Your customers are unreliable narrators — and your segmentation is built on it

Source: www.unsplash.com

A foundational problem

If you have run one in the last few years, you know the shape of it. The personas were vivid. You named them. The deck looked great in the steering committee. Six months later, the marketing team is back on instinct. The personas are still pinned to the wall, but nobody is using them to decide what to write, who to target, or which lever to pull. The agency briefs are written without them. The next planning round repeats the same arguments the segmentation was supposed to settle.

This is the most expensive failure mode in commercial research, and it is a foundation problem rather than an analysis one.

Most segmentations cluster people by how they answered a questionnaire. The variables are attitudes, lifestyle statements, ratings on five-point scales. The clustering algorithm finds groups inside that response space and the personas emerge.

The problem is that the response space and the behaviour space are not the same thing. Two respondents can give nearly identical survey answers while being driven by completely different things. What people say matters to them isn’t usually what actually moves them. People are unreliable narrators of their own decisions. The segments come apart the moment a marketing team tries to act on them, because the architecture under them was never designed to produce action.

That’s why the personas describe customers well in the deck but lose their grip in the room. The vocabulary is fine. The targeting logic isn’t.

A bigger sample won’t help. Neither will a better questionnaire. The starting point is what’s wrong.

Behavioural fingerprint

Predictive segmentation begins with the outcome the business actually cares about. Financial health. Brand affinity. Category readiness. Retention risk. Whatever the commercial brief defines. We compress the survey variables relating to that outcome into a single behavioural index, scored on a clean 0 to 100 scale. Think of it the way you’d think about compressing an audio file into an MP3. Most of the signal preserved, all of the noise filtered out, one number that captures the latent truth in the data.

Then we run the data through a proprietary machine learning model that learns what drives that outcome. Two things come out of it. First, it tells us which drivers actually move the behaviour and which ones only look like they do. Second, every respondent gets a kind of behavioural fingerprint, a unique combination of the drivers that shape how they behave.

Why smart marketers treat research selection like an agency pitch

21 Oct 2025

We then cluster the fingerprints, not the survey answers. The personas emerge from how those behavioural patterns group, not from how the responses do. That’s the difference between describing what customers say and predicting what they do.

When we did this for Intrum across 20 European consumer markets, the strongest single driver of financial vulnerability turned out to be childhood money stress, not current income. Nobody flagged it as important. The model showed it predicted financial fragility three times more reliably than salary. Twenty-one per cent of average earners were in the most vulnerable group. Only 26 per cent of high earners were truly resilient. Income alone explained almost nothing about how people actually behaved with money.

That dataset became four Money Manager personas, deployed operationally across every Intrum market. Same personas in Warsaw and Barcelona, even though Warsaw and Barcelona look almost nothing alike on demographics or income. Same engagement strategies. Same staff training. Calibrated locally only on tone and channel.

The Intrum work covered the first four steps of the framework. What sits on top, and what most marketing teams actually need next, is the layer that turns the segmentation from a static deck into something the team can interrogate.

That layer has two parts. The first is the Impulse Engine, which adds a motivational architecture on top of the behavioural fingerprint. The fingerprint tells you why a customer behaves the way they do. The Impulse Engine tells you what would actually move them. Each persona carries a primary and secondary motivational state, and together they specify the decision style, the offer triggers that work, the tones that land and the message that converts. The marketing team stops getting a description of a customer and starts getting a brief.

Both layers then feed into the Persona Engine. The Persona Engine is an AI surface built on top of the empirical data, and it is where the segmentation comes alive. The fingerprint defines what drives the persona. The motivational architecture defines how they decide and what kind of message lands. The AI generates the language.

The data does the substance. Every claim the persona makes about itself traces back to a measured value in the underlying study, not to model invention.

What that produces is a persona the marketing team can actually talk to. Show it your campaign headline and ask whether it would convert. Test three offer variants against it before you commit creative budget. Ask it to critique copy you have written for it. The answers are grounded in the data, consistent across sessions, and traceable back to the drivers underneath.

Wider gap

For South African marketers running studies across mixed urban and peri-urban markets, the response-versus-behaviour gap is wider, not narrower. Income, language, channel access and category history all distort what people say about themselves. The drivers shift less. The behaviour you can actually predict, plan against and design messaging for sits under the responses, not in them.

Driver-based personas hold up across markets, across channels and across time. Response-based personas usually do not. What people say about themselves shifts with the cultural framing of the question. What actually drives their behaviour shifts a lot less.

So when the marketing team has stopped reaching for your personas, the right question is whether the segmentation was ever pointed at the right thing in the first place. If the input was a questionnaire and the output was a description, the answer is probably no.

A persona that describes a customer is a wall poster. A persona that predicts their behaviour, and tells your creative team how to talk to them, is a decision.