Naming and shipping segments

The table is a group-by: one row per cluster, mean of each feature, plus size and share of revenue. Here are the Mango Lane five, profiled from session 4's validated k=5 run:

Two reading habits. First, always add the share-of-revenue column: 55 VIPs are 18% of heads but a far larger slice of the money, and that asymmetry is usually the single most persuasive number in the deck. Second, read rows for contrast, not description: gifters and lapsed spenders have similar average orders, and only recency and frequency tell you one needs an occasion nudge and the other a rescue mission.

The size column deserves its own read. Our five sit at 55 / 90 / 60 / 45 / 50 - no giant blob, no sliver, every one comfortably fundable, which is part of why session 4's actionability test passed at k=5. Two warning shapes to know. One segment holding most of the pile usually means the features are not separating behavior - everyone looks average on what you measured, so measure something else. A segment in the single digits is a handful of customers wearing a persona; it may be real (whales exist) but it belongs on an account manager's list, not in a campaign tool. Healthy segmentations look like ours: a big middle, distinct wings, nothing too small to matter.

A professional trick: profile on MORE columns than you clustered on. We clustered on R, F, M - but once the segments exist you can compute each segment's average discount usage, category mix, or return rate, none of which the algorithm ever saw. If those held-out columns also differ sharply between segments, that is independent evidence the structure is real - the clusters "know" things they were never shown. If every held-out column looks identical across segments, your clusters may be geometry without meaning. This is the closest thing unsupervised work has to a free validation set.

Before a name ships, one last test: re-run the clustering on a subsample of the pile - do the same five personas reappear? Drop a random 20% of customers, cluster the rest, profile again. Structure that is real survives resampling: you should meet the same loyal core, the same big middle, the same graveyard of lapsed spenders, with wobbling counts but stable characters. Structure that was luck - a border artifact, an initialization fluke - shuffles into different groupings each run, and no honest name can stick to a moving target. This is the unsupervised replacement for the train/test split you lost in session 1.

One honest paragraph our lab owes you: the lab pile is deterministic - seeded, frozen, guaranteed to reproduce - so its stability check will pass by construction. Real pipelines run against live customers, and real segments DRIFT: fresh arrivals mature into steady mid-baskets, VIPs lapse, a pandemic or a product launch rewrites the geometry. Real teams re-run the profiling quarterly and expect the sheet to change. Stability is something you check on a schedule, not something you establish once.

Three outcomes of a stability check, and what each means. All segments reappear: ship, with the check cited as evidence. Most reappear, one dissolves: the dissolving one was probably a border population - consider k-1, or accept it as a soft segment and say so on its sheet. Little reappears: stop shipping. You were about to name initialization luck, and session 9 will show you how convincing that luck can look. The check costs minutes; the credibility it protects took the whole project to build.

Here is the complete Mango Lane sheet for one segment, every field filled:

Notice what the metric field does: it makes the segmentation falsifiable one level up. If reactivation does not move, either the campaign is weak or the segment was not real - and the review date guarantees somebody has to ask which. The sheet is the segmentation's contract with reality.

The sheets are the working artifact; the meeting gets a compression of them. One summary slide: five rows, columns for name, size, the single standout number versus the pile average, campaign, and owner - with the full sheets as appendix pages nobody reads but everybody trusts for existing. Lead with share of revenue where it is dramatic (the VIP row usually is), and give every row its one-sentence pitch from the demo below. The discipline transfers: if a segment cannot survive compression to one honest row, it was not ready to ship - and if a stakeholder challenges a row, the appendix sheet answers with the evidence, the owner, and the review date already attached.