Cohort analysis
Cohort analysis compares groups of customers formed by the period in which they arrived. A cohort brings together the customers whose arrival (sign-up, first purchase, subscription) falls in the same month or the same week. The analyst follows each cohort from its arrival, then compares the cohorts at the same age. The cohort table has one row per arrival period and one column per elapsed period. It shows whether recent customers stay longer or spend more than earlier ones. An average taken over all customers mixes this gap with the ageing of the customer base.
Goal
Cohort analysis is a technique for tracking how a product is used. It splits customers into groups by arrival period, then measures the behaviour of each group at the same age. Croll and Yoskovitz define it as the comparison of similar groups over time.
It answers a question that the average leaves open: do recently arrived customers behave better than earlier ones? The cohort table separates the effect of age from the effect of the product.
Croll grounds the technique in release cadence: a young company that is doing things right ships a release every week. Each release raises the same question. Do the customers who got it from the day they joined stay longer than earlier customers? In an interview he cites a churn rate that falls from 20% to 10% and then to 2% across successive releases.
The technique informs product and investment decisions: keep a change, rework the onboarding of new customers, speed up acquisition or wait until retention holds. In their workshop, Croll and Yoskovitz put retention before growth. The deliverable is the cohort table with a note that dates the product changes and states what the analyst concludes from each reading.
Usage
When to use it
- New release shipped: check whether the customers who arrived after the release stay longer than those who arrived before it.
- Average stable or rising: check that fast acquisition does not hide the departure of older customers.
- Retention to manage: track as an indicator the share of each cohort still active at a fixed age.
- Product used repeatedly: a subscription, an app or an online service that the customer comes back to or pays for every month.
- Internal tool rolled out in waves: compare adoption by the staff trained in March and by the staff trained in June.
- Customer value to estimate: derive a customer's lifetime from the decay of their cohort.
- Dated history per customer available: the system keeps each customer's arrival date and events.
When not to use it
- Effect of a single change to prove: two cohorts also differ by season and channel. Prefer an A/B test.
- Customers to contact one by one: a cohort describes a group. Prefer a scoring model evaluated with the lift chart.
- Question about an attribute unrelated to the arrival date: compare segments at the same date.
Description
Cohort, age and measure
A cohort brings together the customers whose entry event falls in the same period. The entry event brings a customer into the population under study: the sign-up, the first purchase, the first subscription, the first log-in to an internal tool. The period is the grain of the split. It is a day, a week or a month. A customer belongs to a single cohort for the whole analysis.
The age counts the periods elapsed since the cohort's arrival. The arrival period is age 0. The entry event of every customer in the cohort falls in that period. By convention, the share of active customers at age 0 is 100%. For a cohort that arrived in January, month 0 is January, month 1 is February and month 2 is March.
The measure is what the analyst tracks in each cohort: the share of customers still active, the revenue per customer, the number of purchases per customer. Its denominator is the cohort size, which is fixed at arrival. A cell of the table gives the measure of one cohort at one age.
The entry event decides what the table measures. A cohort of sign-ups mixes the conversion to paying customers with their loyalty. A cohort of first payments measures the loyalty of paying customers alone.
Cohort, segment, A/B test
Croll and Yoskovitz place the cohort among four ways of slicing data. The A/B test and the multivariate test change the product for one group and measure the result. We contrast these two tests with the cohort and the segment, which observe customers who all use the same product.
| Method | What forms the groups | Sample question |
|---|---|---|
| Cohort | The arrival period. Each group is followed over time | At month 3, does the May cohort keep more of its customers than the January cohort? |
| Segment | An attribute of the customers at a given date | Do Geneva customers cancel more often than Vaud customers? |
| A/B test | A single element changes between two groups | Does the new home page increase sign-ups? |
| Multivariate test | Several elements change at once | Which combination of price and trial offer converts best? |
A segment splits all customers by an attribute and compares them at a given date. A cohort compares groups that arrived at different dates when they reach the same age. The two combine: the analyst can build a cohort table for the customers of one canton or one channel alone.
In an extract released before the publication of their book, the authors propose launching four acquisition campaigns at the same time and splitting the acquired users into separate cohorts. If the team launches one campaign per week, it waits four weeks to learn which one recruits the most users. The four parallel cohorts give the answer in one week. The season affects all four campaigns in the same way.
Building the table
The example in Croll and Yoskovitz's workshop sets one row per month of arrival and one column per month elapsed since arrival. We derive the procedure from this example. It expresses our judgement.
