Cohort Churn Analysis
When you lose users is as informative as how many
Churn measured per cohort — early-churn signature, time-to-churn distribution and churn by vintage — shows whether losses are structural or cohort-specific.
What you will learn
- Compute cohort churn rates and time-to-churn distributions
- Identify early-churn signatures and their causes
- Separate structural churn from vintage-specific churn
Cohort churn analysis tracks not just total churn but when and which cohorts churn: month-1 churn rate, time-to-churn histogram and churn curves per vintage. Early-churn-heavy cohorts point to activation/onboarding; late churn points to value decay.
What is this?
Cohort churn analysis decomposes churn by vintage and by elapsed time. Instead of a single churn %, it produces: per-cohort churn rates by period (are March signups churning faster than January's?), a time-to-churn distribution (most losses at month 1 vs month 12), and churn curves (survival) per cohort. The pattern — early vs late, structural vs episodic — tells which lever (onboarding? core value? pricing? competitive) deserves attention, something a rate alone cannot.
Why do you need it?
- Heavy month-1 churn is almost always an onboarding/activation problem; fixing it is different from fixing pricing.
- Cohorts that churn at the same rate regardless of vintage point to structural decay, not a specific bad change.
- Time-to-churn informs revenue forecasting: early churn shifts CAC payback and cash-flow timing.
- Survival curves are the rigorous way to compare retention across cohorts.
Mental model
A hospital's readmission curves
Cohort churn analysis
Hospitals track not just 'patients returned' but when and which groups: post-op week-1 returns get a different response than month-12 returns. The timing and the group both dictate treatment.
How it works
- For each cohort, define churn as exit from the product (no activity / no payment) per period.
- Compute churn rate per period: churned-in-period k / active-at-start-of-period k.
- Build time-to-churn distribution: histogram of survival duration for each churned user.
- Plot cohort churn curves (1 − stepwise retention), overlaid by vintage.
- Diagnose: early-heavy + degrading vintages → acquisition/onboarding; uniform late churn → value/pricing structure.
Reading churn shape.
In production
- A fintech sees churn concentrated at month 1 across all vintages; a support-ticket audit ties it to KYC friction, so onboarding becomes the fix target.
- A B2B platform sees a single vintage (the October cohort) churn at 2x — traced to a price change, rolled back for that segment.
- A subscription box company uses time-to-churn to schedule win-back windows (day-30 outreach) with measured lift.
Best practices
- Use survival-style denominators (active at start of period), not gross cohort size, for per-period churn.
- Report monthly and cumulative churn side by side.
- Annotate the churn curve with product/pricing/competitor events for causally-readable history.
- Segment churn analysis by plan and entry channel — 'all users' churn hides the real story.
Common mistakes
- Counting users who merely paused as churned (activation-floor events matter — define exit precisely).
- Median time-to-churn on young cohorts: most haven't churned yet, so the median is an artifact.
- Comparing churn rates across cohorts with different 'active' definitions.
- Reading any single cohort's dip as structural before checking vintage-normalized baselines.
Debugging tips
- If month-1 churn jumps across all vintages at the same date, suspect a tracking or flag change, not a product cliff.
- Time-to-churn outliers (1-day exits) often signal tech issues — crashes, signups with zero value — check the events first.
- Cross-check churn cohort results with the account ledger; payment churn and activity churn rarely agree on timing.
- When two cohorts differ sharply, re-pull their event counts before theorizing — a tenure-based tracking gap is common.
Interview checkpoint
Quick quiz
Month-1 churn concentrated in every recent cohort most strongly suggests:
A churn spike in one vintage only is proof the product is declining overall.