01
Maximum Entropy
We build the most neutral distribution that is consistent with what is known, without hidden assumptions.
The least biased modelling possible
Science & validation
Honest answers, with confidence levels.
Reproducible, traceable results, benchmarked against reality, with limits we state openly. Not a black box.
The science
Three building blocks for honest answers.
02
Probabilistic models
Every answer is a distribution, not a single point. We capture uncertainty and the real diversity of opinion.
We measure disagreement too
03
Machine learning
Continuous calibration: outputs are benchmarked against real-world data and refined through feedback loops.
Verified against reality
Research foundations: F. Pachet & J.-D. Zucker, Maximum Entropy Relaxation of Multi-Way Cardinality Constraints for Synthetic Population Generation, CoRR abs/2603.22558 (2026).
Validation
Same winner. Same runner-up. Same reservations.
ThinkONE tested five advertising claims for a food brand twice: once with real focus groups in the UK, and once with 20 AI personas. Both approaches arrived at the same ranking and the same key reservations.
A second test, on an animated ad concept, confirmed the pattern. Both groups praised a premium, distinctive universe, and both found the story too abstract and short on proof of taste.
Parallel animatic test
Real respondents vs. 20 AI personas
An animated ad concept for a food brand, set in a dreamlike café with a premium, artsy tone, was shown to real respondents and to 20 AI personas.
| Dimension | Focus groups | AI personas |
|---|---|---|
| Attention | Strongly captures attention | Highly premium, artsy |
| Branding | Strong brand attribution | Pack and red codes strongly remembered |
| Understanding | Positive proposition, unclear narrative | Benefit understood, but too abstract |
| Usage context | Needs product usage scenes | Lacks everyday relevance |
| Credibility | Lacks taste proof points | Missing taste and trust proof points |
Shared reading: keep the premium universe and visual memorability, and strengthen the narrative, usage moments and taste proof points.
Validated against reality
Outputs are benchmarked against known real-world data. We measure the gap and show it.
Uncertainty acknowledged
Every result comes with a confidence interval. No false precision.
Transparent and reproducible
A traceable method, so you know why a result emerges and how it was produced.
A complement, not a replacement
Built for exploration, iteration and pre-testing. Fieldwork stays essential for high-stakes decisions.
Limits
When to use it, and when not to
“Synthetic data is a powerful tool for exploration, enrichment and acceleration. It should not replace field research when the objective is to measure reality, uncover the unexpected or provide robust evidence.”
Use it to
- Rapidly explore hypotheses
- Enrich existing personas and past studies
- Pre-test a concept, message or campaign
- Anticipate likely reactions from known targets
- Accelerate exploration before real-world research
- Work on sensitive topics or rare, under-documented populations
Not as a primary source to
- Precisely measure a market or size it
- Produce statistically robust market measurements
- Conduct robust pricing research on its own
- Power a brand tracker or a quantitative U&A study
- Validate a high-stakes decision on its own
- Study weak signals or disruptive, emerging behaviours
Good use: explore and guide decisions. Poor use: replace real-world measurement.
Security & sovereignty
Engage an entire market without a single piece of personal data.
How we protect your dataThe right mix makes better decisions.
The question isn’t “synthetic or real?” It’s “what’s the right mix to make better decisions, faster?”