Customer segmentation
Groups that behave differently, and that can be treated differently
The problem
The segments the team works with were defined once, in a meeting, and usually rely on what was easy to look at: age, location, company size, tenure. They were given names, four or five were assembled, and there they stayed.
The problem is not that they are wrong. It is that they group by what customers are like rather than how they behave, and those two things coincide less than it seems. Two people with the same profile can buy with completely different frequencies, tickets and motivations, and end up in the same box receiving the same message.
The symptom is recognisable. Campaigns perform similarly across all segments, which usually means the segments are not separating anything. And no one revisits them, because there is no moment in the year when that is supposed to happen.
McKinsey estimates that companies with active behavioural segmentation and personalisation generate between 5 and 15% in additional revenue over their existing base. The difference lies in updating the groups when behaviour changes, not in having them.
The solution
Groups built from observed behaviour, not declared attributes. What each person buys, how often, at what ticket, how they respond to a discount, which channel they come from, how long they take to return.
Each group arrives with its profile, which is what allows giving it a name the team recognises and deciding what is done differently with it. A segment without an associated action is a label, and it changes nothing.
How many segments is a business decision. The number is not dictated by statistics: it is dictated by how many different treatments the team can genuinely execute. If marketing can sustain three different messages a month, eight segments is an exercise that nobody will use. It is the conversation that needs to happen before building, and the one that almost never does.
What is validated here, which is different from a predictive model
This is the only one of our cases where there is nothing to predict. There is no future event, no correct answer, and therefore no accuracy percentage. That disorients those who expect a metric and is the most important thing to understand.
A segmentation is validated on two levels: whether the groups actually behave differently and whether the team can do something different with each one. A set of statistically clean groups on which no one can act is the most common failure mode of this technique, and it is not detected by any metric.
It is worth saying what it does not do and what it does. It does not say who is going to leave or who is about to buy — that is what the predictive models are for, and they are used on top of segments, not instead of them. What it does do, and is underutilised, is place a new customer in their group from day one. Still without their own history, they inherit the treatment that has already proved to work with the customers most similar to them, instead of waiting months for them to accumulate enough behaviour.
More on the blog: Dashboard or model: the temporal question that decides where to invest →
How it integrates
Each customer carries their segment as a field in the system, and it is recalculated at whatever frequency makes sense — in most businesses quarterly or half-yearly. The interesting thing is not just the snapshot but the movement: a customer who changes group is saying something, and someone moving from a high-value segment to a low-value one is usually signalling before any other indicator.
The same technique, other applications
Grouping by observed behaviour applies wherever there are many units and they need to be treated differently without being able to examine each one individually.
Product segmentation.
Which SKUs behave similarly in demand and turnover, for differentiated stock and pricing policies.
Supplier segmentation.
Grouping by criticality, delivery behaviour and volume, to define relationship models.
Store or territory segmentation.
Stores or zones that resemble each other, so that the same campaign is not applied to different realities.
Employee segmentation.
Turnover, promotion and retention patterns, for retention policies by profile.