Definition
Cohort analysis segments users into groups defined by a shared start event — usually their signup or first-purchase week — and follows each group's retention, revenue, or actions across the days and weeks that follow. It separates the effect of when someone arrived from how a product or campaign is actually performing.
Where it fits
Acquisition channel → cohort by join date → retention & revenue curve → payback decision
Why it matters
Blended averages hide whether newer users are better or worse than older ones, so cohorts are how you tell real improvement from a growth illusion.
Cohort analysis is one of those techniques that quietly fixes more bad decisions than any dashboard refresh ever will. The idea is simple: instead of averaging every user into one number, you group people by when they started and watch each group move through time on its own track. A January cohort and a June cohort are different stories, and blending them together hides exactly the thing you most want to know — whether the product, the funnel, or the channel is actually getting better.
Why a blended average lies to you
Imagine your overall 30-day retention sits at 22% and has held there for months. It looks stable, even healthy. But that single figure is a weighted soup of every cohort you have ever acquired. If your oldest, most loyal users are propping up the average while every new cohort is retaining at 12%, the blended number tells you everything is fine right up until growth stalls. Cohorts pull those groups apart so you can see the trend line that the average erases.
The same trap shows up in revenue. A rising average revenue per user can come from genuinely better monetization, or it can come from a shrinking denominator as weak users churn out. Only by holding a cohort fixed and watching that group's curve can you tell improvement from survivorship.
How to build one
Pick a start event — usually signup week or first-purchase week — and a metric you want to follow, such as retention, sessions, or cumulative revenue. Lay it out as a triangle: each row is a cohort, each column is weeks-since-start. The newest cohorts only have a few populated columns because their later weeks have not happened yet, which is also the most common reading mistake (see the pitfalls below).
Most teams build this directly in their analytics tool. Product analytics platforms like Mixpanel and Amplitude ship cohort and retention reports out of the box, and on mobile an AppsFlyer cohort view can split the same curves by acquisition source so you see which channels deliver users who stick.
Reading the curves
Two shapes matter. The retention curve tells you whether users keep coming back; a curve that flattens into a stable plateau means you have a real, durable base, while one that decays toward zero means you are renting growth from paid acquisition. The revenue curve tells you when a cohort pays back its acquisition cost. Overlay spend by channel and you can estimate payback period per source, which is the bridge between cohort analysis and the unit-economics work covered in lifetime value and ARPDAU.
This is also where cohorts feed budget decisions. If your paid-search cohorts pay back in three months but your display cohorts never do, that is a clearer signal than any blended ROAS figure, because ROAS averages good and bad cohorts into one ratio. Pairing cohort payback with channel ROAS is a core move in the paid acquisition path.
Common pitfalls
- Treating young cohorts as finished. Their later columns are blank, not bad. Compare cohorts only at the same age.
- Slicing too thin. Daily cohorts for a low-volume app produce noise, not insight. Roll up to weekly or monthly until each cell has enough users to be stable.
- Forgetting seasonality. A holiday cohort may behave nothing like a quiet-month cohort, so annotate the calendar before you draw conclusions.
FAQ
How many users does a cohort need to be meaningful? Enough that a few churned users do not swing the percentage wildly — as a rough rule, aim for at least a few hundred per cohort, and widen the time bucket (weekly, then monthly) until you get there.
Retention cohorts or revenue cohorts — which first? Start with retention to confirm you have a product people return to, then layer revenue on top to find payback. Revenue cohorts on a leaky retention base just measure how fast money drains.
How does this connect to LTV? Lifetime value is essentially the area under a cohort's cumulative revenue curve. Cohort analysis is the raw material; LTV is the summary number you extract from it.
Common beginner mistakes
- Reading a single blended retention number and assuming it describes every user equally
- Cutting cohorts so small that weekly curves are just noise
- Comparing cohorts of different ages as if their later weeks were already known