What is a synthetic audience? A field guide to AI-persona research

A synthetic audience is a population of AI personas built to statistically mirror a real one — the same distribution of ages, incomes, geographies, education levels, political leanings and buying attitudes as the audience you actually need to reach. Instead of recruiting eight people into a room, you generate eight thousand personas that match your census data, and run your message, survey, ad or website through all of them at once.

The idea has moved fast from research curiosity to industry practice. Qualtrics now fine-tunes synthetic-respondent models on decades of human survey data, and a Forbes Technology Council analysis describes synthetic personas as one of the most significant shifts in how insights teams work.

The evidence: simulated people track real ones surprisingly well

The landmark study came from Stanford, Google DeepMind, Northwestern and the University of Washington: researchers built generative agents of 1,052 real individuals, each grounded in a two-hour interview, then asked the agents the same surveys, personality assessments and economic games as the humans they were modelled on.

The commercial research world is seeing similar signals — synthetic panels reproducing real survey toplines with high fidelity when they're grounded in real distributional data rather than a model's vibes. That grounding is the whole game, which brings us to what separates a synthetic audience from "I asked ChatGPT to pretend to be my customer."

What makes an audience synthetic — and not just a chatbot in a costume

Prompting one model to role-play "a busy mum from Brisbane" produces fluent, confident and statistically meaningless output. A synthetic audience is different in four specific ways:

• Stratified, not improvised. Personas are generated to match real distributions — at Lanice AI, 21 census-grade demographic columns plus ~70 attitudinal dimensions, stratified with a documented largest-remainder algorithm against your own data, or built from a plain-language description ("adults in Sydney aged 21–38 who shop online weekly and earn $50k–$100k").
• Diverse by construction. Research keeps showing that when you simply ask an LLM for "diverse personas," the output collapses into a narrow cluster of stereotypes. Assigning demographics and personality traits before generation — and sampling attitudes with a deterministic, population-calibrated sampler — is what prevents that collapse.
• Behavioural, not just verbal. Real audiences don't only answer questions; they scroll past things, pile onto comment threads, share, ignore, and abandon shopping carts. A synthetic audience worth the name lives in a functioning social world and browses real websites, so you observe behaviour, not just stated opinion.
• Auditable. Every persona action should carry its reasoning. If you can't trace why the simulated rural independent rejected your message, you have a black box, not an instrument.

The comparison every insights team is running

Traditional qualitative research isn't broken, but its economics are brutal. A single professionally-run focus group costs $8,000–$15,000 per session once you count facility, recruiting, moderation and incentives — and a standard multi-group study runs 6–12 weeks from brief to report. Head-to-head analyses of AI research versus focus groups keep landing on the same conclusion: the trade isn't either/or. Synthetic audiences do the wide, fast, cheap exploration — every variant, every segment, every "what if" — and human research validates the decisions that matter most.

Test every variant synthetically. Validate the winner with humans. Stop guessing about everything in between.

What synthetic audiences are being used for today

• Message pre-testing — scoring resonance, backlash risk and claim weaknesses on every draft, not just the finalists.
• Synthetic polling — population-scale surveys with cross-tabs and significance testing, where every respondent explains their reasoning.
• Campaign simulation — multi-phase launch storylines with competitive counter-moves and environment shocks played out before launch week.
• Agentic UX research — demographically-grounded personas driving a real browser through your site, narrating why they click and why they leave.
• Academic research — replicable social-dynamics experiments that would be unethical or impossible to run on real platforms.

If you want the deeper mechanics — how stratification, personality vectors and decision traces fit together — the platform overview walks through the full pipeline.