Cohort Analysis: Reading Retention the Way It Actually Behaves

Overall month-one retention has sat at 30% for a year. Flat. Depressing. Then someone builds a cohort table and discovers every new signup cohort retains better than the one before — January keeps 25%, June keeps 38%. The product is improving steadily; the blended number hides it because each month's flood of new users dilutes the maturing cohorts. Flat blended retention plus growth equals hidden progress. Or hidden decay, if the cohorts run the other way — either way, the blend cannot tell you.

How a Cohort Table Works

Group users by the period they started (the cohort), then track each group's retention at age 1, 2, 3… periods. Rows are cohorts, columns are ages, cells are the share still active. Read down a column to compare cohorts at the same age ("do newer cohorts retain better at month 3?"); read across a row to see one cohort's decay curve.

Two construction details matter enormously. First, define "active" crisply — a meaningful action, not "opened the app." Vanity activity definitions produce vanity curves. Second, use complete periods only: a cohort that signed up six days ago has no meaningful "week 2" cell yet, and including partial cells bends every curve upward at the right edge.

What Healthy Looks Like

Retention curves almost always decay fast, then flatten. The shape to hope for is a flattening curve that stabilizes above zero — the flat part is your habitual user base. A curve still falling steeply at month six has found no habit yet. Comparing curve shapes across cohorts matters more than any single number: a new cohort that flattens higher, even from a lower start, is the signature of genuine improvement.

Watch for the two classic distortions. Survivorship in disguise: "average revenue of active users" rises over time partly because low-value users churn first — the survivors were always better. Mix shifts: a viral spike fills a cohort with tourists who churn fast; judging that cohort against organic ones repeats Simpson's paradox in cohort form. Annotate acquisition spikes on the table or split cohorts by channel.

From Table to Decision

Cohorts earn their keep when they change a decision. Three questions they answer well:

• Did the onboarding revamp work? Compare month-1 retention of pre- and post-launch cohorts at the same age. No model needed — just the column.
• When do users decide? The steepest part of the curve is where the habit forms or fails. Put research and onboarding effort there, not evenly everywhere.
• What is a user worth? Multiply the stabilized retention curve by per-period revenue for a bottom-up lifetime value — far more honest than blended ARPU times a guessed lifespan.

Present the table as a heatmap (color by cell value) and the rows as decay curves on one chart — pick the chart by the metric's taxonomy, lines for curves, tables for precise comparison. In KPI Master, dated event data breakdowns give you the raw material; the cohort table itself is one pivot away. Build it once, and the blended number will never fool you again.