Synthetic vs traditional market research compared

Audience simulation lets you investigate responses to material you cannot put before the intended audience. In Artificial Societies’ hyperscaler earnings-call study, a leading hyperscaler tested unreleased numbers and alternative scripts against simulated investors and analysts. The engagement took 48 hours from start to finish, according to the case study. The client expected a margin reset to dominate; simulated analysts concentrated on demand durability and financing commitments. The live call followed that pattern, the study reports.

Our method begins with observed data, including public commentary and anonymised social data. We reconstruct how people reason, check personas against audience benchmarks and prior behaviour, then connect them through relationships and patterns of influence. You can examine an answer at market, segment or persona level. In the earnings-call engagement, the client could rerun scenarios and retest prepared remarks as the numbers changed.

That gives an investor relations team a rehearsal before disclosure. The live-call agreement is a case-study observation, separate from our survey evaluation. Distribution accuracy measures the overlap between simulated and human opinion distributions; it does not establish that a particular investor will react as predicted. You should read the earnings-call account for its decision and design, and the evaluation report for its benchmark.

Traditional market research is the better choice when your question requires a person to experience something. A taste test needs human tasters. Physical usability research and studies following the same real individuals over time also belong in the fieldwork budget. We design simulation for decisions you can rehearse before acting, with direct human validation retained where the decision requires it; Can AI replace market research? sets out why simulation does not remove the need for that validation.

For a publicly available product and 1,000 responses from readily recruited consumers, we would generally recommend a traditional panel. You have an accessible audience and a straightforward question. Buying a simulated audience for that assignment needs a reason beyond the availability of the technology, such as a subsequent confidential narrative test.

Recruitment need not consume the schedule. Nexus Expert Research’s 27 January 2026 guide (opens in a new tab) says pre-screened panels can reduce recruitment from two or three weeks to three to five days. Drive Research’s 29 June 2022 guide (opens in a new tab) gives three to five weeks for focus-group projects, including recruitment and reporting. These are agency estimates for defined methods. Compare the proposed work and its delivery stages before deciding that simulation will save time.

Synthetic and traditional market research differ in what the buyer receives as evidence. A human study records participants’ responses; a simulation estimates responses from a synthetic audience, a model of the people concerned. Biography-prompted large language models, or LLMs, form a separate option: you give the model an invented biography and ask it to answer from that perspective.

The table puts each approach against the same criteria. Its evidence column records a reason to retain human research even where simulation offers a useful rehearsal. Project examples describe those engagements, rather than a delivery promise for your study.

Sources: our earnings-call and consumer-goods case studies; method and January 2026 evaluation; Nexus Expert Research’s January 2026 guide above; Drive Research’s 26 November 2025 cost guide (opens in a new tab). The cost request is a procurement recommendation, not a price estimate.

Biography-prompted LLMs suit a low-stakes question where an aggregate indication is enough. The distinction matters when you need to understand why stakeholders disagree. An invented biography does not provide the observed relationships and prior behaviour used to construct our personas.

Artificial Societies’ January 2026 Survey Evaluation Report records self-contradiction rates of up to 35% for biography-prompted LLMs, under 2% for its personas and about 9% for human panels. This measure checks whether answers contradict one another, rather than whether statements about the world are true. A buyer examining a persona’s reasoning needs that distinction alongside the distribution figure.

The Market Research Society’s 2024 Delphi Group report (opens in a new tab) discusses related problems in work by Angelina Wang and colleagues: demographic prompts can produce stereotypes and flatten differences within groups, the representational problem our article on AI persona bias examines. We would use a biography prompt to explore a tentative hypothesis. For a consequential stakeholder decision, we would ask for observed audience data and checks on the relationships between answers before treating the output as research evidence.

Audience simulation’s cost case is concrete in Artificial Societies’ consumer-goods study: almost 80% savings compared with a human panel with half the sample size. The study used 1,498 simulated health-conscious consumers to compare five product concepts and select two for further work. It reports 48 hours of fielding across two research rounds. That clock covers fielding, rather than the elapsed project including society construction.

The design explains what the budget bought. Artificial Societies compared the concepts in a five-armed randomised controlled trial, with separate experimental groups to avoid cross-contamination, then explored product names and features. The client could examine verbatim reasons alongside selections. This is a worked comparison of research scope and cost, not a general discount to apply to a fieldwork quote.

For a traditional baseline, Drive Research’s November 2025 guide (opens in a new tab) prices an online survey with 400 completed responses at $5,000 to $15,000+. That is the agency’s range. We would compare proposals for the decision you actually face: the audience, alternatives to test and evidence needed at the end. A straightforward consumer survey can justify fieldwork; a confidential experiment across competing concepts gives simulation a different job.

Audience simulation changes who sees an unannounced strategy. Human participants see the material put to them; a synthetic run exposes it to no recruited participant. This distinction matters before a policy announcement or an earnings call, when showing a draft can itself affect the audience. It is a feature of the research design, rather than a guarantee about security.

In late 2025, Artificial Societies’ Teneo engagement tested six technology narratives across societies representing consumers, industry peers and policymakers. The study records 189,756 responses and delivery in less than three weeks, including construction of the societies.

Our ability to model another audience depends on diverse observations of it, whether public or client-provided. Where those observations are absent, research must first establish them; an accuracy benchmark on an observed population cannot establish performance for an unobserved one. For a hard-to-recruit audience with a usable record of behaviour, simulation gives you a way to investigate responses without recruiting its members.

Synthetic research needs evidence at the level of the decision. Artificial Societies’ September 2026 validity framework sets out eight tests covering whether the stimulus caused the response, whether questions measure the intended attitude and whether findings hold for real populations. Matching the proportion selecting each answer leaves a separate issue: whether the same personas hold believable combinations of views. How to evaluate a synthetic research vendor turns those tests into questions you can put to any supplier, including us.

Peer review covers the group behaviour: He, Wallis, Gvirtz and Rathje’s December 2024 paper in the British Journal of Psychology (opens in a new tab) supports the network modelling and says nothing about survey accuracy, as the science page explains.

Human data deserves inspection too. Pew Research Center’s February 2020 study (opens in a new tab) found that 4–7% of interviews in online opt-in panels came from bogus respondents. Among those respondents, 84% passed a trap question and 87% passed a speeding check. Ask to inspect respondent-level data in either method. A human recruitment label is insufficient evidence of response quality, just as a fluent persona answer is insufficient evidence of validity.

A simulation and fieldwork budget should follow the decisions left to resolve. We recommend using simulation to compare alternatives and identify hypotheses, then paying for targeted human research to confirm findings that determine the final commitment. Artificial Societies’ consumer-goods study illustrates the narrowing step: five concepts became two candidates for further iteration. It does not establish that a taste test or a physical usability study has been completed.

You can also begin with traditional research as a baseline, then use simulation to test confidential narratives against that understanding of the audience. Before commissioning either, specify which findings will need human confirmation. We would reserve the fieldwork budget for those questions and scope the simulation around the alternatives you need to explore before choosing what to put before people.

We publish this comparison using our method, evaluation and case studies, agency guides, the cited academic paper, Pew research and the Market Research Society report. Source dates appear below. We compare research approaches rather than rank vendors; engagement outcomes remain specific to their projects, and agency estimates remain attributed to their publishers.

Information is believed accurate as of 17 September 2026. Third-party trade marks belong to their owners; inclusion implies no affiliation or endorsement. Send corrections to support@societies.ai so we can amend the comparison.