Cohort Analysis

The triangle, and how to read it

A cohort table has signup periods down the side and elapsed months across the top. Every row starts at 100% in month 0 and decays rightwards. Because a cohort that started three months ago cannot have a twelve-month figure, the filled cells form a triangle — the newest rows are short and the oldest are long, and the bottom-right of the table is empty by construction.

That shape dictates how to read it. Down a column compares like with like: every cohort at the same age. This is the comparison that answers whether the product is getting better or worse, and it is the reason the table exists. Across a row shows one cohort's decay curve — steep early, flattening later. Reading across tells you when customers leave; reading down tells you whether that is changing.

Cells further right are always built on fewer cohorts and older data, so the right edge of the table is both the most interesting and the least current part of it.

What aggregates hide

A single churn rate is a weighted average across cohorts of every age, so it moves for reasons that have nothing to do with retention. Grow quickly and your base fills with young customers who churn fastest, and blended churn rises while every individual cohort is performing better than the last. Stop growing and blended churn falls, because the base ages into its flat tail.

The worked example above is the common failure: aggregate churn steady at 4% while month-6 retention dropped fourteen points across two quarters. The old cohorts were doing the work. By the time a headline number reacts, three quarters of new customers have already been acquired into a worse experience.

Customer cohorts and revenue cohorts diverge

Run the same table on MRR instead of headcount and the shape can change completely. Revenue cohorts include expansion, so a row can exceed 100% — a cohort that lost 30% of its customers can still be worth more than it was at signup if the survivors grew. That crossover point, where expansion overtakes churn, is the clearest visual evidence that a land-and-expand motion works, and it is invisible in a customer-count table.

The inverse is also worth seeing: a customer cohort holding at 85% while its revenue cohort sits at 70% means the survivors have downgraded, which is a quieter and slower problem than cancellation. Run both tables.

Practical construction rules

• Never re-base the denominator. Every cell divides by the cohort's original size. Dividing by the previous month's survivors turns a retention curve into a period-over-period churn series, which is a different metric that happens to look similar.
• Fix a rule for reactivations. A returning customer either rejoins their original cohort — the more common convention, and the one that keeps cohorts closed — or counts as new. Rejoining can push a cell above a previous month's value, which is correct and surprises people.
• Watch the last row. The most recent cohort is usually incomplete, since customers acquired on the 28th have had two days. Either exclude partial periods or mark them clearly, because an artificially healthy bottom row is exactly where the eye goes.
• Segment before you conclude. A blended cohort table across self-serve and enterprise, or across acquisition channels, averages curves with genuinely different shapes. The insight is almost always in a segment, not in the blend.

What cohorts are for downstream

Cohort data is the honest input to anything that depends on customer lifetime. LTV computed from a single churn rate assumes constant hazard — the same cancellation probability every month — which cohort curves visibly falsify: churn is front-loaded and the tail flattens. Summing a cohort's measured cumulative gross profit and extrapolating only the observed tail is slower, more work, and considerably more defensible. Cohorts also give NRR its natural denominator, since NRR is precisely a one-period cohort measurement.