What is cohort analysis?
Cohort analysis groups users by something they share, usually the week or month they signed up, then tracks each group separately over time. Because every group is followed on its own clock, it separates changes in the product from changes in who you acquired, which a single average always hides.
Worked example
Three signup cohorts of a SaaS product, week 4 retention.
| June cohort, 900 signups | 27% still active at week 4 |
|---|---|
| July cohort, 1,150 signups | 29% still active at week 4 |
| August cohort, 1,400 signups, new onboarding shipped | 38% still active at week 4 |
| Blended average across all three | 32% |
The August cohort is 11 points better than June, and the blended average shows none of it The blended number is dragged down by cohorts that predate the change. Reading down the week-4 column is how you see whether the new onboarding worked, weeks before the average catches up.
Why averages hide what cohorts reveal
A site-wide retention number mixes people who joined last week with people who joined two years ago. Those groups were acquired through different channels, onboarded through different flows and used different versions of the product. Averaging them produces a number that describes nobody.
The problem gets worse as you grow. Rapid acquisition floods the denominator with brand-new users who have had no time to retain, so a company doing everything right can watch its blended retention fall while every individual cohort improves.
Cohorts fix this by refusing to mix. Each group is followed from its own day zero, so week 4 for the June cohort and week 4 for the August cohort are genuinely comparable.
How to read a cohort grid
The grid has two readings and both matter.
Along a row you see one cohort decaying over time. The question is whether the curve flattens. A curve that drops and then goes horizontal means you have found people for whom the product is genuinely durable. A curve sliding toward zero means nobody sticks, and acquisition spend will not fix it.
Down a column you see the same age across successive cohorts. This is the one that tells you whether recent work helped. If week 4 improves cohort after cohort, something you shipped is working, and you know it long before any blended metric moves.
The diagonal is the trap. Bottom-right cells contain the newest cohorts at their oldest ages, which means they are always based on tiny numbers. Ignore anything with a small denominator.
Compared
Acquisition cohorts and behavioural cohorts Two ways of forming the group. Acquisition cohorts are the default; behavioural cohorts answer harder and more useful questions.
| Feature | Acquisition cohort | Behavioural cohort |
|---|---|---|
| Grouped by | When they first arrived or signed up | An action they took, such as inviting a teammate |
| Answers | Is the product getting better over time? | Which behaviour predicts that someone stays? |
| Typical use | Tracking the effect of onboarding changes | Finding the activation moment worth designing for |
| Watch out for | Seasonality and channel mix between cohorts | Correlation read as causation: engaged users do more of everything |
The classic behavioural finding, a social network noticing that users who added a threshold number of connections in their first week almost never left, came from a behavioural cohort, not an acquisition one.
Live demo
See it on real data. The live demo is a fully populated dashboard on demo traffic, with no signup and no card.
Open the live demoGetting cohort analysis right
Match the interval to the product
Daily cohorts for consumer habit products, weekly for team tools, monthly for anything invoiced monthly. The wrong interval makes a healthy product look broken.
Define active as value delivered
Counting logins flatters every curve. Count the action people pay for: a report run, a message sent, an invoice issued.
Ignore small cells
A cohort of 30 with two survivors is not a 7% retention rate, it is noise. Set a minimum size and grey out everything below it.
Annotate the grid
Mark the week a pricing change, a big campaign or a new onboarding shipped. Without annotations you will find a step change and spend a day rediscovering what caused it.
Choose the right analysis tool
Cohort analysis needs a stable starting group and activity measured on the same clock for each group. A dedicated product analytics tool can provide these grids and curves.
mrkr focuses on traffic sources, funnels, visitor journeys and revenue attribution. It does not provide saved cohorts or signup-week retention grids. Its returning-visitor and active-day reports describe observed activity rather than cohort retention.
Anonymous cookieless identity rotates daily, so it cannot support following the same anonymous person over several weeks. Any cross-day analysis needs a suitable persistent identity and the consent that applies to it.
FAQ
What is the difference between a cohort and a segment?
A segment is a group defined by a property right now, such as all users in Germany. A cohort is defined by something that happened at a point in time, such as everyone who signed up in August, and it is then followed forward. Segments slice the present; cohorts follow the past into the present.
How many users do I need for cohort analysis?
Roughly 100 per cohort before the percentages stop swinging on individual users. Below that, widen the interval: monthly cohorts instead of weekly will usually get you there without waiting for more traffic.
Does mrkr draw a cohort retention grid?
No. mrkr does not provide saved cohorts or signup-week retention grids. Returning visits and active days are separate reports about observed activity, not a cohort retention curve.
Can I do cohort analysis with cookieless analytics?
Not across days from anonymous traffic, because the cookieless identifier is rotated daily by design. Cohorts need either an optional per-site cookie mode, which requires consent, or your own user id attached to events once someone has an account.
Start with one line of HTML.Your first visitors show up within a minute.