Skip to content
Faisal Mahmud TusherDigital marketing, SEO & analytics
with Faisal Tusher

AI in Marketing › AI Applications in Marketing

Synthetic Personas and Simulated Respondents

What AI-generated personas and simulated survey answers can and cannot tell a marketer, the few safe uses, and why they never replace real research.
Guideintermediate9 min readUpdated

What synthetic personas and simulated respondents are

The term covers a family of practices that share one feature: a model stands in for people who were never asked. The common forms are:

  • Synthetic personas. A model writes a profile for a segment, with a name, a job, goals, frustrations and buying habits, from a short description of the market or the product.
  • Simulated interviews. A model is told to answer as a particular kind of customer and is questioned as a participant would be, sometimes across several invented personas in turn.
  • Simulated survey responses. A model is given a questionnaire and asked to answer it many times, as different invented respondents, and the answers are tabulated.
  • Simulated focus groups and panels. Several persona prompts are set to react to a concept, a message or an advert and to discuss it among themselves.

Each of these is distinct from using AI on real research material. The AI for research synthesis guide covers a model coding and summarising what real customers said, where every theme traces to an excerpt from a participant. Synthetic material has no participants, so there is nothing to trace it to. The audience research guide builds a Buyer Persona from research and warns against inventing anything to fill gaps; a synthetic persona is, in effect, made entirely of the filler. A simulated panel is no substitute for real group research either; the tutorial on running a focus group covers recruiting real participants and reading their reactions without letting the loudest voice decide.

FormWhat it producesWhat it is sometimes mistaken forWhat it actually is
Synthetic personaA plausible profile of a segmentA research-based personaA hypothesis about a segment, written in persona form
Simulated interviewFluent first-person answersCustomer interview findingsThe model's guess at what such a person might say
Simulated surveyA table of answers and percentagesSurvey resultsVariation in model output, not a sample of a population
Simulated panelA discussion with apparent disagreementFocus group insightThe prompt's framing, voiced several ways

Why a model's answer is not a customer's

A language model generates text that is likely given what it learned from and what it is asked. When it answers as "a busy parent choosing a meal-kit subscription", it draws on how such people are written about in its training text: articles, forums, marketing copy, other people's research summaries. The result is a reasonable portrait of the conventional wisdom about that kind of person. It is not a report from one, and it is not a measurement of how many of them exist or what they do. The guide to how language models generate text explains the mechanism behind this.

Several consequences follow. The answers lean towards the familiar, because the familiar is what is most often written; the unexpected reason a customer actually switched, which is often the most useful finding in real research, is the least likely thing a model will produce. The answers follow the wording of the prompt, so a question that hints at an answer gets it, and a persona described as price-sensitive will be price-sensitive in every answer. The answers are fluent and consistent where real people are hesitant, contradictory and vague, which makes synthetic findings look cleaner and more confident than real ones.

The customer research hub points out that real customers are reliable witnesses to their own recent behaviour and poor forecasters of their future behaviour. A simulated customer is neither. It has no past purchase to recall and no budget to protect, so it can only forecast, and it forecasts what the text it learned from suggests. Asking it whether it would buy a new product or what it would pay produces an answer with none of the grounding that makes even a real customer's answer weak.

A simulated survey has a further problem: its numbers look like a sample and are not one. Running a model through a questionnaire many times produces a distribution of answers, but that distribution describes how the model's output varies with the prompt and its settings, not how opinion varies in a market. No Segmentation analysis, sizing or comparison between groups can be read from it. This guide makes no claim about how closely any particular product's simulated answers match real ones; the point is that a marketer has no way to know, for their own category and question, without doing the real research anyway.

Useful jobs: hypotheses, guides and stress tests

None of this makes synthetic material useless. It makes it preparation rather than evidence, and preparation is a real part of research. The useful jobs are those where the output is a question, a draft or a list that real research or a person's judgement will then test.

Drafting hypotheses. Before research, a team needs a list of what might be true: reasons people might buy, reasons they might not, triggers, alternatives, anxieties. A model can produce a broad first list quickly, which a team then narrows to what the research will test. The research design and brief guide starts research from a decision and its research questions; synthetic hypotheses can feed that step, provided they are written down as hypotheses.

Rehearsing an interview guide. A researcher can question a simulated participant to hear how the guide flows, spot questions that are leading or double-barrelled, and practise follow-up prompts before the first real interview. The simulated answers are discarded; the improved guide is kept.

Pre-testing survey wording. Asking a model to answer a draft questionnaire as different kinds of respondent can surface ambiguous wording, missing answer options and questions that only make sense to insiders. This is a check on the instrument, not on the market. A cognitive pre-test with a few real people remains the standard, and the synthetic pass is a cheap step before it.

Listing objections and stress-testing messages. A model asked how a sceptical buyer might react to a proposition, a landing page or a sales script can produce a list of possible objections and unclear passages. Each item is a candidate for checking against sales calls, support tickets and real interviews, not a finding.

Rehearsing conversations. Sales and support staff can practise handling difficult questions with a simulated customer. The value is the rehearsal, and nobody draws a conclusion about customers from it.

Jobs they must not do

The line is simple: synthetic material must never be the basis for a decision that depends on what real customers think, want or do. In practice that rules out several tempting uses.

