Données synthétiques et études clients synthétiques : quand les utiliser, quand les éviter
TL;DR
AI can now simulate customers, focus groups, and survey responses. This is a real power tool for speed and scale — and a serious trap if used to replace real customer contact. Synthetic methods accelerate hypothesis generation, message pre-screening, and scenario work; they systematically mislead on real preference, novel products, emotional response, and price sensitivity. Use the three-gate test before letting synthetic output drive a decision.
What This Guide Covers
Where synthetic research adds genuine value, where it systematically misleads, the three-gate test that filters when to use it, how to run it well when you do, and the more robust uses of synthetic data for model training and privacy-safe sharing. Built for marketing researchers and product teams tempted to replace expensive customer research with instant AI personas.
Key Takeaways
- Synthetic research accelerates hypothesis generation, message pre-screening, scenario work.
- It systematically misleads on real preference, novel products, emotional response, price sensitivity.
- Three-gate test: reversible decision, downstream validation, familiar territory.
- Synthetic data has stronger uses in model training, testing, privacy-safe sharing.
- Don’t replace customer conversations with simulated ones.
What Synthetic Research Can Do
- Exploratory hypothesis generation — brainstorming likely reactions before running a real test.
- Survey design and pre-testing — catching ambiguous questions before sending to real respondents.
- Message pre-screening — eliminating obviously weak variants before A/B testing with real users.
- Role-play scenarios — training sales or support with simulated difficult customers.
What Synthetic Research Cannot Do
Known failure modes where synthetic output systematically misleads:
- Real preference measurement — LLMs over-index on articulated, rational-sounding preferences. Real consumers are messier and often wrong about their own behavior.
- Novel product reaction — the model predicts based on training data; for genuinely new categories, it’s guessing.
- Emotional or visceral response — synthetic respondents don’t feel irritation, delight, or confusion the way humans do.
- Cultural or subcultural nuance — especially for groups under-represented in training data.
- Price sensitivity — synthetic respondents systematically under-state price sensitivity.
The Three-Gate Test
Before using synthetic research for a decision, ask:
- Is the decision reversible? Reversible decisions tolerate synthetic input; irreversible ones (product launches, rebrands, major campaigns) need real data.
- Can we validate downstream? Synthetic pre-screening followed by real testing is fine. Synthetic as the last step before ship is not.
- Are we in familiar territory? Established categories, known audiences, incremental variations — synthetic is more reliable. Novel products or audiences — much less so.
How to Run Synthetic Research Well
If you’re going to do it, do it right:
- Define the persona precisely — “a 42-year-old working parent with $95K household income in Boston suburbs who uses [brand X] weekly” beats “a millennial mom.”
- Simulate many, not one — 50 diverse synthetic respondents catch distributional patterns one persona hides.
- Ask the same question many ways — phrasing strongly affects LLM output. Consistent answers across phrasings are more trustworthy than single responses.
- Always label the output clearly — “synthetic research” vs. “customer research.” Mixing them in reports will eventually cause a real mistake.
Synthetic Data for Training, Not Just Research
A separate, more robust use: generating synthetic data to train or test other models:
- Test coverage — synthetic edge cases to check how a customer-facing model handles unusual inputs.
- Privacy-safe sharing — synthetic data that preserves statistical properties of real data without exposing individuals.
- Class balancing — augmenting rare categories in a dataset to improve model fairness and accuracy.
- Adversarial testing — generating prompts designed to probe chatbot failure modes before launch.
Common Mistakes to Avoid
- Treating synthetic focus groups as customer substitute. Synthetic output is polished and convergent; real customers are messy and tell you things you didn’t ask.
- Mixing synthetic and real findings in reports. Eventually causes a real mistake.
- Using synthetic for irreversible decisions. The cost is too high.
Action Steps for This Week
- Run one synthetic focus group on a current marketing question.
- Have one real conversation with a real customer on the same question.
- Put the outputs side by side.
- The differences are where synthetic research will mislead you.
Foire aux questions
Can synthetic research replace customer interviews?
No. Use synthetic for pre-screening; real research for decisions.
How many synthetic respondents do I need?
50 minimum to capture distributional patterns. One synthetic persona is anecdotal at best.
Is synthetic data legal under GDPR?
Synthetic data derived from real personal data must follow privacy rules. Pure synthetic from public/aggregate sources is fine.
What’s the best use of synthetic data in marketing?
Adversarial testing of customer-facing AI before launch.
Will AI replace UX research?
No. It accelerates synthesis; live human contact remains the validation step.
Sources et lectures complémentaires
- Riman, T. (2026). An Introduction to Marketing & AI 2E.
About Riman Agency: We design synthetic + real research workflows. Book a research audit.