Retention Cohorts

• Build a period retention matrix from raw events
• Draw the correct conclusion from cohort-row shapes and column trends
• Debate what counts as 'active' and keep it comparable

• A single aggregate retention number can't tell 'new users worse' from 'everyone decays faster' — the matrix separates them.
• It makes product launches checkable: release in July, July's row should flatten sooner.
• Row shapes reveal the decay curve (fast attrition vs mild slope) suggesting different fixes.
• Columns surface seasonal effects while reducing calendar noise per cohort.

Each year's intake has its own attendance trend; the headmaster compares 'same grade, different years' to spot whether this year's class is drifting from school earlier. Rows are classes, columns are grade levels.

• Select the entry event, the active-event definition, and the period (day/week/month).
• Compute for each cohort and each period k the share unique-active / cohort_size (k=0 → 100%).
• Present as a matrix and/or plotted row-by-row as retention curves.
• Compare row shapes and column trends; benchmark decay curve against category norms.
• Drill into underperforming cohorts to distinguish product regressions from acquisition quality issues.

• A job board sees its M0→M1 column drop 20 points for cohorts starting in April when a competitor launched — a market event caught in the matrix.
• A productivity app's June cohort (post-onboarding-redesign) flattens its decay curve: 12% higher M3 retention than May's row — the change is validated.
• A content site uses the matrix as its weekly health check, paging the product lead when any recent cohort dips below its 90-day band.

• Define 'active' exactly (event list + dedupe key) and freeze it for matrix comparability.
• Show cohort counts alongside percentages so thin cohorts don't mislead.
• Use period length deliberately: daily for sticky consumer apps, weekly/monthly for B2B.
• Benchmark the row shape, not just M1: the flattening point (knee) is where retention stabilizes.

• Filling M0 with all signups including users who never actually started using — M0 should be 'active in their first period', often ~signup-or-worse.
• Comparing cohorts with different period lengths.
• Including perpetually-bounded cells (like M3 of a 2-month-old cohort) in trend reads.
• Switching the active definition mid-period and 'discovering' a product win that's actually a metric change.

• A dropped cohort row appearing across ALL cohorts means an active-definition change, not a cohort problem.
• Check cohort sizes before diagnosing shape; a 30-user cohort row has huge sampling noise.
• Reconcile with the raw events table by hand-recomputing one cell for a known cohort.
• Holidays and internal launches must be annotated on the matrix; otherwise analysts can't reconstruct why a vintage deviated.