How our synthetic audiences work
Every number behind our audiences is listed here — what comes from official statistics, what is modelled, how feedback is generated, what we have validated, and what this tool is honestly not.
1. How audiences are built
Audience members are generated by statistical sampling from official population data — not by AI. Each person's age, gender, region, income, education and household situation is drawn from census-derived probability distributions, so the composition of every audience mirrors the real country: the right share of each region, realistic income spread, era-accurate and ethnically representative names.
Large populations are represented by a panel of up to 20,000 fully generated individuals. When you run a focus group or survey, respondents are selected by stratified sampling — the same proportional-quota method used by professional pollsters — so a 12-person group or 500-person survey mirrors the audience's make-up rather than being a random scrape.
2. Data sources — consumer audiences
| Market | What | Source | Vintage |
|---|---|---|---|
| UK | Age structure, gender, regional population | ONS mid-2024 population estimates (table MYE2) | Published Sep 2025 |
| UK | Personal income distribution | HMRC Survey of Personal Incomes 2023–24 (Table 3.1a) combined with ONS adult population; non-taxpayers assigned below the personal allowance | 2023–24, updated Apr 2026 |
| UK | Regional income skew | ONS Annual Survey of Hours and Earnings 2025 (median full-time pay by region) | 2025 provisional |
| UK | Education by age | ONS Census 2021 (England & Wales), highest qualification by age; postgraduate split estimated from DfE Graduate Labour Market Statistics 2024 | 2021 / 2024 |
| UK | Household situation by age | ONS Census 2021 (RM057), Families & Households 2024, Young Adults Living with Parents dataset | 2021–2025 |
| UK | Name-pool ethnicity weights | ONS Census 2021 ethnic group (England & Wales), adjusted for Scotland Census 2022 and NISRA Census 2021 | 2021–2022 |
| US | Age structure, gender, education, living arrangements, race/ethnicity | US Census Bureau, American Community Survey 2024 1-Year (tables B01001, B15001, B09021, B03002) | 2024, released Sep 2025 |
| US | Census-division population shares | US Census Bureau Population Estimates, Vintage 2025 | July 2025 |
| US | Household income distribution and regional skew | ACS 2024 1-Year (B19001 household income; multipliers derived from B19013 division medians) | 2024 |
Data sources — business (B2B) audiences
| UK | Business counts by size (75.1% have no employees) and industry mix | Dept for Business & Trade, Business Population Estimates 2025 (Tables 1 & 5) | Start-2025, published Oct 2025 |
| US | Firm counts by size (82.6% nonemployer) and industry mix | US Census Bureau: Nonemployer Statistics 2023 + Statistics of U.S. Businesses 2022 | 2022–2023, released 2025 |
Where a published table doesn't exactly match our bands, figures are derived from the published numbers and marked as such inside the datasets themselves.
3. What is sourced vs what is modelled
Sourced from official statistics
Ages, gender, regions, incomes, education, household situations, ethnicity-weighted names, business sizes, industries — the composition of who is in the room.
Modelled archetypes (authored)
Personality traits, shopping styles, decision drivers, dealbreakers, KPIs and pain points. No public dataset measures these; they are research-informed archetypes, statistically weighted, and we say so rather than inventing a citation.
4. How feedback is generated
Each sampled respondent is role-played by a large language model that is given the person's full profile and explicit instructions to be honest, not polite: sceptical people are harsh, dealbreakers cannot be argued away, purchase intent reflects the person's actual budget and needs — not how slick the ad is — and respondents are told that real focus groups skew indifferent and critical.
Aggregate scores, sentiment splits and segment breakdowns are computed in code from the individual responses. Only the executive summary is AI-written, from the computed aggregates and representative quotes.
5. Validation
We test the simulator against falsifiable checks. Latest results:
Reliability
The same ad was shown to two independent 50-person samples from the same audience.
Mean purchase intent 4.22 vs 4.54 (within sampling error for n=50); clarity 7.84 vs 7.92. The simulator gives a stable verdict, not a random one.
Sensitivity
A clear, benefit-led ad was tested against a deliberately weak, jargon-filled version of the same product.
Intent 4.22 vs 2.96; clarity 7.84 vs 4.92; negative sentiment 30% vs 46%. Top objections to the weak ad correctly named the problems: "too technical and unclear", "feels gimmicky".
Segment sanity
A river-cruise offer aimed at retirees (coach pickup, no apps, book by phone) was shown to a full-age-range panel.
Purchase intent rose monotonically with age: 1.83/10 among 18–24s (83% negative) to 5.23/10 among the 65+ (46% positive). Demographics demonstrably drive the responses.
6. Honest limitations
- Directional, not calibrated. The simulator reliably ranks creative, surfaces objections and shows which segments respond — but a score of 4.2/10 is not a forecast of a real-world conversion rate.
- AI voices have AI biases. Role-played personas can lean on stereotypes (an 80-year-old may sound slightly more tech-averse than real 80-year-olds) and produce fewer truly left-field reactions than real humans.
- Use it before research, not instead of it. The right role is killing weak ideas and sharpening strong ones in minutes for pennies — before you spend on media or commission human research.
AI Persona — synthetic audience research. Sources last refreshed June 2026.