AI persona market research: simulation explained

An AI persona, one kind of synthetic persona, needs evidence behind its answers. Our method and evaluation distinguish personas grounded in real observations from biography prompting, where a large language model receives an invented personal description. We use observed material to model belief systems, then validate personas against audience benchmarks and prior behaviour before admitting them to a society. A convincing biography alone is a poor basis for committing to a product concept.

Across 1,000 surveys, Artificial Societies’ personas match human opinion distributions with 86% accuracy, where the human ceiling is 91%; the measure and its limits are set out in How accurate are AI personas?.

For a buyer, this establishes a measured comparison with human opinion. Distribution overlap does not by itself establish which people hold which combinations of attitudes; our September 2026 validity framework treats those relationships as a separate test. We would examine both before using a segment’s reaction to guide a decision.

Persona construction starts with published research, public commentary and reviews, alongside anonymised social data and public records of voting and investment. Our method page names LinkedIn, X, Reddit, Facebook, YouTube, Instagram and Amazon among the sources. For the consumer-goods product study, Artificial Societies and its social listening partners gathered over 100,000 anonymised profiles from Instagram, Facebook, TikTok, Reddit and X, plus health forum discussions and reviews of health-adjacent products.

Our privacy policy governs collection. For public-source persona data, we collect information made manifestly public on open social media or professional networks: professional roles, public posts, profile information and aggregate public connections or group memberships. We exclude private communications, restricted accounts and children’s data, as well as financial information, health data and other special-category data. Public availability does not remove those exclusions.

Anonymisation has an explicit exception. Under Article 6(1)(f), the legitimate-interest basis, our privacy policy permits public figures to be processed without pseudonymisation where identity matters to the research. Their identity remains attached. The work concerns their public role, and the persona must be labelled as an AI simulation rather than a direct representation of their actual views. Personas operate in a closed environment; our contracts prohibit using outputs to target, contact or make decisions about identifiable individuals.

Our Behavioural Analysis Engine combines observations to model how an individual reasons. The psychometric step means reconstructing a belief system from those observations so the persona can respond to new information. The product overview names this stage Persona Genesis; our method page places validation against prior behaviour and audience benchmarks before society creation.

The construction depends on the audience. In the consumer-goods study, Artificial Societies combined profiles with similar demographic and psychographic traits into lookalike consumers. The resulting society represented 1,498 health-conscious consumers, segmented by health-consciousness and health literacy. The earnings-call study instead matched personas one-to-one to individuals through public data and social activity. Grounding in real people therefore need not mean copying a single profile into each persona.

Your own qualitative research or survey data can calibrate the personas to the audience you need. This first-party input is optional. We keep it separate for each engagement, store it in the EU and never use it to train our models, as our method page states. For customer data, you remain the data controller and we act as the processor under our privacy policy.

Society Assembly connects personas through real relationships and patterns of influence, after grouping them by shared characteristics. Each society contains 12 to 3,500 personas, according to Artificial Societies’ method and evaluation. The network lets a study examine opinion in its social context: the Simulation Engine combines individual belief systems with the relationships through which people influence one another. Why opinion spreads through networks explains what an audience of unconnected personas leaves out.

A decision can involve distinct audiences. The hyperscaler earnings-call study built four societies around the company’s stock, with the membership defined by each audience’s relationship to it.

Source: Artificial Societies, hyperscaler earnings-call case study.

That design made analyst questioning a defined part of the research, alongside investor reaction and press coverage. We would define membership before discussing audience size. A society representing the analysts who question management has a different job from a population matched to census distributions, as used in our Teneo engagement.

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.

The Simulation Engine starts each run with context based on the audience’s current situation, then presents the personas with stimuli or scenarios. It combines their belief systems with network dynamics to model opinion at individual and population level. Our research methods include surveys and experiments that compare alternative exposures using copies of the same participants.

For the consumer-goods study, Artificial Societies established baseline opinions, then ran a five-armed randomised controlled trial: an experimental comparison of five concepts without cross-contamination between them. We subsequently explored product names and features. The client selected two concepts for further iteration. Its Head of Innovation, Global Consumer Goods Conglomerate, described the value:

“Without Artificial Societies, we wouldn't have the quantitative insights that could help us narrow down product concepts effectively.”

The crisis study used another control: we erased the societies’ memories of messages between tests. This separated a message’s effect from exposure to an earlier message or the order of presentation. You can test alternatives without allowing one experimental exposure to alter the audience for the next.

A simulation returns results you can inspect at market, segment and persona level. Every survey question requests a verbatim reason for the selection by default, as our consumer-goods study describes. You can examine the explanation behind a choice alongside the distribution of choices, rather than treating a favourable total as sufficient evidence to proceed.

The Teneo engagement, conducted in late 2025, returned 189,756 unique responses from 30 questions across six technology narratives. Teneo received a written report and access to examine approval scores, emotional sentiment and verbatim reactions. We identified the strongest overall narrative and refined messaging by audience segment. The deliverable connected a recommendation to responses the client could inspect.

In the earnings-call study, the pre-call briefing challenged management’s expectation that a margin reset would dominate. Simulated analysts concentrated on demand durability and financing commitments; the study reports that the live questions followed that pattern. The client also had access to rerun scenarios and revise script tests as the numbers changed. That account is evidence of a useful rehearsal, separate from the survey evaluation’s distribution benchmark.

AI persona simulation fits consequential concept and positioning decisions, particularly when testing material with a live audience would compromise the work. In the earnings-call study, selective-disclosure constraints prevented testing unreleased figures and scripts with real investors. We kept the drafts and numbers inside our private, secured environment. Our focus is the decision about what to say or develop, rather than low-stakes advertising execution.

Allow for society construction when planning the work. Artificial Societies’ published case studies record whole engagements from 48 hours to three weeks. The consumer-goods study records 48 hours of fielding across two research rounds, with under 24 hours per round on an already-built society. Those are fielding durations, so they should not be used as a promise for building and delivering a new study.

We would bring the decision and the available audience observations into the scoping discussion together. Public material and your own research provide the grounding; where neither offers diverse observations, there is nothing to ground a persona in, and the audience sits outside what the method can model. An observed audience that is hard to recruit is a different proposition from an audience for which there is no evidence to model.

AI persona simulation supplies evidence about opinion for decision rehearsals. It inherits the gap between stated intentions and actual behaviour that our September 2026 validity framework discusses. We would use concept reactions to guide further iteration, as the consumer-goods client did, and treat proof of what people actually do as a separate question your decision may still require.

Edoardo Chidichimo and Felix Wallis address risks inside the research in that September 2026 framework. Language-model personas can agree with a question’s implication without being persuaded by its evidence. We test changes to question wording and response order, as well as changes to the strength of a claim. These checks distinguish a response to the stimulus from a response to how the question was asked.

The framework also examines whether questions measure the intended attitude and whether results hold for real populations. We would require that broader assessment when a decision depends on relationships between beliefs: matching the number of favourable answers cannot establish who is favourable for which reasons.