User engagement cohort analysis

User engagement cohort analysis groups users by a shared starting point, usually the week or month they signed up, and tracks how engaged each group stays over time. Rather than reporting one engagement figure for the whole base, it lays out a grid where each row is a cohort and each column is a period since they joined. Reading down a column shows how a given week of activity compares across cohorts. Reading across a row shows how a single cohort decays or holds.

This structure matters because a blended engagement number can stay flat while the underlying health changes completely. A product can grow its active user count, masking the fact that every new cohort engages less than the one before it. Cohort analysis separates the effect of bringing in more users from the effect of those users actually sticking around, which a single average cannot do.

The analysis is built on a clear definition of engaged, the same discipline that anchors any retention rate measure. Engaged might mean returning, completing a core action, or hitting an activity threshold. Once that definition is fixed, the cohort grid shows whether the experience is getting better or worse for the people who join after each change you ship.

There is no single formula, because cohort analysis is a grid of values rather than one number. Each cell is its own simple percentage: the share of a cohort still engaged in a given period. The work is in defining the cohorts, the engagement signal, and the periods consistently, then computing the same ratio for every cell.

1. 1

   Cohort definition

   The shared trait that groups users, most often the sign-up week or month. You can also cohort by acquisition channel, plan, or first feature used. Every user belongs to exactly one cohort, fixed at entry and never reassigned.

2. 2

   Engagement signal

   The action that marks a user as engaged in a period, such as returning, completing a core task, or crossing an activity threshold. This is the numerator condition and must be identical for every cohort and every period.

3. 3

   Period grid

   The columns of the table, measured as time since the cohort started rather than calendar time. Week zero is the starting period, week one is the following period, and so on. Aligning on time-since-start is what makes cohorts comparable.

4. 4

   Cell value

   For each cohort and period, the count of engaged users divided by the original cohort size, times 100. Fill the whole grid this way to see retention curves stacked across cohorts.

A worked example clarifies the grid. The January cohort enters with 500 users. In week one, 300 of them are engaged, giving 60 per cent. In week four, 180 are engaged, giving 36 per cent. The February cohort enters with 520 users and shows 65 per cent in week one and 41 per cent in week four. Comparing the two rows tells you February users are holding better than January users, which points to a change made between the two intakes. That comparison is invisible in a blended monthly engagement figure.

A cohort grid tells you which cohorts decay and when, but not why. A metric tree decomposes cohort engagement into the drivers underneath each curve and attaches each one to the team that owns it. Together, the grid locates the problem in time and the tree explains the cause.

The first level splits cohort engagement into onboarding quality, ongoing value, and re-engagement. Onboarding quality drives the early periods, where a weak first experience shows up as a steep drop in weeks one and two. Ongoing value drives the middle of the curve, where users decide whether the product earns a place in their routine. Re-engagement drives the tail, where lifecycle messaging and habit loops decide whether lapsing users come back. A cohort that falls off early has an onboarding problem. A cohort that fades slowly has a value problem. The shapes are different and so are the owners.

KPI Tree connects each branch to the action and team that influences it, with RACI ownership so the onboarding branch sits with product, the value branch with the team that ships core features, and the re-engagement branch with lifecycle. When a new cohort decays faster than the last, the change is pushed to the owner of the branch where the curve broke, and the verified impact loop checks whether their fix actually flattened the next cohort. That closes the gap between a dashboard that shows a worsening curve and a team that knows which lever to pull.

Cohort engagement benchmarks depend heavily on how often the product is naturally used and how engaged is defined. A daily-use tool and a quarterly-use tool will have completely different healthy curves, so a single benchmark number is misleading. The ranges below describe the shape of week-four engagement for a monthly active definition. Anchor your target to your own usage frequency and baseline curve.

The single most important read is whether newer cohorts beat older ones. A grid where each successive row sits above the last means the changes you are shipping are improving the experience for the people who arrive after them. That trend matters more than any individual cell, because it tells you the product is compounding rather than coasting.

You do not improve a cohort analysis, you improve the curves it reveals. The discipline is to read where a cohort breaks, fix that stage, and then watch the next cohort to confirm the fix worked before moving on. Because each cohort is a fresh population, the grid is a built-in experiment for every change you ship.

1. 1

   Using calendar time instead of time since start

   If the columns are calendar months rather than periods since the cohort joined, cohorts of different ages are compared at the wrong points and the curves no longer line up. Always measure the grid as time since each cohort started.

2. 2

   Letting young cohorts mislead the read

   A cohort that is only two weeks old has no week-eight data, and reading its short curve as if it were complete invites wrong conclusions. Compare cohorts only at periods every cohort in the comparison has actually reached.

3. 3

   Redefining engaged partway through

   Changing what counts as engaged after some cohorts are already plotted breaks comparability. Earlier rows reflect the old rule and later rows the new one, so the trend down a column is an artefact. Fix the definition and restate history if it must change.

4. 4

   Ignoring cohort mix

   If acquisition shifts toward a lower-intent channel, newer cohorts can look worse for reasons that have nothing to do with the product. Segment by channel or plan so a change in who you acquire is not mistaken for a change in the experience.