Synthetic panel: what it is, how big it is, and when you can trust it
Updated October 2026
A synthetic panel is a group of AI-modelled respondents that answers research questions in place of recruited people. You choose the audience and the panel size, show it a film, concept or message, and get survey-style answers in minutes rather than weeks. It is a pre-test that narrows your options before you pay for a real panel, not a replacement for one.
Cavefish's synthetic audience is a panel of modelled UK viewers, sampled from census data, that watches your campaign film and says how it made them feel, moment by moment, and what they will remember now and the next day.
Why marketing teams are looking at synthetic panels
Traditional panels and focus groups are slow and expensive, so most creative gets tested late, on few options, or not at all. Synthetic panels change the economics: any number of variants can be tested the same day, before media budget is committed. Adoption is already wide. Qualtrics research from 2025, cited by Discuss.io, found 73 percent of market researchers had used synthetic responses at least once, and Qualtrics launched synthetic panels for US consumers in March 2026.
The same sources agree on the catch. A synthetic panel is only as good as the model and data behind it, so a CMO needs to know what the panel can and cannot tell them before a decision rests on it.
How big is a synthetic panel?
Unlike a recruited panel, size is a setting, not a recruitment problem. In Cavefish's synthetic audience you choose 50, 100, 200 or 500 respondents for a panel read, with 200 as the default. The “where does it fit” read runs 500, 1,000 or 2,000 people. A segment is only reported when it holds at least 30 respondents, and bootstrap intervals are shown on every figure.
| Respondents | Typical use | What to expect |
|---|---|---|
| 50 | Quick gut-check on one film or message | Directional only; segment cuts are hidden below 30 respondents |
| 100 | A first read across a broad audience | Enough for age or region splits in large segments |
| 200 | The default panel for a film or concept pre-test | Segments such as 18-34 against 55+ can be reported |
| 500 | Larger panel for finer segment cuts | More groups reach the 30-respondent reporting floor |
| 500 / 1,000 / 2,000 | Where does it fit: the film tested against the whole UK population | Each person is placed in the channels they use, so it shows who the film is for and where to run it |
More respondents narrows the sampling error of the model. It does not remove the model's own bias, which is why the read is labelled modelled and validated against real viewers for large decisions.
Synthetic panel vs real panel vs focus group
| Synthetic panel | Real survey panel | Focus group | |
|---|---|---|---|
| Who answers | Modelled respondents sampled from census distributions | Recruited people | 6 to 12 recruited people |
| Speed | Minutes to hours | Days to weeks | Weeks, including recruitment |
| Sample size | Chosen by the researcher: 50 to 2,000 | Set by quota and budget | Small by design |
| Emotional response | Modelled; weakest area for AI respondents | Self-reported | Self-reported, social pressure |
| Best use | Pre-test, screening, narrowing options | Confirming the decision | Exploring why, in depth |
| Evidence status | Labelled modelled; validate before large spend | Measured | Measured, small sample |
Where synthetic panels go wrong, and what we do about it
Independent testing is blunt. Kantar's experiments found a general-purpose model strongly positive compared with human respondents, much further off on emotional reactions, and not sensitive enough for sub-group analysis. Nielsen Norman Group reran three of its own user studies with AI participants and found the answers too shallow to replace real people. These are the known failure modes, and the design of Cavefish's synthetic audience targets each one.
The recall interview never sees the film. It sees only each respondent's own memory trace: what they noticed, weakened by attention, forgetting and a clutter reel of other ads.
Respondents answer in free text first, and the text is mapped to a scale by semantic similarity rating, giving a spread of answers rather than a single number.
Rejecters and lapsed customers are part of the sample, and the framing tells respondents they can say an ad was boring.
Respondents are drawn from joint census distributions, then given a full life backstory. There are no ethnicity, religion or health selectors.
Every read is labelled modelled, not measured, and states its calibration status. Accuracy against real panels is being established and will be published.
How the Cavefish synthetic audience works
- Choose who is watching. Region or nation, age band, gender, social grade, children in the household, urban or rural, working status, media diet, category use and brand relationship (customer, lapsed, considering, rejecter, unaware). An audience can also be a profession, in the UK or abroad.
- Respondents are sampled, then given a life. People are drawn from ONS Census 2021, Scotland's Census and NISRA tables under the Open Government Licence, then written a backstory: job, household, money pressures, interests and viewing situation.
