experience nuance, and its quality depends on the evidence and checks behind it. A vendor should be able to show you both lists for its own product.
They overlap. Synthetic user research usually means interviewing AI-generated participants about a product: its problems, concepts, names and prices. Synthetic market research is the wider term used by market research and insights teams, and it also covers synthetic survey data and simulated survey panels. Candor does the interview kind for both groups: product teams call it user research, insights teams call it market research.
Synthetic platforms charge by subscription, by study or by interview. Candor has a free plan with interviews at $10 each, and a $75 a month plan with interviews at $7 each, so a 20-interview study costs $140 to $200 in interviews. Check the pricing page for current plans.
Start with a free project, or walk through Candor with us first.