Segmentation

Splitting users or customers into groups that behave differently.

Also known as: segmentation, customer segmentation, user segments

Segmentation splits a population into groups that behave differently enough to deserve different treatment: by plan tier, acquisition channel, company size, behaviour. Averages lie across heterogeneous groups; segments tell the truth piece by piece — and mismatched segments explain most “the data contradicts itself” moments.

overall conversion: 4.2% (flat)
enterprise: 1.1% ↓   self-serve: 5.8% ↑   → two opposite stories, one average

Segments must earn their keep: each one should differ on the outcome you care about and imply a different action. A segmentation nobody uses to decide anything is taxonomy for its own sake. Start with business-natural cuts (tier, channel, tenure) before algorithmic clustering — interpretable segments get used, clever ones get admired.

The classic mistakes:

Segments that describe but do not discriminate. Cutting by attributes that do not move the outcome adds complexity without insight. Validate: do the groups actually behave differently?
Too many segments. Twenty micro-segments, each too small to measure or act on. Merge until every segment is big enough to matter and small enough to treat distinctly.
Confounding the segment with the cause. Enterprise converts worse — because of the segment, or because enterprise gets the slow sales process? Segments describe; causal claims need more.
Static segments for moving users. A “new user” segment means something different at day 1 and day 90. Recompute or graduate users on a schedule.
Simpson’s paradox in reverse. Segments can also mislead when the real driver cuts across them. Check the alternative cut before concluding.

When to go further: behavioural clustering for discovery, cohorts when time since an event is the dimension, RFM for a fast commercial cut. Segmentation is the lens; pick the lens that changes the decision.