AI customer segmentation for e-commerce

Segment size: the 100-person floor below which a segment is a mailing list of one

A segment of 14 people is not a segment. It is a guest list, and mailing it like a segment wastes the targeting that made it worth building.

Segments exist to make messaging relevant at scale, and scale is the part people forget. A segment with 14 people is not a segment. It is a guest list. The statistics do not work at that size: one unsubscribe moves the rate by seven points, one purchase moves the revenue by a fortune, and the campaign report is noise dressed as insight. The floor is around 100 people, and below it the segment should be merged, not mailed.

Why 100 is the floor

The floor is about statistics, not about effort. A segment of 14 people has a 7 percent unsubscribe rate after one unsubscribe. A segment of 100 has a 1 percent rate after the same event. The smaller segment's report is a coin flip, and the campaign decisions made from it are guesses. A segment of 100 is the smallest size where the rates start to mean something, and where a campaign report can be read without a calculator and a prayer.

Klaviyo's predictive features make the same point from the other side: predictive CLV needs 500 or more customers with orders, 180 or more days of history, and repeat buyers. The platform will not even run the model below that size, because the model would be guessing. A segment under 100 is a model that cannot be built.

The segments that fall below it

Three kinds of segment fall below the floor. The hyper-specific slice, built from three conditions that together match a dozen people. The one-purchase wonder, a segment defined by a single product that twelve people bought once. And the VIP list of 12, built by setting the threshold so high that only the top dozen qualify. Each one feels precise. Each one is too small to measure, and precision without measurement is decoration.

The VIP case is the most common, because the threshold feels like a reward. A VIP segment should describe the best buyers, and the best buyers are a group, not a dinner party. If the threshold leaves 12 people, lower the threshold until the group is a segment.

What to do instead

Merge, do not delete. A segment defined by a single purchase becomes a segment defined by a purchase category. A segment defined by one city becomes a segment defined by a region. A segment defined by one product becomes a segment defined by the product's category. The targeting survives in a weaker form, and the statistics start working again, because the group is now large enough to read.

Keep the small segment's rule on file. When the list grows and the segment crosses 100 on its own, split it back out. The merge is a pause, not a deletion, and the rule is the record of what the segment was supposed to be.

The exception that proves the floor

Value can buy an exception. A segment of 20 enterprise accounts, each worth thousands a year, is worth mailing, because one conversion pays for a year of sends. A segment of 20 one-time buyers is not, because one conversion is a rounding error. The floor is about statistical noise, and value can buy an exception, but the exception has to be named, justified, and reviewed. A small segment that exists by exception and is never reviewed becomes a small segment that exists by habit, and habit is how a list fills up with guest lists.

Keep reading

Frequently asked questions

What is the minimum size for a segment?

Around 100 people, as a working floor. Below that, the statistics stop working: one unsubscribe moves the rate by a point or more, one purchase dominates the revenue, and the campaign report is noise. The number is a guideline, not a law, but a segment under 100 needs a reason to exist.

Does a small segment ever make sense?

Yes, when the value justifies it. A segment of 20 enterprise accounts, each worth thousands a year, is worth mailing. A segment of 20 one-time buyers is not. The floor is about statistical noise, and value can buy an exception, but the exception has to be named and justified.

How do I merge small segments without losing the targeting?

Broaden the rule by one step: a segment defined by a single purchase becomes a segment defined by a purchase category. A segment defined by one city becomes a segment defined by a region. The targeting survives in a weaker form, and the statistics start working again.

What happens if I mail a 20-person segment anyway?

Nothing breaks, but nothing can be learned either. The campaign report will show a rate that one person can swing, and the revenue will be dominated by whoever happened to buy. The segment is not harmful. It is unreadable, and unreadable segments accumulate until the list is full of them.