- Set the question and the measure. For a retention question, the analyst tracks the share of customers still active. For a value question, the analyst tracks the revenue per customer. The analyst also defines "active": a purchase in the month, a log-in in the month, a subscription running at the end of the month.
- Choose the entry event and the period. The month suits a monthly subscription. The week suits an app used daily or a team that ships every week. A period that is too fine yields cohorts too small to read.
- Extract the history. The extract holds one row per customer and per event with the customer identifier and the date. The analyst checks it before any calculation: duplicates, missing dates, test accounts, definitions that changed in the source system. Exploratory data analysis provides the method for this check.
- Assign each customer to a cohort. The date of the customer's first entry event fixes their cohort. A customer who cancels and comes back stays in their original cohort.
- Count the size of each cohort.
- Calculate each cell. For each cohort and each age, the analyst divides the number of customers active at that age by the cohort size. For a revenue measure, the analyst divides the revenue at that age by the same size.
- Lay out the table. Cohorts go in rows from the oldest to the most recent. Ages go in columns. A separate column gives the size of each cohort. Recent cohorts have lived through fewer periods and carry fewer cells: the table forms a triangle. One shade per value class helps the eye follow the gaps.
- Read, then date. The analyst reads the table by row, by column and along the diagonal. The analyst then places the releases, the campaigns, the price changes and the incidents on a timeline. Among these events the analyst looks for the explanation of each gap.
Reading the table by row, by column and along the diagonal
A row follows a single cohort over time. It shows the decay: how many customers leave after the first month, then at what age the curve levels off. A row that levels off above zero signals a core of loyal customers. As early as 2008, McClure listed cohort analysis among the retention measures of his AARRR model (Acquisition, Activation, Retention, Referral, Revenue). He linked it to the distribution of visits over time, the rate of decay and the effective customer lifecycle.
To estimate a customer's value, the analyst adds up the cells of a revenue-per-customer row into a running total. The running total at the last observed age is a lower bound on the revenue that one customer of the cohort brings in. On a retention row, the sum of the shares gives a lower bound on the average number of periods during which a customer stays active. We add these two calculations.
A column compares cohorts at the same age. This reading judges the product. If the May cohort retains 74% at month 1 and the January cohort 60%, something changed between the two arrivals. At the same age, ageing does not explain the gap.
A diagonal brings together the cells of one calendar month: month 3 of the January cohort, month 2 of February, month 1 of March. All three fall in April. A dip along a diagonal hits every cohort at once. It points to a dated event: an outage, a season, a price rise, a competitor's campaign.
- Row: one cohort ages
- Column: same age, successive cohorts
- Diagonal: the calendar month of April
Why the average misleads
An overall measure adds up customers of all ages. The pace of acquisition sets the weight of each age in the average. An ageing customer base pulls down the share of active customers, since the older cohorts have lost more customers. Faster acquisition pushes it back up, since the new arrivals have not yet had time to leave. The average moves without the behaviour of any cohort changing.
In their workshop, Croll and Yoskovitz ask which of two companies is doing better. The first presents its average monthly revenue per customer. The second presents a cohort table. Only the table shows whether recent customers bring in more than earlier ones at the same age.
Pitfalls
The moving denominator. A cell divided by the customers active in the previous month gives a survival rate from one month to the next. The table then changes meaning. Every cell of a row takes the cohort's starting size as its denominator.
| Denominator | Month 1 | Month 2 | Month 3 | Month 4 | Month 5 |
|---|---|---|---|---|---|
| January cohort size (400) | 60% | 45% | 38% | 34% | 31% |
| Active in the previous month | 60% | 75% | 84% | 89% | 91% |
The incomplete last period. The cells of the current month understate the activity of every cohort and trace a dip along a diagonal. The analyst stops the table at the last closed month.
The cohort that is too small. In a cohort of 30 customers, a single departure moves the cell by more than 3 points. Between two cohorts of that size, a gap of a few points can be chance. A longer period gives larger cohorts.
The changing mix. Two cohorts in the same column also differ by their acquisition channel, their entry offer and their season. A December cohort recruited through a year-end promotion starts with more volatile customers than a March cohort. A table per channel separates the effect of the channel from the effect of the product. To isolate the cause of a gap, an A/B test compares two groups that arrived at the same time.
The changing definition. A redesign of the tracking can change the meaning of "active" partway through. The columns then stop being comparable. The note in the deliverable dates each change of definition.
