What is synthetic market research? Types and evidence

Across 1,000 real-world surveys, Artificial Societies’ distribution accuracy is 86%, against a 91% human self-replication ceiling, according to its Survey Evaluation Report, January 2026. Distribution accuracy measures how closely simulated and human opinion distributions overlap. The human ceiling provides the reference because people vary their answers when asked again.

The report compares our method with biography-prompted large language models, or LLMs: models asked to answer as someone with an invented biography. It also measures self-contradiction, under the label hallucination rate. Here that means inconsistent answers from a persona; it does not measure errors about facts in the outside world.

Source: Artificial Societies, Survey Evaluation Report, January 2026.

Those results give you a measured basis for considering our method. They also have a boundary: matching the proportion of answers to a question does not establish which people hold which combinations of beliefs. A strategy aimed at a particular segment needs that second level of evidence. Our accuracy guide explains the separate measures, and our validity framework describes how we test the relationships between answers. We would assess a simulation intended to guide strategy on both its population results and the reasoning beneath them.

Artificial Societies’ benchmark results of 1 October 2026 show lower error than all nine off-the-shelf LLMs the Artificial Societies Benchmark tests, on every measure reported, with the same respondents, survey items and scoring code for each system. Across seven survey sources and 68,280 answers, Artificial Societies’ answer distributions needed an average adjustment of 8 percentage points to match the human responses, against 31 for the best LLM, Qwen3.8. That gap is the total variation distance, the share of answers that would have to change to match the real survey; within demographic groups on the 2024 General Social Survey, Artificial Societies’ was 0.063 against 0.222 for the best LLM, which is the segment-level evidence a crosstab relies on. The tests are defined in Artificial Societies Benchmark: A Validation Framework for Synthetic Research (opens in a new tab) (arXiv, September 2026).

Synthetic market research sits within a broader vocabulary of generated data and simulated people. The ICC/ESOMAR International Code, 2025 revision (opens in a new tab), defines synthetic data as “information that has been generated to replicate the characteristics of real-world data”. It defines a synthetic persona as “a digital representation of a person generated to mimic the behaviours, preferences, and characteristics of real people or groups”. Our opening definition applies those ideas to an audience-research task, and a synthetic audience is the simulated group those personas form when a study needs a population rather than one person.

The distinction matters when you evaluate evidence. In Using synthetic participants for market research, 2024 (opens in a new tab), the Market Research Society separates statistical data generation through machine learning from language models simulating research participants. Evidence that generated records preserve a dataset’s statistical patterns does not establish that a simulated participant responds credibly to a new decision.

The same report distinguishes synthetic data from anonymised data. Removing identifying details from a human response leaves an observed response; generating an answer creates a different kind of evidence. For a study using our personas, describe the findings as simulation results so the people acting on them understand how the answers arose.

Our Method & Evaluation page distinguishes three approaches by the balance between individual detail and population scale. Digital twins work from the individual upwards. Synthetic populations work from averages downwards. Our networks of enriched personas connect individual belief systems with relationships and influence, so you can examine both population opinion and differences within it.

Source: Artificial Societies, How Artificial Societies Accurately Models Human Opinion; audience boundary from our published method and audience guidance.

A crosstab compares answers across groups within an audience. If your recommendation depends on how those groups differ, a matching overall average leaves part of the research question unanswered. That is why we combine individual reasoning with population modelling, rather than ask you to choose between them.

Biography-prompted personas are the baseline in our evaluation, separate from the three approaches in the table. A model given an invented biography has a description to imitate; our method begins with observations of a real person. Keep that distinction in view when a demonstration offers convincing dialogue. The types of synthetic market research guide examines the architectures in detail, including the difference between describing a persona and grounding one in evidence.

Artificial Societies builds personas from observations of real people, then models their beliefs and connections. Our method names published research, online commentary and reviews among its inputs, alongside anonymised social data and public records of voting and investment. Sources include LinkedIn, Reddit and YouTube. The audience determines which observations matter; a plausible biography alone is insufficient grounding.

Persona construction uses psychometric methods, which examine patterns in attitudes, to reconstruct how individuals reason. We check each persona against audience benchmarks and prior behaviour before adding it to a society. Society creation then groups personas by shared characteristics and connects them through real relationships and patterns of influence. Artificial Societies’ method write-up specifies 12 to 3,500 personas per society.

Our platform names these stages Persona Genesis, Society Assembly and Simulation Engine. The Simulation Engine combines individual beliefs with network dynamics, introduces a scenario and models opinions at individual, segment and population level. You can use those levels to examine an overall reaction and the groups whose responses differ from it.

