Correction, 2026-10-02: this page still stated some judgments as if they had been measured: that early triage is "the best-fit use", that ranking variants is "a safer use" than predicting numbers, that a synthetic read "is better than a blind guess", that data-grounded agents are "the stronger approach", and that synthetic audiences "will not reliably catch a cultural misread". These are now advice, or tied to the research that supports them. The Verasight errors of 27% to 33% are the four worst-performing examples the report names, not typical results (its overall error was 14.5 percentage points), and our line that synthetic audiences regress toward the most probable answer is now tied to that report's finding of a "systematic tendency to regress predictions toward the mean". An unsourced line about multi-agent systems is gone, the summary of the Paglieri paper now follows its wording (approaches "often" favor the most probable answers), and the claim that the tails are "where breakout creative often lives" is now a caution rather than a fact. The Park paper now carries its current title, and the page now says plainly that 8frame generates the variants but does not run synthetic audiences.
Correction, 2026-10-01: an earlier version described the Stanford generative-agents study as reaching 85% accuracy, 14 points better than persona-only agents. The current version of the paper reports 86% for agents built from interviews and surveys against 74% for agents given only demographics. It also said credible platforms agree with past research 80 to 95% of the time and that the category drew serious funding in 2026, which we could not source, and it overstated what a 2026 persona paper found. The research is now cited with links.
A synthetic audience is a panel of AI-generated personas, built to simulate the responses of a real target segment, that you can query the way you'd query a survey panel or focus group, except no humans are involved.
You describe a target, women 25 to 40 who buy premium skincare, for example, and the system produces or retrieves a set of personas matching that profile. You then show those personas an ad, a message, or a landing page and collect predicted reactions, rankings, and preferences. The appeal is speed and cost compared with recruiting a real panel. The catch is that a synthetic audience predicts responses rather than measuring them, and understanding that difference is the whole game.
How a synthetic audience is built
There are two broad construction methods, and many platforms blend them.
Persona-only generation. The system prompts a large language model to role-play personas matching a demographic and psychographic profile, then queries those personas. This is fast and cheap but the least grounded, because the model is inventing plausible people from its training data rather than from data about real respondents.
Data-grounded agents. The other approach grounds each agent in real interview or survey data about actual people, so the simulated responses trace back to observed human answers rather than pure invention. A study by Joon Sung Park and colleagues took this route with a national sample of 1,052 Americans. In the current version of the paper, on held-out General Social Survey questions, agents built from interviews and surveys reproduced participants' answers at 86% of the consistency people show with their own answers two weeks later, compared with 74% for agents given only demographics.
If a vendor quotes an accuracy figure, ask what it was measured against: matching past aggregate patterns is not the same as predicting the response to your specific new creative.
When you use a synthetic audience
Early creative triage. When AI generation hands you 40 or 50 variants, a synthetic audience can rank and cluster them cheaply, to help you pick the 10 to 15 that get real test budget. Treat the ranking as a shortlist, not a result.
Concept and message screening. Checking whether a value proposition reads as relevant, which of several angles resonates, or how a concept ranks against alternatives, before committing production or media.
Fast, cheap directional reads. When a real panel is too slow or expensive for the decision in front of you, a synthetic read can give you a direction. Treat it as a hypothesis to confirm, not a measurement.
Validity limits
Synthetic audiences fail in specific, documented ways. In a January 2026 Verasight report, LLM-imputed survey answers did badly on multi-answer questions: "Very often, the LLM imputer would only generate responses for one or two options." The worst-performing questions were about personal experiences and behavioral preferences that "lack strong demographic predictors", with mean absolute errors of 27% to 33% on the four examples it names; across all response options the error averaged 14.5 percentage points. The same report found "a systematic tendency to regress predictions toward the mean".
Paglieri and colleagues (2026) note that most persona-generation approaches need detailed data about the target population and often prioritize "density matching (replicating what is most probable) rather than support coverage (spanning what is possible), leaving long-tail behaviors underexplored." A review of 63 studies from 2023 to 2025 by Batzner and colleagues found that only 35% discussed whether their personas were representative, and called for enhanced ecological validity.
The practical upshot: if a panel pulls answers toward the most probable one, it can underweight unusual reactions, and an unusual creative idea may be exactly what you are testing. Don't count on it to catch a cultural misread that a real person would flag. Use it as a filter with blind spots, not as a replacement for in-market results.
Examples
Shortlisting a variant batch (illustrative). A growth team generates 50 ad variants for a new serum. They run all 50 past a synthetic audience calibrated to their buyer, which ranks the hooks and clusters them by predicted appeal. The bottom 35 get cut, the top 15 go to a small-budget Meta test. The synthetic step didn't pick the winner; it made the real test affordable by narrowing the field.
Screening a positioning line (illustrative). A B2B brand tests four value-proposition framings against a synthetic panel of IT decision-makers. Three cluster tightly in predicted relevance and one lags badly. The brand drops the laggard and takes the other three into live testing rather than betting the campaign on the synthetic ranking alone. The panel narrowed the choices; the market confirmed them.
FAQ
Can a synthetic audience replace a real test? No. Use it to narrow a batch, then confirm with real spend or real respondents.
Does 8frame run synthetic audiences? No. 8frame generates the image and video variants; the synthetic panel is a separate tool.
What does a 50-variant batch cost to generate on 8frame? Fifty Nano Banana 2 stills at 1K are 550 credits (about $5.50 at pack rate); video variants cost more, for example 86 credits for a 5-second Kling v3 Standard clip with sound, made from a start image.
Related concepts
- Pretesting ads with synthetic audiences is the full workflow and skepticism guide: generate wide, pretest to shortlist, spend to confirm, and calibrate the panel against your real outcomes over time.
- AI ad variant testing workflow produces the wide batch a synthetic audience is meant to filter. Without volume feeding in, there's nothing for the panel to triage.
Sources
- Park et al., "LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals" (first posted in November 2024 as "Generative Agent Simulations of 1,000 People"; current version revised 2026-06-28; 1,052 participants; on held-out General Social Survey items, 83%, 82% and 86% of participants' two-week test-retest consistency for interview-only, survey-only and combined agents vs 74% for demographics-only): https://arxiv.org/abs/2411.10109, checked 2026-10-02
- Verasight, "Can Large Language Models Replicate Survey Data Across Topics?" (G. Elliott Morris and Benjamin Leff, 2026-01-13; multi-response answers limited to "one or two options"; worst-performing examples named at 33%, 30%, 29% and 27% mean absolute error; overall 14.5 percentage points; "systematic tendency to regress predictions toward the mean"): https://www.verasight.io/reports/synthetic-omnibus-survey, checked 2026-10-02
- Paglieri et al., "Persona Generators: Generating Diverse Synthetic Personas for Arbitrary Contexts" (submitted 2026-02-03): https://arxiv.org/abs/2602.03545, checked 2026-10-02
- Batzner et al., "Whose Personae? Synthetic Persona Experiments in LLM Research and Pathways to Transparency" (63 peer-reviewed studies, 2023 to 2025; only 35% discussed representativeness; calls for enhanced ecological validity): https://arxiv.org/abs/2512.00461, checked 2026-10-02
- Nano Banana 2 and Kling v3 prices, Kling v3's start-image requirement, the credit value at pack rate, and the absence of any synthetic-audience tool: the 8frame app, checked 2026-10-02
A synthetic audience needs variants to judge. Generate them on 8frame, then read the full synthetic pretesting workflow to filter and spend safely.
Try it without signing up: the 8frame canvas opens on a real board. The canvas is free and unlimited; generation is paid from $19/month.