Most segmentation is theatre: the audience is split into groups with names, and then everyone receives the same campaign. Segmentation earns its keep only when the message a segment receives would be wrong for another segment. If you cannot say what a group needs to hear that another group does not, you have not found a segment; you have found a filter.

Segment on what predicts behaviour

Recency, frequency and value of purchases or usage. Product category or use case. Stage in the lifecycle — new, active, at-risk, lapsed. Stated preferences and reasons. Acquisition source, if sources behave differently. Not demographics for their own sake; a thirty-year-old and a fifty-year-old who both bought the same thing last week are the same segment.

Write the difference

For each segment, one sentence: what do they need to hear that others do not. New customers need reassurance and the first step. High-value repeat customers need early access and to be thanked, not discounted. At-risk customers need a reason to stay, specific to what they stopped doing. Lapsed customers need news. If the sentence is the same for two segments, merge them.

Keep it maintainable

Five to eight segments, computed automatically from data every night, with a customer in exactly one lifecycle segment at a time. Segments maintained by hand are out of date within a month. Every journey and campaign declares which segments it targets, and the platform enforces exclusions so a lapsed customer never receives the 'welcome back' and the 'we miss you' in the same week.

Test the segmentation itself

Send the segment-specific message to half of a segment and the generic message to the other half. If the specific message does not outperform, the segment is not real or the message is not different enough. This is the one A/B test that tells you whether the whole apparatus is worth maintaining.