How would you build a cohort retention analysis, and what would you look for in it?

Group users by the period they first did something (usually signup month), then measure the share of each cohort still active in each subsequent period, aligned by periods since start rather than calendar date. Read it two ways: down a column to see whether newer cohorts retain better than older ones, and across a row to see where in the lifecycle users drop off.

First I pin down two definitions, because everything depends on them: what starts a cohort, usually first purchase or signup month, and what counts as retained, which has to be a real action rather than a login. Then it is a matrix, cohort down the side, months since start across the top, with the percentage of that cohort active in each month. Aligning on months since start rather than calendar months is what makes cohorts comparable at all. Reading down a column tells me whether product and onboarding changes are working, so if the month three retention of recent cohorts is climbing, something we did is landing. Reading across a row tells me where the lifecycle problem is. On a subscription product I worked on, the drop was almost entirely between month one and month two, which pointed at onboarding rather than long term value, and that reframed the roadmap. The trap I always flag is the newest cohort, which is incomplete and will look artificially bad if you plot it uncritically.