Picture a food delivery app whose monthly active customers have climbed every month this year. The growth chart looks healthy, the leadership review goes well, and nobody asks a harder question. Then acquisition spend is cut for a quarter and active customers fall off a cliff. The problem was there all along: newer customers were leaving faster than older ones, and fresh sign-ups were covering the hole. Cohort analysis is the tool that would have shown this months earlier.
What a cohort actually is
A cohort is a group of customers who share a starting event within the same time window. The most common version is the acquisition cohort: everyone who placed their first order in January 2026 is the January cohort, everyone who started in February is the February cohort, and so on. You then follow each group forward and ask what share of them is still doing the thing you care about one month later, two months later, three months later.
The key move is that time is measured relative to each customer’s start, not on the calendar. “Month 2” for the January cohort is March; “Month 2” for the April cohort is June. Lining cohorts up by their age rather than by date lets you compare like with like: how did customers behave in their second month, regardless of when that second month happened?
Other cohort definitions are possible. A behavioural cohort groups people by something they did, such as customers who used a coupon on their first order versus those who paid full price, or users who completed onboarding versus those who skipped it. These are often more useful for diagnosis, because they point at causes rather than just dates.
Building the retention table
A cohort table has one row per cohort and one column per period since start. Each cell holds the percentage of the original cohort that was active in that period. Here is an illustrative table for the delivery app, where “active” means placed at least one order in the month.
| Cohort (first order) | Customers | Month 0 | Month 1 | Month 2 | Month 3 | Month 4 |
|---|---|---|---|---|---|---|
| January | 2,000 | 100% | 38% | 30% | 27% | 25% |
| April | 3,000 | 100% | 31% | 22% | 18% | – |
Read across a row and you see one cohort decaying over its lifetime. Read down a column and you compare cohorts at the same age. Read the diagonal and you see a single calendar month, because each step down and to the left lands on the same date.
Now the interesting part. In Month 3, the January cohort has 540 active customers (27% of 2,000). The April cohort also has 540 active customers in its Month 3 (18% of 3,000). In absolute terms the two look identical. In rate terms, April is retaining a third fewer of its customers. If you only ever looked at total active customers, the larger April intake would hide that decline completely.
The shapes worth recognising
Once you have a few months of cohorts, the curves themselves tell a story.
- A curve that flattens. Retention drops steeply in the first month or two, then levels off. The flat part is your core of habitual users. A product with a curve that flattens at 25% has found something people keep coming back for.
- A curve that keeps sliding towards zero. No stable core. Every cohort eventually disappears, which means growth depends entirely on acquisition. This is the pattern that ends badly when marketing budgets tighten.
- A curve that rises after dipping. Sometimes called a “smile”. Customers lapse and then return, perhaps seasonally (festival shopping) or because a re-engagement campaign worked. Common in retail and travel.
- Newer cohorts sitting lower than older ones. The warning sign from the table above. Possible causes include a shift towards lower-quality acquisition channels, heavier discounting that attracts deal-seekers, a product change, or entering a new city where the service is weaker.
Choices that change the answer
Cohort tables look objective, but several decisions sit underneath them, and two analysts can produce different tables from the same data.
The definition of “active”. Logged in, placed an order, or spent above a threshold? For a bank, a login might be meaningless while a transaction is not. Pick the action that represents real value to the business, and write the definition down.
Period length. Daily cohorts suit a mobile game; monthly cohorts suit a grocery app; quarterly or annual cohorts suit B2B software with annual contracts. Too short a period produces noisy cells; too long hides when people actually leave.
Classic versus rolling retention. Classic (or “N-day”) retention counts a customer in Month 3 only if they were active in Month 3. Rolling (or “unbounded”) retention counts them if they were active in Month 3 or any later month. Rolling numbers are always at least as high and keep changing as new data arrives, which confuses people when last month’s report no longer matches. State which one you used.
Incomplete periods. If today is 20 November, the November cell for any cohort covers only part of the month. Either exclude it or mark it clearly. Including it makes the latest column look like a sudden collapse.
Revenue versus customer counts. Customer retention asks how many people stayed. Revenue retention asks how much money stayed, and can exceed 100% if remaining customers spend more over time. Both are valid; they answer different questions.
From table to decision
A cohort table on its own is a description. It becomes useful when you slice it to find the cause. In the delivery example, a sensible next step is to split the April cohort by acquisition channel and by city. If customers who came through a heavy first-order discount show 20% Month-1 retention while referral customers show 40%, the conversation moves from “retention is falling” to “this specific channel is buying customers who do not stay”, which someone can act on.
Cohorts also feed directly into customer lifetime value. If you know how many customers from a cohort survive each month and roughly what they spend, you can estimate what a new customer is worth, and therefore how much it is sensible to pay to acquire one. A falling retention curve quietly lowers that ceiling.
Finally, cohorts are a natural way to judge product changes. If a new onboarding flow launched in June, compare June and July cohorts against March to May at the same ages. This is not a controlled experiment, since seasonality and marketing mix also changed, but it is a far better first look than comparing overall averages.
Where cohort analysis falls short
Small cohorts produce jumpy percentages; a cohort of 80 customers moving from 30% to 25% is four people. Recent cohorts have few columns, so conclusions about long-term behaviour must wait. And cohorts tell you that behaviour differs between groups, not why. Treat a gap between cohorts as a strong lead for investigation, then confirm the cause with segment analysis, customer research or a proper experiment.
Used well, though, a cohort chart is one of the most honest views of a business. Aggregate growth can be bought with marketing spend. Retention by cohort is much harder to fake.
Written by the RP Strativa Team. Figures and the delivery app in the examples are illustrative.
Leave a Reply