Cohort analysis groups users by a shared starting event — usually the month of sign-up — and tracks each group separately across the following periods. It answers how a group’s behaviour develops over its lifetime.
Why averages mislead
An overall figure mixes users who have been there three years with those from last week. In a growing company the new arrivals increasingly dominate that average — and because new users are active at first, overall activity looks good while retention in the older groups quietly decays. That effect is exactly what blindsides companies the moment growth slows: the average concealed the problem for as long as enough new users kept arriving.
How to read a cohort table
The usual presentation is a table: each row a starting group, each column a period afterwards. Two reading directions are informative. Horizontally shows how a single group declines over time — the retention curve. Vertically compares the same elapsed period across different starting groups: do users who started in June retain better after three months than those from January? That vertical reading is the genuinely valuable one, because it shows whether product changes worked.
The shape of the curve
What matters is not the height of the curve but its shape. A curve that flattens after some periods and stabilises at a level shows a core of users for whom the product is durably useful. A curve that keeps falling shows there is no such core — even if the starting value was high. That distinction is the single most informative signal of product-market fit.
Not only by starting month
Cohorts can be built on any shared attribute, and other cuts are often more informative than timing. By acquisition channel, to see which channel brings durable users. By plan, by company size, by the feature used first. The comparison by channel is particularly practical, because it relates acquisition cost to actual durability — a cheap channel with poor retention is more expensive than an expensive one with good retention.
Common errors
Two errors recur. Observation windows that are too short: a retention curve over two months says little, because flattening typically sets in later. And cohorts that are too small: with few users per group, the noise exceeds the effect you are trying to measure.
Practical consequence
One cohort table by starting month and one by acquisition channel suffice for most decisions. Together they answer the two questions that count: is the product getting better, and which channel brings users who stay?