Amey-Thakur Cohort analysis
Blended metrics lie by mixing users of different ages and vintages. Cohort analysis fixes this by grouping users by a shared starting point (usually signup month) and following each group over time, so you see whether the product is actually getting better or just growing over a worsening base.
Method
- Define the cohort by a meaningful shared start. Most often signup or first-purchase period (weekly or monthly cohorts), but any shared origin works (users from a campaign, a feature launch). The cohort's defining event is what you group by; the metric is what you track across their lifetime.
- Track the metric by cohort age, not calendar time. Line up each cohort at "month 0, month 1, ..." from their own start, so you compare cohorts at the same age. This is what separates "our month-1 retention" (a real, comparable number) from a calendar metric that blends all ages.
- Read the cohort table both ways. Down a column (same cohort age across vintages): is a newer cohort retaining better or worse than an older one at the same age? That tells you if the product is improving. Across a row (one cohort aging): the retention/decay curve shape. Both readings carry different insight.
- Use it to expose what blending hides. A flat or rising blended metric can mask every cohort decaying, propped up by growth (new users replacing churned ones). Cohorts reveal this: if each successive cohort is worse at the same age, you have a worsening product hidden by a growth story (see the blending warning in saas-metrics, churn-analysis).
- Look for the retention plateau. Healthy products retain a stable floor of users after early churn (the curve flattens); a curve that decays to zero means no lasting value. Where the curve stabilizes, and at what level, is a core signal of product-market fit (see user-activation, churn-analysis).
- Segment cohorts by acquisition and type. Cohorts from different sources, plans, or first-actions retain differently; segmenting reveals which acquisition or onboarding produces users who stick (a channel whose cohorts decay fast is a quality problem, not a volume win; see funnel-analysis, correlation-causation on selection).
Boundaries
- Cohort analysis shows patterns over time; it does not explain the cause. A worsening cohort trend is a signal to investigate (a product change, a shift in acquisition mix), not a diagnosis (see correlation-causation).
- Cohort tables need enough users per cohort to be stable; thin cohorts are noisy, so widen the period or wait for volume before reading small differences.
- Churn analysis is cohort analysis applied specifically to retention and churn drivers; this skill is the general technique, applicable to any metric by cohort (see churn-analysis for the churn-focused method).