Blog
Notes, op-eds, and field reports
Syntheses, lessons from the field, takes. On substance, on method, on what AI-assisted research changes — and what it does not.
Why Synthetic Personas Caricature Real People: What Pew’s Data Show About Silicon Sampling
Diane and Brenda share one row in any sampling frame, and they do not share the same views. Their synthetic twin is more emphatic than either of them. Measured on Pew’s published data: opinions pushed away from the centre, disagreement cut by almost a third.
Personas synthétiques : pourquoi l’IA caricature les répondants (les données de Pew à l’appui)
Michèle et Françoise ont le même profil, et pas les mêmes avis. Leur jumeau synthétique, lui, est plus tranché que chacune des deux. Mesuré sur les données publiées par Pew : des opinions poussées loin du centre, un désaccord réduit de près d’un tiers.
Intermarché Terroirs: what a storyboard pretest would have shown
We took the film “Ces dimanches-là” back to storyboard stage and showed it to twenty synthetic personas. The emotion lands; proof, products and price still need work.
Use case: a synthetic focus group on French news, without predicting 2027
Eight synthetic personas from four political families read the same two news sheets, first alone, then together. Where they agree, where they split, and what it cannot tell you.
Use case: twenty synthetic personas on their own place in market research
We asked twenty synthetic personas whether synthetic personas can replace human respondents. Not the polite yes you might expect: their conditions, their objections, and what the study cannot say.
Intermarché Terroirs : ce qu’un pré-test sur storyboard aurait montré
Nous avons remis le film « Ces dimanches-là » au stade du storyboard et l’avons montré à vingt personas synthétiques. L’émotion passe ; la preuve, les produits et le prix restent à travailler.
Cas d’usage : pré-test publicitaire de quatre films, en entretiens puis en focus group
Quatre films d’une campagne bancaire, quinze entretiens individuels puis un focus group de neuf entrepreneurs synthétiques : ce que le pré-test éclaire, et ce qu’il ne permet pas de conclure.
Use case: an ad pretest of four films, in interviews and then a focus group
Four films from a bank’s campaign, fifteen one-to-one interviews, then a focus group of nine synthetic entrepreneurs: what the pretest shows, and what it cannot tell you.
Focus group synthétique : que teste-t-on quand les personas ne connaissent pas l’actualité ?
Une actualité documentée, huit personas synthétiques, des réactions individuelles puis une discussion collective : ce que ce dispositif permet d’explorer — et ce qu’il ne permet pas de conclure.
Can AI Stand In for Human Survey-Takers? “Not Really” — but Not Nothing Either
Pew Research Center has published one of the most thorough public evaluations of synthetic samples to date, and its verdict is clear: AI cannot stand in for human survey-takers. The verdict holds. But it answers one question, and Pew's own published data answer a second one.
Jev for market research: we put it in the lab
A new kind of model shipped in September: it does not write, it decides, and it returns probabilities instead of a single answer. We started testing it on our synthetic panel ten days later. Short version: not as a respondent, not yet. But we are not done with it.
Études qualitatives et quantitatives avec l’IA : analyser, interroger, simuler
L’IA entre dans une étude quali ou quanti de trois façons, qui n’engagent pas la même confiance : analyser des réponses humaines, interroger de vraies personnes, ou simuler les répondants. Ce que chacune permet, ce qu’il faut vérifier, et comment les combiner.
Un sondage synthétique peut-il prédire la présidentielle 2027 ?
Des « sondages IA » vont circuler d’ici le printemps. Ce qu’un sondage synthétique sait faire, pourquoi une élection est le cas le plus difficile, ce que dit la loi, et quatre questions pour lire ces chiffres.
Aaru and Simile on validation: what they showed, and the number still missing
Within a month, two of the most closely watched simulation companies published real evidence. Aaru measured itself against surveys repeated on themselves and against actual purchases; Simile built a model that predicts its own errors. Here is what they showed, what neither has measured yet, and what we do, on a much smaller scale, in the open.
Faire une étude de marché avec ChatGPT : ce qu’il fait bien, ce qu’un persona synthétique ajoute
ChatGPT sait structurer une étude de marché, pas interroger vos clients. Entre sa réponse unique et le terrain, les personas synthétiques ajoutent une étape : un panel qui se contredit, et que l’on peut vérifier.
