Testing purchase intent by demographic segment with an AI panel

An overall purchase-intent distribution feels reassuring, right up until you split it by segment and find two opposite realities hiding underneath. A concept that wins over a younger audience can leave an older one completely indifferent, with the overall average giving no hint of it. Ignoring that gap means making a launch decision on a number that, in practice, represents nobody.

The problem with a single average

A purchase-intent average of 3.4 out of 5 can correspond to two very different situations: everyone sitting around 3.4, or half the audience at 4.5 and the other half at 2.3. These two cases call for completely different decisions. In the first, the concept is broadly middling for everyone. In the second, there is a highly receptive segment worth targeting precisely, and another it might be better to leave aside rather than dilute the message trying to please everyone.

What a segmented read reveals

Comparing distributions by age, gender, geographic region or income level, depending on what the panel definition allows, surfaces dynamics that the average hides.

• A segment can show strong purchase intent concentrated on the middle ratings, a sign of real interest held back by a specific objection (often price).
• Another segment can be polarized, with plenty of "definitely" and plenty of "definitely not," a sign of a positioning that divides rather than convinces.
• The verbatims attached to each segment usually explain why, in different words depending on the audience.

This segmented read turns "does the concept work" into a more useful question: "who does it work for, and under what conditions."

What a classic human panel does on this front

The access panels run by established agencies like Ipsos or Kantar rely on large networks of real panelists, with self-reported profiles that allow fine-grained sector and demographic targeting. That is a real strength for large-scale studies, but this agency-mode workflow, with a brief and fieldwork, keeps the timelines and costs typical of the traditional market. GDPR plays a structural role there: consent, personal data hosting and panelist rights need to be managed for every study.

What Panelia changes on this specific point

Panelia simulates hundreds of synthetic respondents per test, calibrated against real human data, and lets you compare several demographic segments on the same concept in about ten minutes, for around one euro per test. Since the run relies on synthetic respondents, no personal data from a human panelist is collected to produce the distribution, which simplifies part of the project's data handling, without exempting Panelia itself from its own obligations regarding its users' data. The method follows a protocol published on arXiv (2510.08338).

This result stays a decision-support tool, not a guarantee: for a very high-stakes segmentation decision, a complementary human test on the most critical segments is still worth running.

In practice

The process is to describe the concept, define the segments to compare, run the test and read the distributions side by side. The gap between segments, when it exists, usually stands out on the first read, with verbatims explaining the difference rather than just an isolated number. Even when the overall average looked perfectly fine, that segmented view is often the moment a launch plan actually gets sharper, sometimes turning a single broad campaign into two smaller, better-targeted ones.