What is audience simulation?

An audience simulation runs in four stages. You decide who belongs in the audience, build a persona for each member from observations of real people, connect the personas, and run a scenario through them. In Artificial Societies’ method, the observations are published research, online commentary, reviews, anonymised social data and public records of voting and investment. A client’s own research is an optional layer. Each persona is checked against audience benchmarks and prior behaviour before it joins a society of 12 to 3,500 personas (method page, read 12 October 2026).

The run itself is an experiment with a baseline and a treatment. For a major consumer goods company facing a crisis, Artificial Societies built 1,174 personas across five societies: policymakers and regulators, media, influencers, consumers and commercial customers (crisis simulation case study). The team surveyed each society’s baseline sentiment and likely purchasing behaviour. It then tested 24 messages and creatives against that baseline and gathered 328,576 responses. Each society’s memory of a message was built up and then erased before the next one, which removed order effects and cross-contamination.

The result can be read for the whole audience, for a segment or for one persona. Artificial Societies’ public Pace the Frontier results, published on 13 September 2026, show the persona level in practice. Their response table holds 4,491 simulated people in three societies. You can filter it by society, role and answer, and each persona gives its reason in its own words.

Audience simulation names an activity. Most neighbouring terms name the thing being simulated or the unit that answers. The distinction that changes what you can ask is whether the simulated people can affect each other.

Sources: AAPOR task force report, May 2026; Argyle and colleagues, Political Analysis, 2023; Peng and colleagues, arXiv 2509.19088, version 5, April 2026. All read 12 October 2026.

A synthetic audience is the closest neighbour: it is the group, and audience simulation is what you do with it. Argyle and colleagues coined silicon sampling in their 2023 paper, Out of One, Many, which conditioned GPT-3 on thousands of socio-demographic backstories from real survey participants. They called the property they tested algorithmic fidelity: the claim that conditioning the model on the right backstory reproduces that subgroup’s answers. Simulated focus groups are one format inside audience simulation, and our comparison of AI focus groups sets them against recruited groups.

The unit decides the question. A sample of independent respondents records first reactions only, so it cannot show where a message spreads or which segment moves first. Artificial Societies connects personas through relationships and patterns of influence for that reason. Its published evidence on spread is still qualitative, and the limits are set out in networked audience simulation.

Use audience simulation where the real audience cannot be asked. The material may not be shareable yet, the people may be unreachable, or the decision may be due before fieldwork could finish. Four uses appear in Artificial Societies’ published case studies.

Message testing

In late 2025, Teneo used Artificial Societies to test six technology narratives for a major US company (Teneo case study). Three societies of more than 5,000 AI personas gave 189,756 responses in under three weeks. The method behind it is explained in message testing.

Crisis communication

In Artificial Societies’ crisis study for a major consumer goods company, 24 messages and creatives were tested in under two weeks (crisis simulation case study). Evidence-backed messaging about the company’s steps to resolve the issue was the best path. Rehearsing a response this way is the subject of crisis communication simulation.

Investor communications

Selective-disclosure rules stopped a hyperscaler’s investor relations team from showing unreleased numbers to real investors. The company expected its margin reset to dominate the earnings call. Artificial Societies’ simulated analysts largely looked past it and pressed on demand durability and financing commitments, and the live call ran the same way (hyperscaler earnings-call case study). Our article on investor communications testing explains how an equity story is tested.

Public affairs

For a global transportation leader, Artificial Societies built 1,500 AI personas of Washington, D.C. opinion leaders. Each persona was matched one-to-one to a real individual using public data and social activity, and together they gave 250,000 responses on competing policy narratives (transport-policy case study). The wider practice is covered in public affairs research.

Accuracy depends on the measure, so a figure needs its measure and its human comparator beside it. Artificial Societies’ January 2026 Survey Evaluation Report puts distribution accuracy at 86% across 1,000 real-world surveys (method and evaluation page). Distribution accuracy is the overlap between simulated and human answer distributions. In the same report, biography-prompted language models reach 67%. The same report puts the human self-replication ceiling at 91%, the rate at which people repeat their own answers when asked again.

The Artificial Societies Benchmark adds harder tests. Its framework paper sets eleven tests across external, internal and construct validity (Chidichimo, Jung, Wallis and He, arXiv 2609.30030, September 2026). Those three ask whether results match real respondents, whether each simulated respondent is stable and coherent, and whether answers relate to each other as people’s do. In a comparison published on 1 October 2026, Artificial Societies ran its own simulation on the benchmark with the same respondents, items and scoring code as nine off-the-shelf LLMs. The run covered 2,796 synthetic respondents and 68,280 answers from seven survey sources.

