A single retention number lies to you. “60% of users are still active” sounds fine until you realize it’s blending the loyal customers you won two years ago with the ones who signed up last week and are about to vanish. Cohort analysis takes that mushy average apart, grouping users by when they arrived so you can see whether the product you’re shipping today actually holds onto people better than the one you shipped six months ago.
What cohort analysis is
Cohort analysis is a method of grouping users by a shared starting point in time and then tracking how each group behaves over the days, weeks, or months that follow. The most common grouping is acquisition date, all the users who signed up in January form one cohort, February another, and so on, but a cohort can be any shared origin event: first purchase, free-trial start, or the month they downloaded your app.
The power comes from the comparison across cohorts over time. Instead of one retention figure, you get a curve for each group, and laying those curves side by side reveals whether your changes are working. If your March cohort is still 40% active at week eight while your January cohort had dropped to 25% by then, something you did between January and March is paying off.
How it differs from cluster analysis
These two are constantly confused, so here’s the clean distinction. Cluster analysis groups people by similarity across many behavioral attributes and the algorithm discovers the groups for you; it’s typically a snapshot. Cohort analysis groups people by a shared moment in time that you define up front, and its entire purpose is to watch that fixed group change as time passes. In short: clusters answer “what kinds of customers do we have right now?” while cohorts answer “is what we’re doing getting better or worse over time?”
Reading a cohort table
The classic output is a triangular table. Each row is a cohort (say, sign-up month), each column is the time elapsed since they joined (month 0, month 1, month 2), and each cell shows the percentage still active or the revenue they generated. Two patterns matter most when you read it:
- Read across a row to see how a single cohort decays over time. A steep early drop tells you the problem is in onboarding, not in the long-term product.
- Read down a column to compare the same life stage across cohorts. If “month 1 retention” is climbing as you move down to newer cohorts, your recent changes are improving how well you hold onto people.
What we use it for
From our agency experience, cohort analysis is the most honest way to judge whether a change actually helped, because it isolates groups that experienced that change from groups that didn’t. A few applications that consistently earn their keep:
- Measuring onboarding and product changes. Ship a new welcome flow, then watch whether cohorts that signed up after it retain better than the ones before. The cohort table shows the effect cleanly where a blended average would hide it.
- Comparing acquisition channels. Group users by where they came from and track retention. What we consistently see is that the cheapest channel to acquire often produces the worst-retaining cohort, a truth that only shows up over time, never on day one.
- Tracking revenue maturity. Layer spend onto each cohort to see whether newer customers are reaching the same lifetime value as older ones, faster or slower.
When we run this for clients, the most useful first chart is almost always retention by acquisition month. It answers a question every executive cares about, “are we actually getting better at keeping customers?”, in a single picture.
Getting it right
The biggest pitfall is reading too much into young or thin cohorts. A cohort that’s only two weeks old can’t tell you anything about three-month retention, and a cohort of forty users will swing wildly on noise alone. Give cohorts enough size and enough time before drawing conclusions. Watch out, too, for seasonality, a holiday-season cohort may behave nothing like a summer one, so compare like with like before you credit a product change.
Frequently asked questions
What’s the difference between a cohort and a segment?
A segment groups users by an attribute that can change (current plan, country, device). A cohort groups users by a fixed historical event, usually when they started, and that membership never changes. That permanence is what lets you track the exact same group over time.
How do I actually build a cohort analysis?
Pick the defining event (most often sign-up date), choose the metric you care about (retention, revenue, orders), pull the data, then arrange it so each row is a cohort and each column is time-since-start. Tools like Google Analytics and most product-analytics platforms can generate these tables for you.
How long should I track a cohort?
Long enough to cover the behavior you care about. For a habit-forming app, the first few weeks reveal most of the story. For a considered annual purchase, you may need a year or more before a cohort’s true retention shape is clear.
Related terms
- Cluster Analysis — groups users by behavioral similarity discovered by an algorithm, rather than by a shared start date.
- Retention Rate — the core metric most cohort tables track over time.
- Churn Rate — the inverse of retention; cohort curves make churn timing visible.
- Customer Lifetime Value — cohorts let you see how lifetime value matures for each group of customers.
- Google Analytics — a common tool for building cohort reports from real user data.