  • Sizing or prioritising segments. Choosing which segments to target, the second step of the STP framework, depends on real differences in size, need and value between groups. Synthetic personas have no size and their differences are whatever the prompt described.
  • Choosing between concepts, messages or names. A simulated panel's preference reflects the model and the prompt. Real preference needs real people, and often a live test.
  • Setting prices or reading willingness to pay. Real stated willingness to pay is already unreliable; a simulated one has nothing behind it.
  • Validating a Go-to-Market Strategy plan. A launch built on synthetic reassurance has skipped the step that would have revealed a problem.
  • Reporting percentages. No figure from a simulated survey should appear in a report, a deck or a claim, because it describes no population.
  • Quoting a customer. A simulated answer presented as a quote is a fabricated quote, however it is labelled later.
  • Replacing research the team cannot afford. If real research is out of reach, say the question is unanswered and use the cheapest real evidence available, such as reviews, support logs, search data or a handful of conversations, rather than a confident synthetic substitute.
Illustrative example — not a real result

Preparing a tea subscription survey with synthetic respondents

Orrinby Tea Co., an invented loose-leaf tea subscription, plans a survey of 600 lapsed and current subscribers about pausing versus cancelling. Before fieldwork, the researcher asks a model to answer the 24-question draft as 6 different invented subscribers. The synthetic pass finds 4 problems: a question about "box frequency" that the model reads in two ways, a price question with no option for "I share a subscription", a matrix of 14 rows that every simulated respondent answers the same way, and an insider term, "tin swap", that the model guesses at. All 4 are fixed, and 5 real subscribers then complete a cognitive pre-test, which finds 2 further problems the model did not. The simulated answers themselves, including a confident 70% preference for pausing, are deleted. The survey goes to the real sample, and the deck says the questionnaire was checked with a model and pre-tested with subscribers.

Labelling and reporting synthetic material

The risk with synthetic material is less that a team uses it than that it travels. A persona written as a brainstorm ends up in a strategy deck; a table of simulated answers is pasted into a slide beside real survey results; a fluent quote from a simulated interview is repeated in a pitch. Each step strips away the context that it was never research. Labelling is what keeps that context attached.

The professional codes published by ESOMAR and the Market Research Society, as the site's marketing research guides describe them, expect findings to be reported honestly, including the limits of the method. Synthetic material meets that expectation only if it is labelled where it is used. A workable rule set:

  • Label at the point of use. Every page, slide, table or document that contains synthetic material says, where the reader will see it, that it was generated by a model and is not research with customers.
  • Keep it separate. Synthetic outputs never share a chart, table, file or dataset with real responses. Separate storage also prevents them being swept into a later analysis by mistake.
  • Say what it was for. "Used to draft the interview guide" or "used to list objections for testing" tells a reader how much weight it can bear.
  • Record how it was made. Keep the prompts, the date and the tool, so anyone can see what framing produced it.
  • Never present it as a persona from research. A synthetic profile is titled as a hypothesis, and any research-based persona says which study it came from.

A team that already keeps a register of AI uses, as the governance section recommends, should list synthetic research there with its permitted jobs, so that the boundary is written down before someone is tempted to cross it under deadline.

Common mistakes

  • Treating fluency as validity. A clear, confident answer from a simulated customer is not evidence that any customer thinks it.
  • Putting synthetic and real data side by side. Once they share a chart, nobody can tell them apart.
  • Reporting simulated percentages. A distribution of model outputs is not a share of a market.
  • Quoting simulated respondents. A generated sentence in quotation marks is a fabricated quote.
  • Using synthetic personas to choose targets. Segments without size or real differences cannot be prioritised.
  • Skipping the real pre-test. A synthetic pass on a questionnaire comes before testing with people, not instead of it.
  • Letting the prompt decide the answer. A persona described as price-sensitive will always say price matters; that is the prompt talking.
  • Calling it research. Labelling synthetic material as research, even informally, is how it ends up driving decisions it cannot support.

Frequently asked questions

Can AI-generated respondents replace a customer survey?

No. A simulated respondent produces an answer that is plausible given the model's training text and the prompt, not one drawn from the people who buy in your category. It cannot tell you how many customers hold a view, what they would pay or which concept they prefer, because there is no sample behind it. Use simulated answers to prepare a survey, for example to find ambiguous questions, and then field it with real people.

Is a persona built by AI from our own research data synthetic?

It depends on what the model did. If it summarised coded findings from real interviews and every statement in the persona traces to evidence, it is a research output that a model helped write, and the research synthesis checks apply. If it filled gaps, added details or wrote the persona from a description of the market, those parts are synthetic and should be labelled as assumptions or removed.

How should synthetic material be labelled in a presentation?

Clearly, on the slide or page where it appears, not in a footnote at the end. Say that it was generated by a model, that it is not research with customers, and what it was used for, such as drafting hypotheses or testing a questionnaire. Never place synthetic answers in the same chart or table as real responses, and never quote a simulated respondent as if a customer had said it.

Terms used on this page

Frameworks

Where this page is used

Also related

Sources

  1. ESOMAR — ESOMAR
  2. Market Research Society — MRS

Related