Cohort Analysis: How to Understand Why Customers Stay or Leave
Cohort analysis reveals the retention patterns hidden inside your aggregate numbers. Here is how to read one, build one, and act on what you find.
Your overall retention rate is 78%. Is that good? It depends entirely on which customers are staying and which are leaving — and that is precisely what your aggregate retention number cannot tell you. A company with 78% retention might be successfully retaining its most valuable customers while bleeding its newest ones, or it might be holding onto a legacy customer base while its recent cohorts churn rapidly. The aggregate hides both stories. Cohort analysis reveals them.
A cohort is simply a group of customers who started at the same time — signed up in the same month, made their first purchase in the same quarter, or upgraded in the same week. By tracking each cohort independently over time, you can see how retention patterns change across different periods of your company's history. The result is one of the most powerful diagnostics available to a product or business team.
How to read a cohort retention table
A standard cohort retention table has acquisition cohorts running down the rows (January, February, March...) and months since acquisition running across the columns (Month 0, Month 1, Month 2...). Month 0 is always 100% — that is the starting point. The numbers across each row show what percentage of that cohort was still active in each subsequent month.
| Cohort | Month 0 | Month 1 | Month 2 | Month 3 | Month 4 | Month 5 | Month 6 |
|---|---|---|---|---|---|---|---|
| January | 100% | 68% | 52% | 41% | 38% | 35% | 34% |
| February | 100% | 71% | 55% | 43% | 41% | 39% | — |
| March | 100% | 76% | 62% | 51% | 49% | — | — |
| April | 100% | 79% | 65% | 61% | — | — | — |
| May | 100% | 82% | 69% | — | — | — | — |
| June | 100% | 84% | — | — | — | — | — |
This table tells a clear story: retention has been improving significantly each month. The January cohort retained only 34% by Month 6, while more recent cohorts are holding at 69–84% through their first two months. Something changed — probably a product improvement, a better onboarding flow, or a change in the acquisition channel that brought in better-fit customers. The aggregate retention number would have masked this entirely.
The three patterns to look for
Once you have a cohort table, you are looking for three specific patterns. Each points to a different problem and requires a different fix.
A steep early drop: If Month 0 to Month 1 retention is consistently below 50%, you have an onboarding problem. Customers are signing up and not finding value quickly enough to stay. This is the most common pattern in early-stage SaaS and the most treatable — better onboarding, faster time-to-value, and more proactive early engagement all help.
Consistent decline across all cohorts: If retention keeps declining at the same rate month over month across all cohorts, you have a product value problem. Customers try the product, some stay initially, but there is not enough ongoing value to keep them engaged. No amount of onboarding improvement will fix this — the product itself needs to deliver more reasons to return.
A "smile" curve that flattens: The best retention curves flatten after an initial drop. Customers who survive the first two or three months tend to stay. The goal is to maximize the number of customers who reach that inflection point. If your curve never flattens — if retention keeps declining at the same rate at Month 6 as at Month 1 — you do not have a stable retained base, and LTV calculations built on that cohort will be misleading.
The most important number in a cohort table is not the percentage at Month 6 — it is the slope between Month 0 and Month 2. That is where most customers are lost, and where most interventions have leverage.
Segmenting cohorts: where the real insight lives
Cohort analysis becomes truly powerful when you segment cohorts by acquisition channel, pricing tier, customer size, or geography. A cohort table that shows overall retention at 60% after three months might reveal that customers from organic search retain at 75% while paid social customers retain at 40%. That single insight reframes the entire marketing allocation conversation.
The highest-value first segmentation is almost always by acquisition channel. Customers who find you through referrals, content, or organic search typically retain at materially higher rates than those acquired through broad paid campaigns. If your cohort data shows this pattern — and it usually does — it is a strong argument for investing in content and referral programs even when paid acquisition appears cheaper on a cost-per-acquisition basis. CAC looks cheap until you build retention into the model.
What to do with a declining cohort table
If your most recent cohorts are retaining worse than older ones, that is an urgent signal. It means something about your product, your acquisition, or your onboarding got worse — and it got worse recently. Check what changed. Did you open a new acquisition channel that brings less-qualified customers? Did you ship a feature that disrupted an existing workflow? Did you change pricing in a way that attracted customers who were not a good fit?
Improving cohort retention is almost always a product and onboarding problem before it is a marketing problem. The sequence of interventions that tends to work: first, identify the actions that retained customers take in their first two weeks that churned customers do not. Then, redesign onboarding to drive those actions earlier. Then, use early behaviour as a signal to identify at-risk customers and intervene before they churn, rather than after.
Long-tenured cohorts look great because the customers who stayed are, by definition, your most satisfied and best-fit customers. When you look at the Month 24 retention of a two-year-old cohort, you are seeing a population that has already been heavily filtered by churn. Do not use the behaviour of long-retained customers as a model for what new customers will do — they are a selected sample, not a representative one.
Cohort analysis is most useful when it is reviewed consistently rather than run once and forgotten. In FabricLoop, product and growth teams often maintain a monthly note in their shared group with an updated cohort snapshot — the table refreshed with the latest data, a one-paragraph interpretation of what moved and why, and a task list of interventions being tested against the at-risk cohorts. When the analysis lives alongside the tasks that respond to it, the gap between insight and action narrows substantially.