Can I use synthetic findings in a client deliverable?

Yes, if you label them. State on the cover that the audiences are synthetic, keep the language directional, and attach the validation plan. Agencies do this routinely with exploratory work; the problem is never the method, it is the unlabelled slide.

How is this different from just prompting ChatGPT?

Three things: the profiles persist and remember across a session so round five is still the same conversation; they are grounded in one market's demographics, prices and register rather than a global average; and the session is captured as a transcript and an executive report you can hand to someone else. A single chat gives you none of that.

The first question is free with no account, one every 48 hours. €7.99 buys 60 responses — 10 questions to a panel of six — valid for 30 days as a one-off purchase. €59.99 a month covers 600 responses across up to 10 focus groups with parallel rooms and cited market evidence.

AI consumer research, used the way it actually works

AI consumer research covers two very different things, and conflating them is how teams end up presenting simulated output as consumer behaviour. One is simulated respondents — synthetic panels that generate answers. The other is AI-moderated interviews with recruited humans. This page is about the first, what it is good for, and the discipline that keeps it honest.

Ask one question free See pricing

The two things the term covers

Synthetic panels generate responses from models configured as personas. Nothing a synthetic panel says came from a person, which is precisely why it is fast and cheap and why it can never be evidence.

AI-moderated research interviews real recruited participants, using a model to run the conversation, probe and summarise. The participants are human; the moderator is not. That output is evidence, with all the recruitment cost that implies.

Both are sold as 'AI consumer research'. Before you buy either, ask one question: does a human being appear anywhere in this pipeline? The answer determines what you are allowed to claim afterwards.

Where simulated respondents genuinely help

The wins are all in the part of the process that is currently done badly or not at all, because it is too slow to be worth doing.

Concept screening at volume — ten ideas down to three before anyone is recruited.
Discussion-guide piloting, which almost nobody does today and which recovers half of a two-hour session.
Language discovery: the words a category is actually described with, before you write the survey.
Objection mapping across buyer profiles, including the sceptical and edge-case ones a small recruit would miss.
Market sense-checking for a country you have no presence in yet.

How to report it without overclaiming

The single most damaging habit is turning six answers into a percentage. Six people are not sixty per cent of anything, and a number implies a precision the method does not have.

Not "68% would buy" — instead "the panel surfaced three recurring objections".
Not "this validated product-market fit" — instead "simulated reactions helped sharpen the survey".
Not "customers prefer A over B" — instead "the configured profiles prioritised A, to be tested".
Not "we validated pricing without human testing" — instead "pricing hypotheses to be tested in field".
Say on the cover page that the audiences are synthetic. It costs nothing and it stops someone citing it as fieldwork six months later.

The validation plan is part of the deliverable

Every finding from a synthetic session should leave the room with a named next step: the interview that will confirm it, the survey question that will size it, or the experiment that will settle it. Without that, a directional signal quietly becomes a fact through repetition in slide decks.

Where the profiles disagree with each other is usually more informative than where they agree. Unanimous synthetic agreement often means the question was not sharp enough to divide anyone.

Grounding is what separates useful from plausible

A model asked to be a consumer will produce fluent, plausible, average output. That is the failure mode: it reads like signal and is noise.

What makes a simulated answer worth reading is specificity — a real cost of living, real shops, a real argument happening in that country this month. Audience Seems injects the current economic and social climate of Spain, Mexico or Chile before each session, and each profile reasons with the money and the references of that market rather than a translated average.

What it will not do

It will not forecast demand, size a market, establish representativeness, prove causation, or measure real willingness to pay. It will not replace human participants where the decision carries legal, medical, financial or safety consequences, and it cannot report a sensory reaction to a physical product.

If a vendor tells you otherwise, that is the moment to ask for the validation study — and to notice whether one exists.

Frequently asked questions

No. It replaces the informal exploration that happens before research is commissioned, and it is useful again afterwards for iterating without re-recruiting. The commissioned study still happens; it is just better aimed.

Keep reading

Synthetic audiencesSynthetic audiences grounded in the demographics, economy and everyday language of Spain, Mexico and Chile. Six profiles react live to your concept. First question free.Synthetic focus groupRun a synthetic focus group with six AI participants grounded in Spain, Mexico or Chile. Ask, follow up, export the transcript. First question free, no sign-up.LATAM researchTest concepts on Mexican and Chilean consumers with synthetic audiences grounded in local income, retail and price habits. No local recruiter. First question free.

Last reviewed: 2026-08-21. Participants are AI-generated synthetic archetypes, not real people.