Cas d’usage : vingt entretiens individuels avec des personas synthétiques sur leur propre place dans les études
Une étude qualitative exploratoire, publiée de bout en bout : la question posée, le panel de vingt profils, le dispositif, les thèmes, les désaccords, les verbatims et un entretien lu en entier. Puis ce que l’étude permet de dire, et ce qu’elle ne permet pas.
Les répondants synthétiques sont-ils fiables ? Nos chiffres, huit marchés
Les fournisseurs annoncent 85 à 95 % de précision. Nous publions un autre chiffre : à quel point des répondants simulés sont en désaccord entre eux, mesuré face aux enquêtes nationales de huit marchés, et les quatre choses qu’il ne prouve pas encore.
How Reliable Are Synthetic Respondents? Our Own Numbers, Eight Markets
Vendors quote 85 to 95 percent accuracy. We publish a different number: how much simulated respondents disagree with each other, measured against national surveys in eight markets — and the four things it still does not prove.
Synthetic Personas vs Digital Twins: One Word Changes the Evidence You Owe
They are sold as the same product. They are not the same claim — and the stronger claim now has the worse published record.
Synthetic Respondents: What a Vendor's Claim Actually Tells You
Three providers say “synthetic respondents” and mean three different products. Here is how to tell which one you are being sold — and why the most common accuracy claim answers the wrong question.
Faire une étude de marché avec l’IA : ce qui tient, ce qui ne tient pas
L’IA fait aujourd’hui quatre choses très bien dans une étude de marché, et deux choses pas du tout. Savoir lesquelles évite de prendre une note de synthèse pour une mesure.
Under-dispersion, and what we did about it
On 1 September we published a defect in our own panels: simulated respondents who share a profile agree far more than real people do. We have measured it against 1,016 real respondents, corrected it, and started rolling that correction across every market we serve.
Données synthétiques et sondages : simuler n’est pas mesurer
Peut-on remplacer un sondage par des données synthétiques ? Les quatre usages reconnus par la profession, les écarts déjà mesurés, et le test qu’un acheteur peut faire lui-même.
Personas synthétiques : peuvent-ils remplacer les répondants ? Je leur ai posé la question
J’ai interrogé vingt-huit répondants synthétiques sur leur propre remplacement. Leur réponse unanime est instructive — et cette unanimité l’est encore plus.
Données synthétiques pour les études marketing : une matière première, trois usages
Assistant de connaissance, persona synthétique ou jumeau numérique : la même base peut nourrir ces trois usages, à condition de ne pas confondre restitution, simulation et prédiction.
Why ESOMAR Asks Twenty Questions
In a preregistered audit of 37 language models, a Gaussian copula with no AI in it scored 0.688 against the best model's 0.714 — both far below what humans manage with each other. No demo would have shown you that. That gap is the whole reason a buyer needs a questionnaire rather than a presentation.
Under-dispersion, and a test anyone can run
In our answers to ESOMAR's 20 Questions we committed to publishing what we found on domain fine-tuned models, either way. Here is the result — and, more usefully, the method that produced it. It runs on published data, costs nothing, and scores our own work unflatteringly in places.
Calibration, and what a buyer can check
A commentary priced the calibration layer at $2 billion. ESOMAR's own figures say 8% of research leaders consider synthetic data reliable enough for a business decision. This note argues that calibration does not require a private panel — and reports what we found when we audited our own claims against ESOMAR's 20 Questions, including three we were making publicly that were not true.
85% of What? The Number the Synthetic-Research Market Keeps Getting Wrong
One figure has become the industry's shorthand for “synthetic respondents work”: 85%. It comes from a serious study, and it does not mean what the market has made it mean. It is not 85% accuracy — it is 85% of the consistency people manage with themselves. That distinction changes what you can reasonably buy.
What Is a Synthetic Panel? A Reference Guide to Synthetic Respondents and the Fifth Generation of Market Research
A reference guide to synthetic panels: what they are, how synthetic respondents are actually built, what the industry is really doing, what the science says, and where they create decision-grade value — framed as the fifth generation of market research.
The Fake AI Monet, and Why Your Synthetic Persona Will Get the Same Treatment
A viral hoax — a real Monet passed off as AI-generated and torn apart by online critics — quietly replicates twenty years of psychology research. The same bias is now distorting the synthetic-research debate. The real question is not 'is it human or synthetic?' The real question is: does it help you decide?