The Segment Describes. The Persona Decides

Long before it was my job, I was already sorting people into categories. One of the first frameworks I met was a color-based profiling tool: fill in a questionnaire, come out red, blue, green or yellow. Mine came back analytical, to nobody's surprise. What made it useful had nothing to do with how accurate it was. Sharper instruments followed, Myers-Briggs among them, most of them more precise. The color survived because a whole office could use it as shorthand, one person telling another to keep a pitch short because of how a colleague scored, and both understanding exactly what was meant. It was easy to apply on the spot. That, not accuracy, tends to decide whether a way of describing people ever gets used.

Set that against what most segmentation delivers. A typical attitudinal segmentation manages to be accurate and nearly unusable at once. It lands in an awkward middle: too broad to tell a marketer what to actually say, too intricate to survive as everyday shorthand, rigorous but inert. It hands over a description. The person who commissioned it increasingly wants a decision.

From Description to Decision

I did not reach this conclusion from inside the research industry. After leaving TNS Kantar, I spent years in adjacent fields, first in programmatic advertising, where a live auction settles in milliseconds what to show a person and what to spend, then in retail, where purchase and loyalty data feed the systems that shape what a shopper sees next. In both worlds the pattern was identical.

Building a model was routine, telling the business what to do was the whole point, and describing what had already happened was merely the raw material for both. Coming back to research, it felt a step behind: still handing over the description when the buyer now wanted the recommendation. That gap is why I started Knowsis, a decision intelligence consultancy whose job is to keep what research already does well and add the layer above it, the part that says which signals matter and what to do about them.

Start with the Outcome, Not the Questionnaire

Segmentation is where the difference shows up most clearly. As a purely statistical exercise it is easy. Put survey responses into a clustering routine, take whatever groupings emerge, give them names. The arithmetic is fine. The trouble is that it has grouped responses, and how someone answers a questionnaire is a weak stand-in for what actually drives them. Two people can give near-identical answers for completely unrelated reasons. Build the groups on the answers and they come apart the first time a marketing team tries to use them.

Our approach starts from the other end. We treat segmentation as a behavioral question first and a statistical one second, and it hinges on a step most projects skip: choosing a dependent variable, the outcome you are actually trying to explain, before going anywhere near clustering. Frequently the business already has one. Net Promoter Score, category spend, retention, satisfaction, any metric it already lives by can serve as the anchor. Sometimes the outcome is not something the business captures directly. Things like financial wellbeing or brand affinity are real but buried across a dozen questions, so before grouping anyone we model that outcome into a single index and score every respondent on it. Either way the analysis ends up with one behavioral target to aim at, rather than a wall of attitudinal variables.

The rest follows from there. A model trained on that target does two jobs. It separates the drivers that genuinely shift the outcome from the ones that merely correlate with it, and it places each respondent on those drivers as an individual profile. Those driver profiles are what we group on. The segments then reflect what produces the behavior, not the language people reached for to describe themselves.

What Held Across Twenty Markets

We ran this for Intrum, the European credit management group, on a financial-health study covering 20 markets. Since financial health is not something respondents report directly, it became the modelled outcome. The driver that mattered most was not income: it was the money stress people carried from childhood, which forecast adult financial fragility about three times better than salary did, though nobody had nominated it as important.

Among average earners, 21% sat in the most fragile group; among high earners, only 26% were genuinely secure. The study resolved into four money-management personas that Intrum put into operation in every market it covers. The striking part was the stability. The personas behaved the same way in Warsaw as in Barcelona, cities that share almost nothing on demographics, because driver structure carries across borders that demographics cannot.

Why It Is Finally Usable

A segmentation this precise runs straight back into the problem I opened with: it is accurate and it is costly to act on. Carrying a detailed profile in your head while you draft a brief or a subject line is precisely the effort that made the richer frameworks unusable. What has changed is how cheaply that detail can be turned into something a person can use in the moment.

Hand the profile to a language model and it will give you the plain instruction a strategist once worked out by hand, only now anchored in how people behave rather than in what they claimed. The one thing that has to be right is the order of operations. A confident model will present a flawed profile with the same assurance as a sound one, so the substance has to be fixed and constrained before any language is generated on top of it. Reverse those two and you have built an articulate way to be wrong.

I have been categorizing people for as long as I can remember, and never because the category itself mattered. The label was only ever a route to understanding what drives someone and what they are likely to do next. A conventional segment tells you who a person is. A segment built from the start around the behavior you care about tells you what they will do and how to move them. The first is a description. The second is a decision, which is what the buyer was paying for in the first place.