Cohort Analysis

Cohort analysis groups users by a shared starting event — usually the week or month they signed up — and tracks each group separately over time. The technique reveals patterns that a single aggregate metric hides, especially when product changes or marketing shifts have changed who's joining.

Why aggregate metrics mislead

A product can have rising aggregate retention while every individual cohort's retention is falling — if new cohorts are growing faster than old cohorts are decaying. This is Simpson's paradox applied to growth metrics. Cohort analysis prevents the trap by separating each group's curve.

How to read a cohort table

The canonical view is a triangular table:

• Rows: cohorts (one per signup week/month)
• Columns: weeks/months since signup
• Cells: % of cohort still retained (or paying, or returning, depending on the metric)

Best-in-class teams look for two patterns:

1. Curves flattening — does the retention drop slow down or stabilize? A flat tail means the product found durable value for the surviving users.
2. New cohorts above old cohorts — are recent cohorts retaining better than older ones at the same age? If yes, the product is improving. If no, recent product changes may have hurt.

When cohort analysis fails

It fails when cohorts are too small to be statistically meaningful (typically under 100 users) or when the shared event isn't actually predictive of behavior (random sampling cohorts is rarely useful). It also fails when teams use it as a vanity exercise without acting on the curve shape — the technique is a diagnostic, not a metric for dashboards.

Related

• Retention — what cohort analysis usually measures
• Activation — the gating event that determines if a cohort begins
• Churn — the leakage shape each cohort eventually shows