AI market research: choosing the right method for the question.

In short: AI market research helps teams build customer understanding faster and make it available across the business. Which method fits depends on the question. Use observed behavior when you need to know what people do, reported preference when you need to know what people say and think, and modeled inference, such as customer-specific AI personas, to explore needs and directions on demand. Heatseeker combines customer-specific personas with live market evidence so teams can apply customer understanding to positioning, propositions, concepts and campaigns. Behind every live Heatseeker experiment is a synthetic experiment. Comparing the paired results shows how closely the model's responses correlate with what the audience actually does.

Three kinds of evidence

Customer research produces three different kinds of evidence. They support different claims.

• Observed behavior is what people did: clicks, sign-ups, purchases, switching. It is the strongest basis for a claim about behavior, and it describes the context and measure of the study.
• Reported preference is what people said in a survey, interview or focus group. It is rich for understanding language, concerns and reasoning. It does not always match later behavior, which is why the say-do gap matters for decisions about what people will do.
• Modeled inference is what a model estimates from patterns in data, as customer-specific AI personas do. It is fast and available on demand. Its value depends on what it is built from and how it has been evaluated. See synthetic personas.

A good method choice starts with the decision and asks which claim it needs to rest on.

Choosing a method by question

What large CPG teams need from AI customer research

Enterprise CPG teams usually carry many brands, many audiences and many markets, and customer understanding has to be useful to brand, innovation, insights and media teams at once. That shapes what to look for:

• Reusable understanding. Research done once should stay available for the next brief, rather than ending in a report.
• Audience specificity. Personas built from your own research and customer data, not general knowledge about a category.
• Evidence you can inspect. What each answer rests on, and how the persona has been compared with observed behavior.
• Fit with existing research. Surveys, trackers and studies you already pay for should inform the model, not be replaced by it.
• Enterprise controls. Security attestations, data handling and the ability to share personas and workspaces across teams.

Heatseeker is an enterprise customer intelligence platform built around those needs. See the platform and the customer stories.

Understanding customer segments, including in beauty

Beauty and personal care brands often serve segments that differ in routines, concerns, occasions and what "good" means. To understand different segments:

1. Bring together the research and customer feedback you already have for each segment.
2. Explore each segment's needs, motivations and trade-offs with customer-specific personas, within the audience each persona represents.
3. Identify the segment questions that carry the most decision risk, and add live market evidence there.
4. Apply the understanding to the decision: a proposition, a product concept or a campaign brief.

This guide uses public information only. For customer examples we can discuss publicly, see customer stories.

Where live market evidence fits

Live experiments put options in front of real audiences and observe the response. They add behavioral evidence where it matters most: an important gap, an unfamiliar question or a calibration need. The findings feed back into the personas, so understanding builds over time. For why customer assumptions deserve evidence, see our comparison of AI models with observed market responses, a selected study of general models that does not evaluate Heatseeker's personas.

Every live Heatseeker experiment has a synthetic experiment behind it. We compare their results to measure how closely the model's responses correlate with observed market responses. That comparison is part of the workflow, giving your team evidence for the decision and an ongoing check on the model.

1. Synthetic response: what the model predicts.
2. Live response: what the audience does.
3. Paired results: how closely they correlate.

The question can start with a customer need, a pain point, a feature or an emerging trend. Design the stimulus around that question, give the alternatives a fair comparison and measure the response the decision depends on. A click shows a response to the stimulus; use sign-up or purchase measurement when the decision depends on those actions.

The method builds on field experimentation. Bertrand and colleagues varied real loan offers and measured demand. Ding, Grewal and Liechty showed how giving choices real consequences improved prediction in their conjoint studies. Gordon and colleagues demonstrated why allocation and comparison design matter when measuring advertising effects.

Anonymized public example: a bank was certain its brand should be built on human-led customer support. A concept test against real buyers in market showed a different customer need, speed to set up an account, drawing the stronger response in every market, before the rebrand. See concept testing.

Questions to ask any AI market research provider

• What is it built from? Your own research and customers, or general knowledge?
• How has it been evaluated? Against what, for which tasks and audiences?
• Which claims does the evidence support? Needs and directions, response to a concept, or purchase?
• Whose model gets smarter? Is your data used to train a shared model?

Working on a new positioning direction? Here's how to test it before committing campaign spend.