Cohort analysis

Cohort analysis is a method of grouping customers or users by a shared starting characteristic, usually the month they signed up, and tracking how that group behaves over the weeks and months that follow. Instead of looking at one blended average across your whole base, you watch each group age on its own timeline. A cohort that joined in January is followed through February, March, and April, and a cohort that joined in February gets its own row beneath it.

This matters because a single blended number hides what is really happening. If overall retention rate is flat, that could mean every cohort is steady, or it could mean older cohorts are decaying while newer ones improve and the two cancel out. Cohort analysis pulls those two stories apart. It answers the question that a headline average cannot: are the customers we acquire today better or worse than the ones we acquired last year.

The most common form is a retention cohort, where each cell shows the share of the original group still active in a given period. But the same structure works for revenue, for feature adoption rate, or for any behaviour you can attribute back to a joining date. The grid is the same. Only the value inside each cell changes.

There is no single number for cohort analysis. The output is a grid, and each cell is a simple ratio. To build it, fix a cohort by its starting period, then for each later period count how many of that original group are still active and divide by the original size. Repeat for every cohort and every period.

1. 1

   Define the cohort grouping

   Choose the shared event that places a user into a cohort. Most often this is the signup month, but it can be the first purchase week or the quarter a contract started. Be consistent so every user lands in exactly one cohort.

2. 2

   Record the original cohort size

   Count the members who joined in the starting period. If 400 customers signed up in January, the January cohort size is 400 and that figure becomes the denominator for every later period in that row.

3. 3

   Count active members in each later period

   For each subsequent period, count how many of those original 400 are still active, defined the same way every time, whether that is a paid subscription, a login, or a purchase.

4. 4

   Calculate the cell value

   Divide active members by original size and express it as a percentage. If 280 of the January cohort are active in March, the March cell for January reads 70 percent. Plot every cohort row to form the grid.

Reading the grid is a two-direction exercise. Reading across a single row shows the retention curve for one cohort as it ages, which tells you how sticky the product is once someone is in. Reading down a column compares cohorts at the same age, which tells you whether the customers you acquire are getting better or worse over time. A healthy product shows curves that flatten rather than fall to zero, and newer cohorts sitting at or above older ones at the same age.

A cohort grid tells you which group is decaying. A metric tree tells you why, and who can do something about it. The two work together. The cohort surfaces the anomaly, and the tree decomposes the retention of that cohort into the drivers that teams actually control.

The first level splits the retention of a struggling cohort into its causes: the quality of acquisition for that period, how quickly those users reached value, the depth of their early engagement, and the friction they hit when their first bill came due. Each of those decomposes further. Acquisition quality breaks into channel mix and ideal-customer fit. Time to value breaks into onboarding completion and first meaningful action. Each leaf maps to a team.

This is the gap between a dashboard and a decision. A cohort chart shows a cliff in the March group, but it does not assign the work. The tree connects the cliff to the onboarding flow product owns, the channel mix marketing owns, and the early-life outreach customer success owns. With RACI ownership on each branch, the accountable owner is pushed the drop the moment it appears rather than discovering it a quarter later in a review.

Cohort benchmarks depend heavily on business model, so the useful comparison is the shape of the curve rather than a single percentage. The question is not just how high retention sits at month one, but whether the curve flattens into a stable plateau or keeps sliding toward zero. A flattening curve signals a durable base. A curve that never flattens signals that the product has no natural retention floor.

When you compare your own cohorts down a column, a useful rule is that newer cohorts should sit at or above older ones at the same age. If month-3 retention for recent cohorts is consistently below month-3 retention for cohorts a year ago, acquisition quality or onboarding has degraded even if the blended retention number looks stable. That column comparison is where cohort analysis earns its place over a single headline figure.

Improving cohort analysis means two distinct things: making the analysis itself sharper, and acting on what it reveals to lift the curves. The first is about discipline in how you build cohorts. The second is about routing each finding to the team that owns the lever.

The decomposition matters most when a cohort underperforms. KPI Tree lets you take the worst cohort, break its retention into the acquisition, onboarding, engagement, and renewal branches, and connect each branch to the team that owns it. When the March cohort cliff appears, the verified impact loop then checks whether the onboarding fix product shipped actually lifted the next cohort, rather than leaving you to guess whether the intervention worked. That closes the distance between spotting a decaying group and proving you fixed it.

1. 1

   Reading only the blended average

   Collapsing all cohorts into one retention number defeats the purpose. The blended average can stay flat while older cohorts decay and newer ones improve. Always keep cohorts separated so the two stories stay visible.

2. 2

   Letting members switch cohorts

   A cohort is fixed by its starting event. If you move users between groups when they upgrade or lapse, the denominator shifts and the retention curve becomes meaningless. Hold the original group constant for the life of the analysis.

3. 3

   Changing the active definition midway

   If active means a login in one period and a purchase in another, the cells are not comparable. Pin a single definition of active before you build the grid and apply it identically across every cohort and period.

4. 4

   Ignoring cohort size

   A cohort of 12 users that drops to 8 looks like catastrophic churn but is just noise. Always show the original size next to the curve so small, statistically thin cohorts are not over-interpreted.

5. 5

   Stopping at observation

   Spotting a decaying cohort is only half the job. Without decomposing the drop into its causes and assigning an owner, the analysis describes a problem nobody is tasked with solving.