Reading a metric by group of users who arrived at the same time, instead of one blended average that can hide two opposite trends.
In practice
Cohort analysis needs no new measurement: it's the same events, grouped by arrival date instead of averaged together. That regrouping alone changes what you can read, because a blended average mixes groups that sometimes move in opposite directions — two cohorts at the same thirty-day rate can end up at 21% and 6%. What you look for on a cohort curve isn't its height but its shape: a curve that eventually flattens says a share of every cohort found a lasting reason to stay, and that's the closest thing to a measurable product-market-fit signal. A curve still falling at period six describes a product renting its users rather than accumulating them.
The formula
Retention(cohort C, period n) = (members of C still active in period n) ÷ (size of C at period 0)
- Cohort
Everyone who arrived in the same window — the January sign-ups, the users from one campaign. What defines a cohort is when its members started, which is what makes two of them comparable at the same age.
- Period n
Age, not date. Period 1 for the January cohort is February and for the April cohort is May — reading them side by side means reading them at the same age, not in the same month.
- Still active
Whatever "using the product" means for you, defined once and never changed mid-analysis. Logging in is the weakest version; the action your product exists for is the useful one.
The same arithmetic works on revenue, on orders, on anything per cohort. Retention is just the version everybody draws first, because its shape is the one that answers whether the product works at all.
Worked example
Two cohorts, the same D30, opposite businesses
- January cohort: 25% still active at D30, 22% at D60, 21% at D90, 21% at D120. The curve flattens.
- April cohort: 25% at D30, 17% at D60, 11% at D90, 6% at D120. The curve slides toward zero.
- A single blended "retention rate of 25%" describes both of them exactly, and hides the only thing that matters: one product keeps a fifth of everyone it ever acquires, the other keeps nobody and is running on new sign-ups.
- The flattening level is the number you can multiply: at 1,000 sign-ups a month, the January cohort adds about 210 durable users every month, and they stack. The April cohort adds a spike that is gone by spring.
Nothing here needed a new measurement — the same events, grouped by arrival date instead of averaged, turned one meaningless number into a decision about whether to spend on acquisition at all.
Orders of magnitude
- The signal to look for is the shape, not the level: a curve that flattens at all means a group of users found a lasting reason to stay, and that is the closest thing to a measurable product-market-fit signal. A curve still falling at period six has not found it yet, whatever height it starts from.
- Levels vary so widely by category that comparing them across products is close to meaningless: a daily-use consumer app flattening at 20-30% by month three is often cited as strong, while a B2B tool bought by a company would be in trouble at that level. Compare your cohorts to each other, not to someone else's chart.
- A cohort small enough that one user is worth more than a percentage point is not a cohort, it is an anecdote with a chart. Below roughly a hundred members, read the direction and ignore the decimals.
How to improve it
- Plot by sign-up cohort before anything else. A single blended line is an average over groups that can be moving in opposite directions, and it is the default in most dashboards.
- Follow at least six periods. A flattening cannot show in three, and three periods is exactly long enough to mistake a slow slide for a plateau.
- Cut cohorts by acquisition channel too, not only by date. Two channels with the same CAC and different curves are two different decisions, and the blended view makes them look like one.
- Mark your product changes on the chart. A cohort that behaves differently is only informative if you can say what was different about the product it met.
In the Tour
Do you track a retention rate (D7/D30 or similar)?
- Yes, tracked and reviewed regularly20 pts
- We can pull it, but rarely look7 pts
- No0 pts
The Tour asks whether you track a retention rate at all — D7, D30 or similar — for 20 points, against 7 for watching it loosely and 0 for not tracking it. Cohorts are what turns that yes into something usable: a team can answer 20 with one blended figure and still be unable to say whether the product is getting better or worse, because a blended rate mixes the cohort that arrived last week with the one that arrived last year. If you track a rate but have never plotted it by cohort, you are closer to the 7-point answer than the score suggests.
Questions people ask
What is the difference between a cohort and a segment?
A cohort is defined by when its members arrived; a segment by what they are — country, plan, company size. They answer different questions and are strongest together: segmenting cohorts is how you find out that the curve only flattens for one type of customer, which is usually the most useful finding in the whole exercise.
Weekly or monthly cohorts?
Match your product's natural usage cycle, and pick the grain that gives you cohorts big enough to read. A daily tool wants weekly cohorts; something used at month-end can only be read monthly, and reading it weekly produces a sawtooth that looks like a retention problem and is a calendar.
What does a flattening curve actually mean?
That a share of each cohort stopped leaving. Those users found a reason to stay that does not wear off, and their number is what acquisition multiplies. It is the difference between a product that accumulates users and one that rents them.
How big does a cohort need to be?
Big enough that losing one member does not move the line by a visible amount — around a hundred is the usual floor. Below that, widen the window rather than chase precision: two months of sign-ups read together beat four weekly cohorts that each swing ten points on noise.