There is a public example of how answers become inspectable. Our September 2026 Pace the Frontier results page presents a survey run in Radiant, our client-facing application. Its response table includes society and role filters, answer filters and each persona’s verbatim reason. We label those numbers as forecasts of group reactions, with a clear distinction from polling real people.

Synthetic market research is useful when you need to test how opinion changes before making a decision. Our method establishes an audience baseline, exposes personas to a message or scenario and examines the response through the group. For strategic communications or investor communications, that makes the research question concrete: how a proposed decision changes the audience’s starting position, including through social influence.

Confidential material is a particular reason to simulate. Our earnings-call case study describes testing unreleased numbers and draft scripts that could never be shown to a real investor. The audience receives the material inside a simulation. You can explore alternative decisions before choosing what to put into the world, including options you would not take to a live audience.

Teneo’s own account of the engagement is in our Teneo case study; it is a client’s account of one engagement, rather than a promise about every research task.

Our work also addresses public affairs and hard-to-reach audiences. Difficulty recruiting people need not rule out research where diverse observations of them exist. Simulation needs observations even when recruitment cannot reach the people concerned, which makes the available evidence a practical starting point for scoping the study.

Compare synthetic market research companies by the decision you need to make and the evidence behind their proposed method. NIQ’s guidance states: “Synthetic feedback should be trusted only when the supplier has access to data that validates its accuracy” (Levanti and Verret, updated February 2026 (opens in a new tab)). Ask for the test population, the questions and the meaning of the score before treating a headline percentage as evidence for your study; how to evaluate a synthetic research vendor sets out the questions in full.

For interview work, Synthetic Users describes a late-February-2024 comparison of eight human and eight synthetic interviews with UK school teachers about classroom technology. Its science post (opens in a new tab) reports qualitative similarity through thematic overlap, depth, coverage and alignment. That is evidence about interview content, whereas our distribution measure concerns the balance of survey answers.

For wealth research, EY’s account of its work with Aaru, 7 October 2025 (opens in a new tab), reports a 90% median correlation across 53 single-select questions against research involving 3,600 affluent investors in more than 30 markets. The account comes from a commercial partnership. Neither that correlation nor interview similarity should be read as interchangeable with distribution overlap.

For vendor-specific questions, see our Artificial Societies alternatives comparison and AI market research tools guide. We would start a strategic opinion-shift study with the networks approach and require evidence at segment level before accepting a recommendation based on an audience average.

Synthetic market research needs evidence that simulated beliefs behave coherently as well as matching survey totals. Artificial Societies’ September 2026 validity framework, by Edoardo Chidichimo and Felix Wallis, sets out eight tests. They examine whether a stimulus caused a response change, whether questions measure the intended attitude and whether findings hold for real populations. These are the questions behind internal, construct and external validity.

The tests include rephrasing questions and changing answer order. They also distinguish changing the presentation from changing the substance: removing evidence supporting a claim should weaken its reception. This matters because a language model can agree with the suggestion in a question rather than respond to the proposed decision. We test for stable beliefs instead of relying on how persuasive the answer sounds.

A second test compares a simulation with published experiments on real people. In September 2026, Artificial Societies reproduced four published UK experiments on misinformation, fact-checking and prebunking, and compared its synthetic population with evidence from more than 25,000 real Britons, having hashed and time-stamped the design before the first survey call (Simulating Misinformation, Fact-Checking, and Prebunking). Fact-checks cut belief in false COVID-19 claims by 46% in the synthetic panel against the control group, where the original study reported 45%. The same replication found a small bias towards vaccine-positive content in one control group, and the post reports that miss alongside the matches.

The network modelling rests on a peer-reviewed paper co-authored by Artificial Societies’ James He and Felix Wallis in the British Journal of Psychology (opens in a new tab) (He, Wallis, Gvirtz and Rathje, December 2024); it does not validate the survey accuracy figures, and the science page sets out what it found.

Synthetic market research should leave physical experience to research with people. We would commission human participants for taste trials, hands-on packaging tests and usability research that requires interacting with an artefact. In-person ethnography also belongs here. Our scope is modelling opinion and reactions to decisions; a simulated answer cannot provide the experience those studies investigate.

Observations are a second boundary. Our personas require diverse evidence about the audience, whether public or client-provided. Where that evidence is absent, we would begin by obtaining it. Results measured against observed audiences provide no basis for assigning the same accuracy to an unobserved group.

There is also a distinction between stated opinion and action. Our validity framework explains that synthetic research inherits the intention–behaviour gap from surveys: what people say and what they do can differ. Use simulation to examine a decision’s likely reception, and use observed behaviour when the decision requires evidence of action.