Definition (canonical)
Cohort analysis groups customers by when they started (signup month, first purchase) and tracks each group's behavior separately over time. It answers the question aggregates hide: is the product getting BETTER for new customers — or is a growing top line papering over decaying quality?
Why averages deceive and cohorts don't
Blended metrics mix generations: rapid signup growth means young users dominate the average, and young users always look engaged — so blended retention can RISE while every individual cohort decays. The company feels healthy right up until growth slows and the rot surfaces. Cohorts unmix the signal: January's users tracked as January's group, month after month. One real-world pattern this catches constantly: a pricing or onboarding change helps NEW cohorts while old ones churn on legacy behavior — blended numbers show mush; cohort rows show exactly when the fix landed and how much it's worth.
Reading the triangle
The classic layout: rows = start cohorts, columns = periods since start, cells = % retained (or revenue retained, or LTV accumulated). Three reads matter: down a column — are newer cohorts better at month 3 than older ones were? (product improvement — the core question); across a row — where does each cohort's curve flatten? (the plateau IS your long-term business; no flattening = no PMF); the diagonal — recent calendar events (outages, campaigns) cut diagonally across all cohorts at once. Revenue cohorts beat logo cohorts for business decisions: 80% of customers retained can mean 60% or 110% of revenue retained (NRR logic).
Setup without a data team
A spreadsheet suffices for years: export customers with start date and activity/payment per month → pivot into the triangle → conditional-format as heatmap. Discipline points: pick ONE activity definition and freeze it (changing definitions mid-history destroys comparability); monthly cohorts for most businesses, weekly only at high volume; annotate the triangle with launches and price changes — the chart becomes the company's actual history book. Then wire the two decisions cohorts are FOR: which acquisition channels produce durable cohorts (cut the pretty-but-leaky ones), and whether LTV assumptions match realized cohort curves.
FAQ
Behavioral vs acquisition cohorts? Start-date cohorts are the default; behavioral ones (grouped by action taken, e.g. "used feature X in week 1") diagnose WHAT drives retention — use both, in that order.
How many periods before a cohort is readable? Three points sketch a curve, six confirm it — resist redesigning the product off a two-month-old cohort.
Do cohorts apply outside subscriptions? Everywhere repeat behavior matters: e-commerce (repeat purchase by first-order month), marketplaces, even hiring (retention by start quarter).
Related: Retention · Churn · LTV · Product-Market Fit Reviewed: 2026-07-16 · Sources: analytics practice canon