Cohort analysis is an analytics method that groups people or entities by a shared starting event within a defined time window, such as the month of signup or first purchase, and then tracks each group’s behavior over time measured from that starting point. In analytics, it is the standard way to answer whether newer customers behave better or worse than older ones, a question that overall totals cannot answer.
The method separates two effects that aggregate metrics mix together: growth from new users and behavior of existing users. A product can show rising monthly active users while every cohort retains worse than the one before, because heavy acquisition hides the decay. Cohort analysis makes that pattern visible, which is why it sits at the center of retention, lifetime value and product-market fit assessments.
How a cohort table works
A cohort analysis is usually presented as a cohort table or triangular heatmap. Each row is one cohort, each column is a period since the starting event, and each cell shows a metric such as the share of the cohort still active. Because recent cohorts have had less time, the table forms a triangle.
| Signup cohort | Users | Month 0 | Month 1 | Month 2 | Month 3 |
|---|---|---|---|---|---|
| January | 1,200 | 100% | 42% | 33% | 29% |
| February | 1,450 | 100% | 45% | 36% | 31% |
| March | 1,600 | 100% | 47% | 38% | Not yet observed |
| April | 1,900 | 100% | 44% | Not yet observed | Not yet observed |
Reading down a column compares cohorts at the same age, which shows whether product changes improved retention. Reading along a row shows how one cohort decays over time. Reading along a diagonal shows all cohorts in the same calendar month, which exposes seasonal events or outages that hit everyone at once.
Defining cohorts and metrics
The usefulness of cohort analysis depends on three explicit choices: the event that defines membership, the period length and the metric measured in each period. Changing any of them can change the conclusion.
| Cohort type | Grouping rule | Question it answers |
|---|---|---|
| Acquisition cohort | Date of signup, install or first purchase | Do newer customers retain better than older ones |
| Behavioral cohort | An action taken, such as using a feature in week one | Which early behaviors predict long-term retention |
| Channel cohort | Acquisition source or campaign | Which channels bring customers who stay |
| Plan or segment cohort | Pricing plan, region or company size | Where retention and revenue are strongest |
Period length should match the natural usage cycle of the product. Daily cohorts suit games and messaging apps, weekly cohorts suit most consumer software, and monthly or quarterly cohorts suit subscriptions and business software with long billing cycles. Common metrics include user retention, revenue retention, orders per user and cumulative revenue per user.
Retention definitions that change the result
Retention sounds simple, but there are several incompatible definitions. Teams must agree on one and state it on every chart, because the numbers can differ widely for the same data.
| Definition | A user counts as retained when | Effect on numbers |
|---|---|---|
| Classic or period retention | Active in that specific period | Strict, can fluctuate |
| Rolling or unbounded retention | Active in that period or any later one | Higher, smooth, never increases over time |
| Revenue retention | Measured as revenue kept from the cohort | Can exceed 100% through upgrades |
| Logo retention | Customer account still paying | Ignores expansion and contraction |
Net revenue retention above 100% means expansion from remaining customers outweighs churn and downgrades, a pattern common in business software with usage-based or seat-based pricing. In consumer products, healthy user retention curves typically flatten after an initial drop; a curve that keeps falling toward zero suggests the product has not found a durable audience.
Cohort analysis compared with related methods
Cohort analysis complements rather than replaces other analytics methods. The key distinction is that it indexes behavior by time since a starting event, not by calendar time or by funnel step.
| Method | Unit of analysis | Main output | Typical use |
|---|---|---|---|
| Cohort analysis | Groups by shared start date or trait | Behavior curves over relative time | Retention, lifetime value, product changes |
| Funnel analysis | Sequence of steps in one journey | Conversion and drop-off per step | Onboarding and checkout optimization |
| KPI tracking | Totals per calendar period | Trends of headline metrics | Operational monitoring |
| Predictive analytics | Individual users or accounts | Probability of future behavior | Churn prediction, targeting |
Pitfalls in cohort analysis
Cohort analysis is easy to compute and easy to misread. Most errors come from small cohorts, inconsistent definitions or confounded comparisons.
- Small cohorts produce noisy percentages: a cohort of 80 users can move several points by chance, so compare cohorts only when sizes are sufficient or show confidence intervals.
- Incomplete periods in the latest cohort look like a sudden drop; exclude periods that have not fully elapsed.
- Mixed acquisition channels can make a cohort look worse simply because a cheaper, lower-intent channel grew that month.
- Changing definitions of active over time break comparability between old and new cohorts.
- Survivorship bias appears when revenue per user is computed only over retained users rather than the original cohort size.
Cohort tables show correlation in time, not causation. A better March cohort may reflect a product change, a seasonal audience or a campaign. To attribute an improvement to a specific change, combine cohort analysis with a controlled experiment whenever possible.
Stripped to the essentials
Cohort analysis is like comparing school classes by the year they started. Instead of counting all students together, each starting class is followed on its own. The question is how many of each class are still there after one year, two years and three years.
For a business, the classes are customers who joined in the same month. If newer groups stay longer than older groups, the product is improving. If they leave faster, something is wrong, even when the total number of customers still grows.
Where it is applied
A meal-kit subscription service sees total subscribers rising every month and assumes the business is healthy. An analyst builds a monthly acquisition cohort table of order retention for the past year.
The table shows that cohorts acquired during a heavy discount campaign keep only about one in five customers after three months, while cohorts acquired through referrals keep roughly twice that share. Growth in totals had been driven by expensive discount acquisition that mostly churned after the promotion ended.
The company shifts budget from deep discounts toward referral incentives and adds a recipe-selection step in the first week, a behavior that the behavioral cohort analysis linked to higher retention. Over the next two quarters, new cohorts retain noticeably better at month three, and marketing spend per retained customer falls.