Retention, Cohort by Cohort.

A cohort retention grid groups customers by the month they joined, then tracks each group's retention over the following months. The result is a triangular heatmap: the rows are vintages (January cohort, February cohort, etc.), the columns are months-since-joining (M1, M2, M3...), and the cell color encodes the retention percentage. This is the single most-cited diagnostic in growth analytics because it surfaces issues that month-over-month aggregates hide.

Aggregate monthly churn averages across all vintages. If a recent cohort is churning faster than the older base, aggregate churn looks stable while the future is breaking. Cohort retention exposes this immediately: the bottom-left of the grid (recent cohorts, early months) is where new product or marketing changes show up first. Every serious growth analyst checks this view before trusting an aggregate retention number. Amplitude's explainer and Mixpanel's guide both treat cohort analysis as table stakes for product analytics.

The canonical PMF signal, popularized by Lenny Rachitsky's retention benchmark essays, is a retention curve that flattens. A curve that keeps decaying (10%, 5%, 2%, 1% across M1, M3, M6, M12) signals leaky activation or weak habit formation. A curve that flattens at 30-40% by M3 and holds means the customers who stayed have stuck. Best-in-class consumer SaaS hits 35-45% M12 retention; best-in-class B2B SaaS hits 80%+ M12 logo retention.

If a specific vintage drops sharply at M2 or M3, look at what changed in onboarding, pricing, or product during their first weeks. If the drop is uniform across all vintages at the same age, you have a structural retention problem (poor habit moment, weak feature stickiness) and it's an activation engineering problem rather than a marketing one. If the drop is on recent vintages only, it's likely an acquisition-quality issue: paid channels are pulling lower-fit users who never had the intent.

Monthly once you have at least 6 monthly cohorts with 100+ users each. Below those thresholds the grid is too noisy to read. Pair with the funnel analyzer when you suspect the leak is acquisition or activation, and with the LTV:CAC calculatoronce you can quantify the dollar impact of fixing the cohort that's breaking.