Can you understand Gen Z, young parents or the soon-to-retire without asking a single human being? Short answer: with synthetic personas you can explore, test and rough out ideas very quickly, but you can't conclude. For generational research, synthetic is an accelerator. It doesn't replace observing real behaviour.
This is no longer a theoretical question. Stratégies, the French trade title, reports how Patrick van Bloeme, co-founder of Novatris, which became Harris Interactive and then Toluna, has launched Limpid, a platform for qualitative research without human respondents. Companies can use it to run a complete study on synthetic personas, "from brief to deliverables", in a few hours and autonomously. When someone who spent thirty years running research institutes makes that bet, agencies would do well to take a close look.
What does respondent-free research promise?
According to Stratégies, Limpid lets you study behaviours, usage and reactions to concepts, messages or situations. You can simulate a sudden price rise, a new health crisis or life moments such as moving house or retiring.
The platform also sets out its own limits, which is to its credit: it doesn't claim statistical representativeness and doesn't produce quantitative surveys. In other words, it is a tool for testing scenarios quickly, not for measuring a market.
For an agency, the appeal is obvious. A pitch brief lands on Monday, the recommendation is due on Friday: there is no time for classic qualitative fieldwork to happen. Synthetic fills the gap.
Why does a fluent answer look so much like the truth?
This is where the trap snaps shut. A synthetic persona answers in impeccable, coherent, nuanced language. It has a first name, an age, habits. All of it looks like knowledge.
Call it the data illusion effect: the more precise and polished a piece of information looks, the more credit we give it, regardless of how solid it actually is. Synthetic research pushes that effect to the maximum, because the form is perfect by design.
Yet a synthetic persona lives through nothing. It plays back regularities learned from vast amounts of text. Data reveals patterns, but it cannot define humanity. A statistical pattern about "older people" is not a 67-year-old woman hesitating over whether to switch banks.
Is generational research a minefield?
Yes, for a simple reason: generations are already stereotypes. "Gen Z" doesn't exist as a homogeneous bloc. A student in Lyon, an apprentice in rural France and a young Parisian executive have little in common beyond their year of birth.
A language model also learns, and learns a great deal, from what is said about generations: articles, opinion pieces, studies, debates. So it risks knowing the discourse about Gen Z better than the way young people actually express themselves. The result: a synthetic persona can make the cliché more convincing instead of correcting it.
Imagine an agency preparing a campaign for a health insurer targeting the newly retired. The synthetic panel confirms what everyone already believes: caution, a need for reassurance, wariness of all things digital. Meanwhile, the real newly retired are posting about their travels, their volunteering projects and their grandchildren. The reassuring campaign falls flat.
Synthetic, self-reported or observed: which method for which job?
| Approach | What it captures | Representativeness | Good use | Trap |
|---|---|---|---|---|
| Synthetic personas | Plausible reactions, simulated by a model, in a few hours | Not claimed | Exploring scenarios, roughing out messages | Mistaking fluency for truth |
| Quantitative survey | What people say, at scale | Possible if the sample is built for it | Measuring, sizing, tracking over time | The gap between what people say and what they do |
| Qualitative interviews | Deep motivations, personal stories | No, but real depth | Understanding the why | Samples too small to generalise from |
| Observed published content | What people actually post and share | Limited to the audiences who post | Anchoring insights in real behaviour | Forgetting the silent audiences |
No method wins on every row. The right question is not "which one should I choose?" but "in what order should I combine them?".
How do you combine synthetic and real without kidding yourself?
- Start with what is observed. Analyse what the target generation actually publishes in your category: topics, words, tensions, formats. That is your foundation.
- Build personas from that content. Group posts by profile and by concern, rather than starting from an age bracket.
- Use synthetic to explore. Put scenarios, concepts or hooks to synthetic personas to prune the weak routes quickly.
- Test every conclusion against reality. Every claim that comes out of synthetic research must find an echo in real posts. No echo, no insight.
- Measure when you need to measure. If the decision commits a significant budget, validate it with a representative sample or a real-world test.
- Document the method in the recommendation. Tell the client what comes from synthetic, observed and self-reported sources. They will thank you the day the question comes up in the boardroom.
What advertisers are going to ask you
As synthetic research spreads, one question will come up in every presentation: "Where does this insight come from?" An agency that answers "from a panel of AI-generated personas" will have to say a lot more. An agency that can show the real posts behind each conclusion won't have that problem.
That is the logic of Defy Monitor: every AI conclusion points back to the real posts it rests on, and strategy is rebuilt from what audiences and competitors actually publish, not from a survey. Audience content is clustered into personas automatically, which avoids starting from an age bracket. So the Defy platform isn't opposed to synthetic research: it gives it an anchor.
Synthetic is a good tool, used in its proper place. The danger isn't the machine. It is the temptation to believe that a well-written portrait is worth as much as an observation.
