AI focus groups compared: people and simulated audiences

AI synthetic focus groups use simulated participants drawn from a synthetic audience. In our method, observations of real people ground the personas, and relationships connect them into a society. The distinction matters when your question concerns how people influence one another: an independent interview and a discussion between connected participants ask different things of a model. Our method page describes both the individual reasoning and the network of influence.

Across 1,000 surveys, Artificial Societies’ personas match human opinion distributions with 86% accuracy, where the human ceiling is 91% (Survey Evaluation Report, January 2026); the measure and its limits are set out in How accurate are AI personas?. These figures establish the survey comparison; they should not be applied to the accuracy of a group discussion.

Artificial Societies’ Interactive focus groups bring together 6 to 12 simulated participants in curated discussions, with live thematic analysis, as described on its solutions page. The analysis covers what participants say, their internal monologues and interpersonal dynamics. Internal monologues here belong to the simulated personas; they are part of the model’s account of reasoning, rather than observations of a real person’s private thoughts.

The discussion starts with the society’s construction. We draw on published research and observed material such as online commentary, reviews and anonymised social data. Psychometric methods reconstruct belief systems, and we validate each persona against audience benchmarks and prior behaviour before connecting personas through relationships and patterns of influence. Our method page sets out those stages.

Follow-up interviews are a separate method in our research toolkit. You can question an individual about their survey answer or contribution to a discussion. This gives the researcher a route from a group theme to the reasoning of a participant, without treating the group’s apparent agreement as a sufficient explanation.

A society contains 12 to 3,500 personas, according to Artificial Societies’ method page. The discussion group is a selection within that audience, while surveys can examine the wider society. Keep those counts separate when scoping research: the number in a conversation and the population available for survey work describe different parts of the study.

Audience simulation extends the research you can conduct before sensitive material reaches real people. Traditional focus groups give you direct observation of human interaction. Recruited groups have the advantage when your evidence depends on watching people handle physical materials. Our wider comparison of synthetic and traditional research covers the choice beyond group discussions.

Product descriptions: our solutions and method pages, and Remesh’s homepage (opens in a new tab). The traditional-group comparison concerns direct observation of human interaction.

Traditional focus groups are the right commission when you need to watch people handle physical materials or respond spontaneously to others in the room. Body language and rapport are part of that evidence. We would send you to a recruited group for that task. Krueger and Casey’s Focus Groups: A Practical Guide for Applied Research (opens in a new tab), published by SAGE in 2014, is the practitioner reference for planning this work.

Group composition also deserves attention beyond a total headcount. In their 2011 review (opens in a new tab), Carlsen and Glenton cite research finding that two groups of four participants produced more information than one group of eight. That result concerns human groups. It gives a researcher a reason to question a proposal built around participant volume alone.

For simulation, the corresponding commissioning question is whether observations of your audience exist. Our method needs diverse observations, from public material or your own data. Where those are absent, we would begin by collecting human evidence. Research that establishes who an audience is gives a later simulation something to model; a benchmark from an observed population cannot establish accuracy for an unobserved one.

Recruitment makes human focus groups a commitment of money and calendar time. Drive Research’s November 2025 cost guide (opens in a new tab) gives $7,000 to $20,000+ per group. Its June 2022 timing guide (opens in a new tab) estimates three to five weeks for a project, including set-up, recruitment and reporting. These are an agency’s published estimates, rather than a price or timetable for your study.

Carlsen and Glenton’s 2011 review of 220 published studies (opens in a new tab) found a median of five groups per study, with median minimum and maximum group sizes of five and eight. The number of people heard is therefore a question to inspect in the design, alongside which perspectives the recruitment covers. Their review also examined claims of saturation, the point at which further discussion stops adding relevant information. Of 28 studies that used saturation to justify sample size, 15 did not convincingly describe an iterative process of collection and analysis ending at saturation. The weakness was in the justification reported, not a finding that the discussions had no value.

