Group users by when they joined and follow each group. It is the only way to tell real improvement from a growing pile of signups.
12 min read · Level 3
Cohorts are how you tell real improvement from a growing pile of signups masking churn — and how you prove a change worked, which is otherwise nearly impossible in a company where several things change every month. The annotation is what turns the table from a description into a lesson: an improvement you cannot attribute is an improvement you cannot repeat.
A cohort is a group of users who joined in the same period. Cohort analysis follows each group separately over time, so you can see whether the product is getting better for successive groups.
Because aggregate numbers hide almost everything worth knowing. A company can look like it is growing while every single cohort is collapsing — the growth is just new signups arriving faster than old ones leave.
Cohorts are also how you prove a change worked. If the March cohort retains better than January's, something you did between them helped.
Cohort M1 M2 M3 M4 M5 Jan 100% 38% 24% 19% 18% Feb 100% 41% 27% 22% 21% Mar 100% 52% 38% 35% — Apr 100% 58% 44% — — May 100% 61% — — —
Read it two ways.
Down a column — M2 retention improved from 38% to 61% across five cohorts. Something changed for the better, consistently.
Along a row — each cohort's curve flattens. January settles around 18–19%; March around 35%. Flattening is the signal that a stable group has genuinely adopted the product.
What this table says: the product got substantially better around March, and the improvement held. If you know what shipped in March, you know what to do more of.
An aggregate retention number would have shown none of this.
First-time founders often look at a single aggregate retention figure and cannot tell whether things are improving. The whole point is comparison between cohorts.
The second mistake: reading too much into a small or recent cohort. A cohort of eleven users tells you very little, and this month's cohort has not had time to churn yet.
The third: only cohorting users, never revenue. Revenue cohorts show expansion and are how you see net revenue retention above 100%.
Weekly for fast products, monthly for slower B2B. Match it to your natural usage cycle.
What percentage of that original group is still active in each subsequent period?
Is month-2 retention getting better for later cohorts? That is your product improving.
Does the curve level out above zero? That is the signal of genuine adoption.
Mark the month you changed onboarding. Without annotations you see the improvement and cannot attribute it.
Revenue cohorts reveal expansion, and NRR above 100% is the strongest signal in SaaS.
Monthly, once you have at least a few dozen users per cohort and six weeks of history.
Not useful with very small cohorts or a very short history. Below about thirty users per cohort the noise exceeds the signal.
The most useful reading of a cohort table is vertical, not horizontal: month-two retention across successive cohorts. Rising from 38% to 61% over five months says the product genuinely got better for later users — a claim that no aggregate number and no single cohort could support.
Read down for improvement, along for adoption. Both are in the same table and founders usually read only one.
Teams that keep cohort tables without noting what changed each month can see that something improved and cannot say what. The annotation costs one line a month and is the difference between a chart and an explanation.
Write what shipped beside each cohort. Unattributed improvement cannot be repeated deliberately.
Create a free account to read the rest and apply it to your own startup.