A synthetic respondent is a research subject played by a language model — a written identity, situation, and manner of speaking that a model is asked to hold to while it answers questions about something you show it. The output is transcript-shaped: reactions, objections, hesitations, the occasional refusal to engage with a premise.
The interesting question is not whether that output reads as plausible. It does, immediately and effortlessly, and that is the problem rather than the proof. The question is which decisions it can carry. The honest answer is narrower than the product demos suggest and wider than the flat dismissals allow.
Two different things wear the same name
The first family is statistical. A model is conditioned on real survey respondents’ demographic backstories and asked to answer the next question the way that person would. Argyle and colleagues called the result a silicon sample and reported that a properly conditioned model reproduced response distributions from human subgroups closely enough to be worth studying, a property they named algorithmic fidelity (Political Analysis, 2023). The ambition there is measurement: get a distribution out that resembles the one a survey would have produced.
The second family is qualitative. A panelist is an authored document — a specific invented person with a specific situation, written to be played by a model rather than summarized in a slide — and the output is a set of reactions, not a number. Nobody should claim the transcript is representative. The claim is that the objections in it are objections a real person could plausibly raise, and that hearing them early is cheaper than hearing them late.
Conflating the two is where most of the argument about synthetic respondents comes from. A qualitative panelist library is not making a measurement claim, so measurement critiques miss it; a simulated survey is making one, so it inherits every burden of proof a survey has.
What the published evidence actually says
Three findings are worth carrying into any use of this method.
Conditioning works better than it has any right to, within limits. The silicon-sample result above is real and has been replicated in adjacent settings. A model conditioned on a described person produces language that correlates with what people like that say.
The extremes get exaggerated. Bisbee and colleagues prompted a model to adopt described identities and rate sociopolitical groups; the averages tracked the human baseline reasonably well, but the generated responses overstated the extremity and the certainty of division compared with real people holding the same attributes (Political Analysis, 2024). Synthetic respondents are more confident and more polarized than their originals. In a focus group that shows up as an objection stated too crisply, which is flattering to a research team and misleading about how hard the objection actually is.
Steering does not fix whose views a model defaults to. Santurkar and colleagues measured the alignment between model opinions and those of sixty US demographic groups and found substantial mismatch that persisted even when models were explicitly steered toward a group (arXiv, 2023). A panelist document is a steering instruction. It shifts the output; it does not relocate the model’s center of gravity.
Set against that, the comparison case is not perfect either. Opt-in online panels carry measurable shares of bogus respondents — roughly 4% to 7% depending on the source, answering in patterns that bias estimates rather than merely adding noise (Pew Research Center, 2020) — and the resulting distortions concentrate in exactly the young and Hispanic subgroups researchers most often want to read (Pew Research Center, 2024). “Ask real people” is the right instinct, and it is not automatically the same thing as “get real answers”.
What synthetic respondents are good for
Four uses survive contact with the evidence above, and they share a shape: the synthetic session decides what to do next, and something else decides what is true.
Surfacing objections to positioning. A panelist written with documented distrusts will name the objection your copy invites. Whether that objection is widespread is a different study; that it exists at all is worth knowing before the launch.
Cutting weak variants. Which of four landing pages deserves real traffic? A synthetic pass will not rank them reliably, but it will often eliminate one for a reason nobody had noticed, and a revised draft can be shown again the same afternoon.
Pressure-testing the question. Run the discussion guide or the survey before it costs a recruitment budget. Questions that produce nothing produce nothing quickly, and the ones that produce only agreement are usually badly framed.
Briefing a team on an audience it has not met. A specific written subject travels better through a team than a demographic bracket, because people argue with a person and nod at a bracket.
What they cannot do
They cannot report what a market did. They cannot establish incidence, size a segment, or supply a percentage anyone should defend. They cannot tell you a fact about the world their author did not know. They cannot be cited as evidence of demand, and a synthetic quotation in a pitch deck, presented as a customer voice, is a fabrication regardless of how carefully the panelist was written.
