What is Cohort Analysis?

Cohort analysis is a form of behavioral analytics that divides users into groups (cohorts) sharing a defining event or attribute, then measures an outcome for each group at successive time intervals after the defining event. The most common cohort is acquisition-based: everyone who signed up in a given week or month, tracked for activity, retention, or revenue in each subsequent period. Behavioral cohorts group by an action instead (users who enabled a feature, who came from a channel, who hit the activation event), and compare their trajectories with those who didn't. The output is usually a table or a set of curves: rows for cohorts, columns for periods, cells for the outcome. The term borrows from epidemiology, where a cohort study follows groups defined by exposure over time, and the logic is the same: hold the starting point constant and watch what happens next.

Reading a Cohort Table

Read across a row to see one cohort's trajectory: the January signups retained 60% in month one, 45% in month two, flattening at 35%. A curve that flattens is the signature of a retained core; one that decays toward zero is a product people try and leave, the reading that product-market fit assessments depend on. Read down a column to compare cohorts at the same age: are recent cohorts retaining better at month two than older ones did? That is the only honest way to know whether a change improved retention, because it compares users at the same stage rather than mixing new users with veterans. Read the diagonal to see calendar effects: a bad month affecting every cohort at once (an outage, a pricing change, a holiday) appears as a diagonal stripe.

What It Answers That Averages Can't

Is retention improving? A blended monthly active user count can rise while every new cohort retains worse than the last, because growth in signups masks the decay; cohorts by column expose it.

Did the change work? Cohorts who joined after a redesign, compared at the same age with those who joined before, the cohort version of a baseline comparison, with the usual caution that other things changed too.

Which behaviors matter? Behavioral cohorts (did versus didn't) with divergent retention curves nominate the actions worth engineering toward, as correlations to be tested.

Which segments and channels bring users who stay? Cohorts by acquisition source, plan, or region, compared on the curve rather than on signup volume.

When does churn happen? The shape of the curve locates the drop: early (an onboarding problem) or late (a value or competition problem), which points research at the right stage.

Doing It Well

1. Define the cohort event and the outcome precisely ("active" means what?), and keep them fixed.
2. Choose the period to match the product's rhythm: daily for consumer apps, weekly or monthly for B2B tools.
3. Wait for cohorts to mature before judging them; a two-week-old cohort has no month-three number, and the empty cells are not zeros.
4. Watch cohort sizes; a small cohort's percentages wobble, and the margin applies to behavioral data as much as to surveys.
5. Separate internal and test accounts, and any cohort with a known anomaly.
6. Pair the divergent cohorts with research: when one cohort's curve differs, the explanation is in the users. Interviews with the well-retained cohort about what they do, and a recorded study of the poorly-retained one attempting the core workflow (a Ballpark study fielded to a customer list drawn from that cohort), turn a curve into a cause.

The Bottom Line

Cohort analysis groups users by when they started or what they did, and follows each group over time, so that retention, the effect of changes, and the behaviors that matter can be seen instead of averaged away. Read rows for trajectories, columns for improvement, diagonals for calendar shocks; define events precisely, let cohorts mature, mind their sizes, and research the ones that diverge. The average tells you how the product is doing; the cohorts tell you whether it's getting better.