Track groups who share a starting characteristic — usually their signup month — over time. It separates real product improvement from the illusion created by a flood of new users.
▸ Try the interactive toolCohort analysis tracks the behaviour of specific groups of users over time, where each group shares a defining characteristic at a point in time. The most common is acquisition date — users who signed up in one month are a cohort, the next month another — and you watch how each group behaves in the weeks and months after.
Its essential power is separating real change from composition change. A blended, all-users metric can rise simply because you acquired a flood of new users, masking the fact that retention is actually getting worse. Cohort analysis fixes the starting point for each group, so you can see whether the product is genuinely improving — whether March's cohort retains better than January's — rather than being fooled by growth that papers over decay.
Group users by a shared start point (usually signup period). Track each group's behaviour over the weeks/months after. Comparing cohorts shows whether the product is actually improving — not just whether the blended number is moving.
Blended metrics lie by aggregation. A headline number can look healthy while the underlying product is decaying, because new-user volume hides falling retention.
| The shortcut | What it costs | What it gives you instead |
|---|---|---|
| Blended metrics hide decay | Overall numbers rise on new-user volume while retention falls. | Cohorts isolate each group, exposing the real trend. |
| Can't tell if changes worked | Did the product improve, or did the mix change? | Comparing cohorts before and after a change reveals true impact. |
| Averaging across tenures | New and old users blended into one meaningless number. | Cohorts compare like-for-like by time since start. |
| Survivorship illusions | Looking only at remaining users flatters the picture. | Cohort retention curves show who left, not just who stayed. |
Usually acquisition period (signup month/week), but it can be any shared starting trait — the feature they first used, the plan they joined on. The characteristic defines what you can learn.
Follow each group's key behaviour — retention, usage, revenue — across the periods after their start. The result is a curve per cohort, not a single number.
The core move: does a later cohort behave better than an earlier one at the same age? Improving curves mean the product is genuinely getting better; flat or declining means it isn't, whatever the blended number says.
When a cohort's curve improves or worsens, tie it to what changed for that group — a feature, an onboarding tweak, a different acquisition channel. This is how you attribute impact.
Early retention can look fine while later periods decay. Read the whole curve — the shape over time is the real health signal.
A team watched its monthly active users climb steadily and felt confident the product was healthy. The blended number rose every month, so nobody looked closer.
A cohort analysis broke the spell. Each new signup cohort was actually retaining worse than the one before — the rising headline number was pure acquisition volume masking steadily worsening retention. The product wasn't getting healthier; it was acquiring fast enough to hide that it was getting sicker. Only fixing the starting point per cohort made the decay visible, while there was still time to address it.
The deliverable is a cohort grid — each row a starting group, each column a period after start — read for whether later cohorts outperform earlier ones.
| Blended metric shows | Cohort analysis shows |
|---|---|
| MAU is rising | Whether each cohort retains better or worse |
| A single health number | The trend across acquisition groups |
| Growth (maybe masking decay) | Real product improvement, isolated |
| Who's here now | Who stayed and who left, by cohort |
A blended metric answers “are the numbers up?” Cohort analysis answers the question that actually matters: “is the product getting better?” — and the two answers are often opposite.
The trap cohort analysis defends against is one of the most common in product: confusing growth with health. When acquisition is strong, the blended number rises regardless of whether the product is improving or decaying, so a team can feel successful right up until growth slows and the underlying retention collapse is suddenly exposed. By fixing each group's starting point and comparing like-for-like, cohort analysis strips out the composition effect and shows the truth early — whether March's users are sticking around better than January's. That's the difference between noticing decay while you can still fix it and discovering it when the growth that hid it runs out.
Aggregates hide composition change. Always check the cohort view.
Retention can look fine at week one and collapse by week eight. Read the whole curve.
A single cohort in isolation tells you little. The comparison is the insight.
A changed curve is a clue, not a conclusion — connect it to what changed for that group.
Cohort analysis adds the time-and-group dimension to raw behavioural data (Tool 09).
When a cohort's curve shifts, qualitative research reveals why (Tools 05, 08).
Comparing pre- and post-change cohorts confirms whether a change actually improved retention.
Cohort analysis and retention curves are core to the metrics and growth-accounting work in Module 5.
Imagine a product whose total active users rise every month. List two ways that rising number could coexist with the product actually getting worse.
Now describe what a cohort table would show in that situation — and how it would reveal the decay the blended number hides.
If you can see how growth and decay coexist in the same metric, you understand exactly why blended numbers can't be trusted alone — and why cohort analysis is the antidote.