Customer decision
Mintel
Synthetic respondents showed 0.79–0.88 Jensen-Shannon distribution similarity from demographics alone, with no choice data as input. This is distribution agreement, not predictive accuracy.
Decision
Mintel, a global market research firm, had already run a choice-based conjoint on a 50ml facial moisturizer the traditional way: field a survey to 1,000 human respondents, then field it again to a second 1,000 as a blind validation cohort. The question was whether synthetic respondents could reproduce that result closely enough to trust, and Mintel set the bar in the open beforehand: 80 to 85% would be compelling enough to carry internally.
Interpretation
The model was given demographics only, no preference data and no choices from the real survey, and still tracked the human distributions at 0.79 to 0.88 Jensen-Shannon similarity across attributes. It also disagreed in a specific, legible way. Human demand fell off sharply between $34 and $48, where the model assumed a smoother decline, so the two diverged at the top of the price range. On attribute importance the model put proof and actives on similar footing, which Mintel judged more intuitive than its own data.
Comparison
The comparison behind the decision.
| Comparison | What the source records |
|---|---|
| Human comparison | A 50ml moisturizer conjoint with two waves of 1,000 human respondents; the second wave was a blind validation cohort. |
| Model comparison | The same product attributes and choice task, using demographics only. No human preference or choice data went into the model. |
| Distribution similarity | Reported Jensen-Shannon similarity ranged from 0.79 to 0.88 across attribute distributions. This is distribution agreement, not a percentage of correct predictions. |
| Price divergence | Human demand dropped sharply between $34 and $48; the model assumed a smoother decline. The source reports divergence at the upper price points. |
Mintel POC review, 2026-04-09 (Circleback meeting ps6_KyX1Vp-cqSLq6Vkgz)
Numbers
Figures with provenance.
0.79–0.88
Jensen-Shannon similarity between synthetic and human choice distributions
Inspect provenance
2 × 1,000
human respondents: one wave to calibrate against, a second as blind validation
Inspect provenance
Zero
preference or choice data given to the model; demographics only
Inspect provenance
Model-predicted
Modeled and observed evidence.
Predicted
Choice distributions and attribute-level preferences for a 50ml facial moisturizer, generated from demographics alone with no preference data and no choices from the real survey.
Observed
Mintel's own conjoint, fielded to 1,000 human respondents and repeated on a second 1,000 as blind validation.
Agreement
0.79 to 0.88 Jensen-Shannon similarity across attribute distributions. Mintel previously stated an 80 to 85% target for internal buy-in. Distribution similarity is not prediction accuracy, so these scores do not establish whether that target was met. The two diverged at the top of the price range, where human demand drops sharply above $34 and the model assumed a smoother decline.
Mintel POC review, 2026-04-09 (Circleback meeting ps6_KyX1Vp-cqSLq6Vkgz)
Limits
What this record establishes.
Raw human and model distribution series are not available in the approved public record, so this table does not reconstruct curves or bars. Similarity does not establish predictive accuracy, sales lift or agreement on treatment-effect size. The price divergence is part of the result.
Causality
Attribute-level utilities are identified, not associated. The same design that identifies them in the human data identifies them in the synthetic data, which is what makes the two comparable at all.
- Intervention
- Product attributes varied across choice tasks: format, actives, proof claim, and price, in 10 tasks across 2 blocks plus 2 shared holdout tasks.
- Outcome
- Choice share across the moisturizer set.
- Design
- Choice-based conjoint replicated on synthetic respondents, with hierarchical Bayes fit to both the synthetic and human datasets and full distributions compared rather than point estimates.
The best comparison we've done across all our work together.