They also cannot surprise you the way a person can. A real respondent occasionally answers from a life you did not model. A panelist answers from the document, and the document was written by someone with your assumptions in it.
The failure modes to design around
Flattening. Put ten panelists in one conversation and they converge more than ten people would, because every answer is conditioned on the ones before it. Keep them apart, choose panelists who genuinely conflict on the dimension under test, and treat unanimity as a sign the cohort was badly chosen.
Agreeableness. Models drift toward the answer the questioner seems to want. A question that carries an implied preference gets it back, every time. Fix the questions in advance and ask the same ones every round.
Premise echo. A panelist handed a positioning statement will often reuse its vocabulary. If the transcript is full of your words, you have measured your own copy.
Confidence inflation. Per the finding above, discount the certainty rather than the content. “I would never pay for that” from a synthetic respondent is closer to “I would hesitate” from a person.
Summary drift. A model asked to summarize twelve transcripts can report a consensus the transcripts do not contain. Take the counts from what each panelist actually said, not from the summary.
Replayability. If the panelist text can be edited in place, last month’s finding cannot be reproduced. Version the documents and freeze them.
How Klingbar runs a focus group
The panelist library holds 197 authored panelists — 107 consumer and 90 business — published as documents rather than generated on demand. None is a real person. Every panelist carries an Esri Tapestry segment, business panelists also carry a 2022 NAICS code, and every panelist scores the same 19 marketing channels; the guide to a panelist taxonomy covers that layer. Published panelists are immutable and versioned, so a finished study can be replayed against the exact text its panelists were given.
Focus groups are a capability of a crew’s marketing persona, and each step answers one of the failure modes above.
The cohort is built on klingbar.com. You choose six to sixteen panelists from the library, weight and tag them, add a block of product context, and say in a sentence who the cohort is shaped like. The cohort is versioned and rendered into the crew’s prompt tree, so the box runs exactly the text you saved.
The rounds run on your box. The persona’s focus-group task runs there, not in a browser. Each round is one isolated engine call per panelist on a Claude Code login, in a fresh, empty working directory holding only the stimulus, with no connectors and no secrets. No panelist ever reads another’s answer. From round 2 a panelist sees its own prior reaction beside the revised creative, and nothing any other panelist said. A group runs at least three rounds and at most five.
The questions hold still. The same core questions are asked every round, so round 2 is comparable with round 1. Every answer ends with closing lines in a fixed vocabulary: a stance (positive, mixed, or negative), whether the panelist would act, and from round 2 whether the revision addressed what it objected to.
The counts come from the panelists. After each round a separate call writes the synthesis: what landed, the objections, the panelists’ own language, and a brief for the next revision. Its 0–100 score and its resonant and dissent counts are then taken from the panelists’ closing lines — never inferred from the prose — and a panelist whose call failed is left out rather than counted as a dislike. The persona revises the creative between rounds, and every round, transcript, and synthesis is visible under the crew on klingbar.com.
Real users decide. The winning creative does not become a conclusion; it becomes a candidate for an experiment. Where your app runs Statsig and reports conversions the way the persona expects, the persona proposes a launch card, and once an approver approves it, the experiment starts there and real traffic settles the question. During early access the marketing persona is set up with us, not self-serve. The walkthrough follows one focus group from cohort to experiment, and focus groups describes the capability.
Questions worth asking a vendor
ESOMAR’s 20 Questions to Help Buyers of AI-Based Services is the standard checklist, and four of its themes matter most here. Where did the panelist text come from, and was any real person’s data used to build it? Which model played the respondent, and can that be pinned for a repeat run? Can a finished study be reproduced against the exact inputs it used? And what does the vendor claim the output measures — because a vendor that will not answer that one in a sentence is answering it in the pricing page instead.
Where this leaves you
Use synthetic respondents as a rehearsal instrument. They are cheap, immediate, and specific enough to change a draft, a concept, or a launch sequence. Recruit real people when the decision rests on how many, how much, or whether: incidence, willingness to pay, and anything you will defend as a number. Then report which of the two you did.