A dominant participant can also make agreement appear broader than it is by suppressing minority views. Our Interactive focus groups make simulated interpersonal dynamics available for analysis; the researcher still needs to examine how a theme arose. A statement attracting agreement and a belief held independently are different findings to take into a decision.

Audience simulation is worth considering when the material itself prevents recruitment. Our hyperscaler earnings-call case study describes an investor relations team preparing remarks and financial guidance that it could not test with real investors because of selective-disclosure constraints. We built societies around the stock and tested unreleased numbers and alternative scripts inside a private environment.

The simulated analysts focused on demand durability and financing commitments, although the company had expected the margin reset to dominate. The case study reports that questions on the live call followed the simulated pattern. That is a concrete account of a decision rehearsal, without turning the agreement into an accuracy percentage or a finding about focus-group discussions.

Teneo’s engagement shows the breadth available beyond a discussion. In late 2025, Artificial Societies supported Teneo on a project for a major US company, testing six technology narratives across three societies, as the Teneo case study describes. The project gathered 189,756 responses and delivered a report with access to explore approval scores and verbatim reactions.

The Teneo work used surveys and experiments. For your own project, we would distinguish the discussion you want to explore from the wider tests needed to decide between narratives, then agree which outputs will support that decision.

No, not when the evidence you need comes from people in the room: watching them handle a physical product, reading body language and rapport, or studying an audience that has left no observations to ground a persona in. Artificial Societies’ September 2026 post on simulation validity states that “synthetic audiences will never be a true substitute for the people they seek to represent”, and it has published no accuracy figure for group discussions. Simulation belongs before or beside a recruited group: rehearsing material you cannot show real people, or narrowing the options so the human sessions concentrate on the strongest ones. Our article Can AI replace market research? sets out the wider evidence, including the Nielsen Norman Group critique we agree with.

Remesh (opens in a new tab) describes live online groups with real customers or employees and AI analysis of responses. Listen Labs (opens in a new tab) describes AI-moderated interviews with recruited humans. If you need human testimony with AI handling the interview or analysis, those descriptions fit the task. Outset (opens in a new tab) describes both AI-moderated human interviews and individual digital twins, so the proposal should specify which produces your evidence.

Synthetic Users (opens in a new tab) describes simulated individual interviews and surveys using personas generated from an audience description. Our method instead constructs personas from observations and connects them through a social network. Consider that difference when the question moves from an individual’s response to how opinions spread between people.

Synthetic Users’ science post (opens in a new tab) reports 85 to 92% synthetic-organic parity from a late-February-2024 comparison of eight human and eight synthetic interviews about classroom technology with UK school teachers. Its score combines thematic overlap, depth, coverage and qualitative alignment. It measures interview similarity; Artificial Societies’ survey distribution figure measures a different quantity. Comparing the percentages as a league table would discard what each test actually measured.

A convincing transcript gives you something to investigate, but fluency alone does not establish stable beliefs. Our September 2026 validity framework describes how a persona can sound plausible without maintaining a consistent position. It also discusses sycophancy: agreement prompted by the wording of a question rather than persuasion by its substance.

Artificial Societies’ September 2026 validity framework sets out eight tests, which How to evaluate a synthetic research vendor applies to buying decisions. These include rephrasing questions and changing response order, alongside checking whether removing supporting evidence weakens a claim’s reception. For a proposed study, we would ask to inspect how its design distinguishes a response to the message from a response to the way the question was asked.

Our data-quality check addresses human and synthetic data. Internal quality and external accuracy are separate judgements: the check assesses the former and does not establish the latter. Ask for respondent-level evidence when discussing the analysis, so that a coherent theme can be examined against the answers from which it was drawn.

We publish this comparison using our product descriptions, evaluation report and case studies, alongside our published research, peer-reviewed work, focus-group literature, agency estimates and vendors’ own accounts. Source dates appear in the list below. We excluded unsupported product details and offer no ranking based on unlike accuracy measures. This comparison does not report hands-on trials of other products.

Information is believed accurate as of 17 September 2026. Third-party trade marks belong to their owners; naming a vendor implies no affiliation or endorsement. Corrections can be sent to support@societies.ai.