How we segment

The most useful thing a segmentation does isn't statistical. It gives a company one way of talking about its audience.

We call that the primary segmentation. When marketing, product and the executive team use the same names for the same people, their arguments get sharper, because they're arguing about the same thing. Findings travel from one team to the next. New people pick it up quickly. Without it, each department keeps its own picture of the customer, and nobody can add the pictures up.

So the first question isn't which clustering method to use. It's which segmentation the business will actually speak.

The data can come from four places, and each sees something the others miss.

A survey of the whole market, clustered on what people need from the category. It's the slowest and the most expensive, and it's the only one that sees people who don't buy from you yet. That's why it's usually the right primary segmentation when growth is the question.

Your own customer data: transactions, engagement, how people use the product. Cheaper and richer in detail, but it can only see the people you've already got.

What people follow and talk about in public online. Fast, and good for keeping up with shifting tastes and planning communications. It's a supplement, not a foundation.

Language models reading large amounts of public conversation. Quick and cheap, and useful for forming hypotheses before any data is collected. Then we test those hypotheses with real people.

Whichever source carries the work, we cluster on needs rather than behaviour where we can. Behaviour tells you what people do. It rarely tells you why.

A gaming project made the point. Players who engaged heavily with certain game modes looked like one group. They weren't. Their engagement came variously from a need for social connection, for competition, for exploration or for creativity. Same behaviour, four different reasons. A segmentation built on the behaviour alone would treat them as one group and serve them the same way. Pulling the needs apart took careful feature design and a lot of iteration on the clustering models.

That project taught a second lesson. Behavioural data is rarely designed to capture motivation. When a product is built around specific needs, and the tracking measures whether each feature meets the need it was built for, needs-based clustering gets much easier and much more useful.

The end point is the same whatever the method: a clear view of your Now, Next, Not Yet and Never audiences.

What to do on Monday

• Ask three senior colleagues to name your audience segments. If you get three different lists, you don't have a primary segmentation.
• Check what your current segmentation is built on. If it's only customer data, it can't see who might come next.
• Pick one high-engagement behaviour in your data and list the different needs that could be driving it.
• The next time a feature is specified, add tracking for the need it's meant to serve, not just the click.

A needs-based segmentation sits at the heart of Now+Next.

← All points of view