Cohort Analysis
Also known as: retention cohort, cohort retention, cohort table, retention analysis, birth cohort, acquisition cohort
Cohort analysis groups individuals (customers, users, patients) who share a common defining characteristic (typically when they started using a product, when they were first treated, or when they made their first purchase) and tracks their behaviour over time.
By comparing cohorts that started at different times, you can see whether product changes, seasonal effects, or external factors influenced outcomes, and you avoid the misleading aggregate statistics that combine groups at very different stages of their lifecycle. Cohort analysis is fundamental to understanding retention, churn, and lifetime value.
Cohort Analysis, explained at three levels
At home
Cohort analysis is about following groups of people over time, starting from the same point. Imagine a gym tracks when members joined and measures whether they are still members each month. Members who joined in January 2023 are one cohort; members who joined in February 2023 are another. Looking at each cohort's retention separately (how many from each cohort are still active after 1, 3, 6, 12 months) reveals whether the gym is getting better at keeping members: a new cohort with much better 6-month retention than older cohorts shows that something improved. Mixing all cohorts together in one average would hide this improvement.
At work
Cohort analysis is essential for subscription and SaaS businesses because aggregate metrics hide important dynamics. If a company's total monthly active users is growing, it could mean existing users are more engaged, OR it could mean lots of new users are joining but existing users are churning rapidly. Cohort analysis separates these: you can see whether users who joined 12 months ago are still active (retention) independently of how many new users joined this month. A business with growing new user acquisition but declining cohort retention has a problem that aggregate MAU growth conceals.
At home
The standard cohort retention table: rows represent cohorts (typically defined by signup month), columns represent months since signup (0, 1, 2, 3...), and each cell shows the percentage of the cohort still active at that time point. Month 0 is always 100% (by definition). Reading down a column shows how each successive cohort performed at the same time point (improving or declining?). Reading across a row shows how one cohort's retention evolves over time. The diagonal elements (same calendar month across different cohort rows) show the current state: what fraction of each cohort is active today. A well-formatted cohort table immediately reveals improving or declining product-market fit across cohorts.
At work
Cohort analysis beyond retention: the same framework applies to revenue cohorts (tracking revenue per user over time for each acquisition cohort), conversion cohorts (tracking purchase conversion rates over time after first visit), and engagement cohorts (tracking feature usage intensity over time). Revenue cohort analysis reveals whether customer lifetime value (LTV) is changing across acquisition cohorts: are recently acquired customers spending more or less over time than customers acquired 12 months ago? This directly informs how much you should spend on customer acquisition (maximum CAC is a function of expected LTV).
At home
Age-period-cohort effects in demographic and economic analysis: observing the same outcome across time involves three confounded effects. Age effects: behaviour changes as individuals age (people become more conservative with money as they age). Period effects: the same calendar period affects everyone (a recession affects all cohorts simultaneously). Cohort effects: the cohort you were born in or joined affects outcomes (people who entered the workforce in a recession have lower lifetime earnings than those who entered in booms). Separating these three effects in data is analytically challenging (they are perfectly multicollinear in balanced panel data) and requires additional assumptions or design choices.
At work
Cohort analysis for measuring product impact: when you ship a major product change, users who join after the change constitute a new cohort whose behaviour can be compared to pre-change cohorts at the same tenure points. This design approximates a natural experiment: post-launch cohorts experienced the new product from their first day; pre-launch cohorts experienced the old product from their first day. Comparing 30-day and 90-day retention for pre- and post-launch cohorts (at the same points in their tenure) isolates the product change's effect from seasonality and market changes. This is more rigorous than simply comparing aggregate metrics before and after launch.
Real-world examples
At home
Sophie ran a small fitness app as a side project and checked the cohort view in its analytics dashboard. She noticed that users who joined in January (New Year resolution cohort) had very low 60-day retention: most had dropped off by March. Users who joined in September (back-to-school cohort) had much better 60-day retention: they were still active in November. This pattern matched her own experience: she was more likely to stick with exercise habits she started in autumn than January resolutions. The cohort analysis revealed that her product's value wasn't uniform across acquisition periods: seasonal timing of joining affected habit formation.
At work
Jerome ran a SaaS product team. His aggregate monthly active user chart was growing steadily and looked healthy. A cohort retention analysis told a different story: cohorts acquired in the first year had 40% 12-month retention; cohorts acquired in the most recent year had only 22% 12-month retention. The aggregate growth was masking serious retention deterioration: more new users were joining, but each cohort was churning faster. This finding triggered an investigation into product changes and onboarding degradation that the aggregate metrics had completely hidden.
Common confusions
- Many analysts use aggregate monthly active user or customer count metrics and assume they reveal product health — Aggregate active user counts mix together users at very different lifecycle stages (day 1 users vs. year 3 users) and mix together different acquisition cohorts that may have very different retention patterns. A growing aggregate MAU can simultaneously be experiencing worsening retention if acquisition growth exceeds churn. Declining MAU can occur even when retention is improving if acquisition slows. Cohort analysis is the tool that separates these effects and reveals whether underlying product-user relationships are improving or deteriorating, independent of acquisition volume changes.
- Some assume that a cohort with fewer users than expected at a time point necessarily indicates churn — Users counted as 'lost' in cohort analysis may have paused, temporarily become inactive, or be in a different product tier, not necessarily churned permanently. The definition of 'active' matters significantly: a user who logs in once every 60 days counts as 'active' under a 90-day definition but 'inactive' under a 30-day definition. Additionally, some users return after apparent inactivity (resurrection rate). For businesses with natural usage patterns (quarterly or annual software, seasonal service), daily or monthly active user definitions may misclassify many retained users as churned.
For analysts
Cohort analysis is one of the highest-value analytical frameworks for subscription and recurring revenue businesses. When reviewing a client's customer analytics, asking 'can you show me cohort retention by acquisition month?' often reveals dynamics invisible in aggregate metrics. The diagnostic questions: are recent acquisition cohorts retaining better or worse than older cohorts at the same tenure points? Is the shape of the retention curve flattening early (quick initial churn then stability) or declining continuously (ongoing churn throughout the user lifecycle)? Are high-cost acquisition channels producing cohorts that retain better or worse than low-cost channels? These questions directly inform product, marketing, and unit economics decisions.
History & origin
Cohort analysis originated in demographic and epidemiological research in the early 20th century, tracking mortality and health outcomes for birth cohorts (groups born in the same year). The method was adopted in social science research (tracking outcomes for school cohorts: students who started in the same year) and economic research (tracking income for cohorts entering the job market in different years). Consumer internet companies brought cohort analysis into product analytics in the 2000s-2010s: tracking retention curves for users who signed up in different months revealed how product improvements (or deteriorations) affected user behaviour across different starting groups. Andrew Chen, Dave McClure, and Alistair Croll popularized cohort analysis as a core SaaS metrics practice through startup education content in the 2010s.