- They watch the film. Films of up to 10 minutes are read from what is measured in the film (faces, voice, words, on-screen text, cuts) plus a read of the frames. Each respondent says what they felt and whether they were watching, half-watching or scrolling past, depending on the channel.
- They are interviewed about what they remember. Unaided ad and brand recall, who the ad was for, the message they took away, and the same questions again the next day, from each person's weakened memory trace only.
- You get a decision-ready read. Felt-emotion curves, attention and drop-off points, recall and correct branding, results by segment, and three to five fixes. Share it with your agency or board by read-only link.
Modelled audience, not measured viewers. Every screen and export carries that label. The read is currently marked as not yet checked against real panels, so use it as a pre-test and confirm large decisions with real viewers. Cavefish pairs it with EchoDepth facial coding of real viewers, and reports modelled and measured results side by side, never merged into one figure.
Questions CMOs ask about synthetic panels
What is a synthetic panel?
A synthetic panel is a group of AI-modelled respondents that answers research questions in place of recruited people. Each respondent is a generated persona, built to match a target audience, that reacts to a stimulus such as a film, concept or message and answers survey-style questions. It is used for fast, low-cost pre-testing before a real panel or focus group.
How big is a synthetic panel?
It is set by the researcher rather than by recruitment. Cavefish's synthetic audience runs a panel of 50, 100, 200 or 500 modelled respondents (200 by default), and a 'where does it fit' run of 500, 1,000 or 2,000 people to read the whole UK population. A segment is only reported when it holds at least 30 respondents.
How accurate are synthetic panels compared with real panels?
It depends on the question, the model and the data behind the personas. Independent tests are mixed: Kantar found general-purpose models overly positive, weak on emotional reactions and unreliable for sub-groups, and Nielsen Norman Group found AI-generated users too shallow to replace real participants. Cavefish labels its read as a modelled audience, not measured viewers, and states that it is not yet checked against real panels.
Can a synthetic panel replace focus groups or survey panels?
Not for decisions that carry large budget or risk. Most research bodies and vendors position synthetic panels as a pre-research layer: screen ideas, sharpen the questionnaire and prioritise what to test, then confirm the shortlist with real people. Cavefish sells its synthetic audience the same way, as a pre-test that comes before a real panel.
What is the difference between a synthetic panel and a digital twin?
A synthetic panel is a set of generated respondents sampled to mirror a population. A digital twin is built from one organisation's own research, such as its interviews and survey data, and answers within that evidence. Cavefish's synthetic audience is a population panel: respondents are sampled from UK census distributions rather than from one company's data.
What are the limits of synthetic respondents?
The main failures are agreeing too much, clustering around the average, stereotyping thin personas, remembering too much, and being unreliable for small sub-groups. Cavefish's design answers each: the recall interview never sees the film, respondents answer in free text that is mapped to scales, personas carry a full backstory, and results below 30 respondents are hidden.
What can a CMO use a synthetic panel for?
Pre-testing a campaign film or message before media spend, comparing creative routes, checking who a film fits and which channel suits it, and forecasting what viewers will remember and whether they will link it to the brand. It is a screening tool: it narrows many options to a few before real-viewer testing.
Can I share the results with my agency or board?
Yes. A synthetic audience report can be shared by read-only link with no account needed. The shared view shows the summary and where the film fits, and does not show the modelled people's individual profiles.
Sources
- Kantar: What is synthetic sample, and is it all it's cracked up to be?
- Discuss.io: Are synthetic respondents reliable for market research?
- Qualtrics: new market research capabilities, X4 2026
- Nielsen Norman Group: Synthetic Users: If, When, and How to Use AI-Generated “Research”
What is a synthetic panel and can marketers trust it?
A synthetic panel is a group of AI-modelled respondents that answers research questions in place of recruited people, so a marketing team can test a film, concept or message in minutes. Marketers can trust it as a pre-test that narrows many options to a few, and should not use it alone for large budget decisions, because independent tests find AI respondents too agreeable, too average and unreliable for sub-groups. Cavefish's synthetic audience runs 50 to 500 modelled UK respondents per panel, or up to 2,000 to read the whole population, shows recall and emotion by segment, and labels every result as modelled rather than measured.
See a synthetic audience read on your own campaign
Send a cut of your next film. We return how a modelled UK audience felt, what they remembered, and where it fits.
Request a synthetic audience read