AI considerations
A language model writes the SQL query or the Python script that assigns each customer to a cohort and calculates the cells. The database or the script runs this calculation. The model calculates no cell. The analyst checks the result by recounting one cohort by hand: its size, then two or three cells. The analyst looks for a calendar month taken as the age and for a cell divided by the previous month's active customers.
The model also flags the columns where one cohort departs from the others and the diagonals that dip. It proposes explanations that the analyst checks in the release log and the campaign calendar. The model knows neither the releases shipped nor the incidents: an explanation without a date in the log remains a hypothesis.
The per-customer history contains personal data. It stays in the database. To write the query, the model receives the table structure without the data. For the reading, it receives the table aggregated by cohort.
Examples
A Vaud cooperative sells monthly subscriptions to vegetable boxes ordered through an app. In April 2026 it reworked the onboarding of new subscribers. The table gives the share of each cohort still subscribed at the end of each month since sign-up. It stops at 30 June 2026.
| Cohort | Sign-ups | Month 0 | Month 1 | Month 2 | Month 3 | Month 4 | Month 5 |
|---|---|---|---|---|---|---|---|
| January | 400 | 100% | 60% | 45% | 38% | 34% | 31% |
| February | 420 | 100% | 58% | 44% | 37% | 33% | |
| March | 450 | 100% | 61% | 46% | 39% | ||
| April (new onboarding) | 600 | 100% | 72% | 60% | |||
| May | 700 | 100% | 74% | ||||
| June | 800 | 100% |
At month 1, the January to March cohorts retain between 58% and 61% of their sign-ups. The April and May cohorts retain 72% and 74%. At month 2, the April cohort retains 60% against 44% to 46% for the three earlier cohorts. Over the same period, the share of all sign-ups since January still subscribed falls: 80% at the end of February, 69% at the end of March, 65% at the end of April, 63% at the end of May, 63% at the end of June. The January to March cohorts age and pull this average down. Read on its own, the average suggested that the new onboarding had failed.
Before attributing the gap to the new onboarding, the analyst checks that the acquisition channel of the April and May cohorts stayed the same.
Visualisations
The cohort table has one row per cohort and one column per age with one shade per value class. The diagram from history to table shows how the analyst assigns a customer to a cohort and then to a cell. The diagram of the three readings shows what a row, a column and a diagonal tell. The chart sets the overall average beside the retention curves by cohort and shows the trend the average hides. The comparison of cohort, segment and tests is a table. A two-row table computes the January cohort's rates with two denominators.
Cost
| Phase | Level | Rationale |
|---|---|---|
| Preparation | Medium | Set the measure, the entry event and the definition of "active" with the product manager. Obtain the extract and then check it. |
| Execution | Low | A query or a pivot table builds the table once the extract is clean. |
| Documentation | Medium | Write the note that dates the releases, the campaigns and the changes of definition. Refresh the table at each closed period. |
Tooling
A spreadsheet is enough for a few thousand customers. One column calculates each customer's month of arrival. Another calculates the age of each event. A pivot table places the cohorts in rows and the ages in columns. Conditional formatting colours the cells.
Beyond that, an SQL query groups the events by month of arrival and by age. In Python, the pandas library does the same calculation with a group-by followed by a pivot table. The business intelligence tools Power BI and Tableau display the result as a heat map and refresh it at each data load.
The product analytics tools Amplitude and Mixpanel deliver a ready-made cohort report. Their default settings fix the entry event and the definition of activity. The analyst checks these settings before reading the report.
Sources
- Alistair Croll, Ben Yoskovitz, Lean Analytics workshop, Dublin City University, April 2014, slides "Slicing and dicing data" and "Which of these two companies is doing better?": the definition of the cohort, the four ways of slicing data, the example of the two companies and its cohort table, retention placed before growth. Croll and Yoskovitz wrote the book Lean Analytics (O'Reilly, 2013). We derive the procedure for building the table from their example.
- Alistair Croll, Ben Yoskovitz, Lean Analytics, extract released before publication, 2012, note 5: the four acquisition campaigns run at the same time and split into cohorts.
- FourWeekMBA, Lean Analytics, interview with Alistair Croll, 24 June 2019: weekly release cadence as the reason cohorts matter, churn falling from 20% to 10% and then to 2% across successive releases.
- Dave McClure, Startup Metrics for Pirates, Startonomics SF, 2008, Retention slide: cohort analysis among the retention measures.