Source: Artificial Societies, Artificial Societies Is More Accurate Than State-of-the-Art LLMs, 1 October 2026. Read 12 October 2026. The post names the best model only for the first row.

Total variation distance is the share of answers that would have to move for two distributions to match, so 0.063 means about 6 answers in 100. The subgroup row matters to anyone who reads crosstabs, the breakdowns by age, region or education where simulated populations are least reliable. Artificial Societies designed the benchmark and ran the comparison itself. The framework paper and the benchmark are public, so anyone, including a competitor, can apply the same tests to another simulation.

Simulations also miss. In Artificial Societies’ September 2026 misinformation study, factual posts moved the synthetic control group about four points towards vaccination, while 1,000 real Britons showed no significant shift. The same study matched a human effect closely: fact-checks cut belief in false COVID-19 claims by 46% in the synthetic British public, against 45% in the published human experiment. To judge any vendor’s claim, ask which measure it uses and against which human comparator. Ask whether the test used data the model had not seen, which subgroups it covered, and whether it reproduces the size of a change as well as the level of an answer. Our guide on how to evaluate a synthetic research vendor turns those questions into a checklist.

Audience simulation descends from agent-based modelling, in which a computer runs many individual agents and the researcher watches group patterns emerge. Epstein and Axtell’s Growing Artificial Societies (Brookings Institution Press and MIT Press, 1996) is an early example. In its Sugarscape model, agents with traits such as vision and metabolism move across a grid of sugar. Group formation, cultural transmission, combat and trade arise from simple local rules. Those agents are abstract, so they explain how patterns form without modelling any particular audience.

Language models made it possible to give each agent the views and voice of a kind of person. Argyle and colleagues’ silicon samples followed in 2023. In the same year, Park and colleagues’ generative agents paper placed 25 agents in a small simulated town (arXiv 2304.03442). From one agent’s intention to throw a Valentine’s Day party, the agents spread invitations over two days and coordinated to arrive together. In December 2024, a British Journal of Psychology (opens in a new tab) paper co-authored by Artificial Societies’ James He and Felix Wallis reported that 33,299 chatbots in a simulated online society formed communities around a common language. Clustering with similar others in this way is the human tendency called homophily.

Artificial Societies is a London company, founded in October 2024, that simulates how real audiences respond by modelling them as networks of AI personas. It is distinct from the academic term “artificial societies” (agent-based social simulation, Epstein & Axtell, 1996): the company builds simulation software that models how real audiences form, spread and change opinions, so organisations can test consequential decisions before making them. What are Artificial Societies? covers the company and the method.

An audience simulation can fail at three points: what goes in, what is tested, and how the output is used. At the input, the simulation needs observations of the audience. Without a public trace or first-party research, there is nothing to build personas from, and the run has no grounding. In what is tested, a simulation can take messages and creatives but not a physical or sensory trial, such as a taste test, which needs real people handling the real product.

Three limits sit at the output. A simulation reports what people would say, and what people say is often a weak guide to what they later do; simulated answers inherit that intention–behaviour gap from surveys. A simulated result is also not evidence of public opinion. The American Association for Public Opinion Research (AAPOR) task force report of May 2026 says AI-generated cases are not research participants. It requires them to be identified as AI-created in any purported study of public opinion, and says words such as poll and survey imply human respondents.

The third is a question about absolute levels, such as the exact share of a country that believes a claim today. Artificial Societies’ misinformation study compared relative differences between groups rather than absolute rates. Run the simulation to choose between options before the decision, and run a human study when the decision needs the people themselves.

The seven companies below sell software that simulates an audience, and they describe the same idea with different nouns. We publish this page and sell audience simulation, so Artificial Societies is one of the seven. They are listed alphabetically, each in its own description.

Sources: each company’s homepage, read 12 October 2026; Artificial Societies’ method page. Descriptions are paraphrased from those pages and carry no verdict on capability.

The offers differ in what grounds a persona. Electric Twin starts from a client’s own customer data, and Aaru describes real-world behaviour and outcomes as its grounding. Artificial Societies starts from observations of the specific people in an audience. For published validation across vendors, see our guide to public opinion simulation; the AI market research tools guide covers the wider category.