E-commerce customer segmentation

"Our customers" is rarely one group, but most marketing treats them like one. We used behavioral data to find the real groups inside a retail customer base, and what to do differently for each.

Type
Applied study
Domain
Retail & e-commerce
Techniques
Clustering, behavioral analytics
Scope
Segments into action

The problem

An online retailer's customers are wildly different from each other. Some buy weekly, some bought once two years ago, some only show up for sales. But marketing usually sends everyone the same message, aimed at an average customer who doesn't really exist.

The goal of this study was to find the real groups inside the customer base, then figure out what to do differently for each one.

Same six CRISP-DM steps as every project. Here's how they went.

what we learn feeds back into the business questions Businessunderstanding Dataunderstanding Datapreparation Modeling Evaluation Deployment
The six CRISP-DM steps. Every project here ran through them.

Start with the business

Keeping an existing customer is much cheaper than winning a new one, and different customers need different things to stay. A loyal regular doesn't need a discount. A customer who's gone quiet might come back for one. Sending both the same email wastes money on the first and loses the second. So the goal wasn't just "find groups." It was to find groups a marketing team can treat differently.

Get to know the data

The data was purchase history: what people bought, how often, how recently, and how much they spent. Simple data, but it took real preparation, because a customer's behavior over time has to be summarized into one profile before any grouping makes sense.

Build the model

We used clustering, which finds groups of customers who behave similarly instead of starting from assumptions about who the customers are. The groups come from the data. That tends to surface segments a demographic guess would miss, like the occasional big spender, or the loyal regular who never spends much.

loyal core drifting one-and-done
Three different groups, each worth a different next move.

Check it against the goal

The test for a segmentation is practical. Are the groups clearly different from each other? Do they stay stable over time? And does each one suggest a different action? If two groups would get the same email, they aren't really two groups.

Make it something people can use

The deliverable was a plain-language description of each group: who they are, why they buy, and what to do next for them. Keep this group happy, grow that one, win back the third. Something a marketing team can act on with the tools they already have.

The clustering takes an afternoon. Making the groups mean something a business can act on is the real work.

This was an applied study. We're happy to discuss how a segmentation project would be scoped and what data makes it worthwhile.

Treating one-size-fits-all customers as one size?

There are usually distinct groups worth serving differently.

Book a free discovery session