Cohort Analysis Basics for Growth Marketers

Averages hide almost everything interesting about growth. A monthly active user count can rise while every new group of customers churns faster than the last, simply because signups are increasing. Cohort analysis fixes this by following groups of users from a shared starting point, so you can see whether the product is genuinely getting better at keeping people.

What a cohort is

A cohort is a group of users who share a starting event within the same period. The most common choice is signup date: everyone who signed up in the first week of June forms one cohort, everyone from the second week forms another. You then measure what share of each cohort is still active, or still paying, one week later, two weeks later, and so on.

Other starting events can be useful too. You might group by first purchase, by the date someone upgraded, or by the week they first used a new feature. The question you want to answer decides the starting event.

Reading a cohort table

A retention cohort table puts cohorts in rows and time since start in columns. Each cell shows the percentage of the original cohort that was active in that period.

cohort      size   W0     W1    W2    W3    W4
2026-06-01   412   100%   46%   38%   35%   34%
2026-06-08   398   100%   49%   40%   37%
2026-06-15   455   100%   52%   43%
2026-06-22   430   100%   55%
# illustrative numbers for reading practice, not benchmarks

Reading across a row

Following one row shows how a single cohort decays over time. The sharpest drop usually happens between week zero and week one, when people who were only curious leave. What matters next is whether the curve keeps falling or levels off.

Reading down a column

Comparing cells in the same column shows whether newer cohorts retain better than older ones at the same age. In the example, week-one retention improves from cohort to cohort, which suggests that something, perhaps an onboarding change, is helping.

Reading the diagonal

Cells along a diagonal share the same calendar week. If every cohort dips on the same diagonal, the cause is likely external, such as a holiday, an outage or a seasonal pattern, rather than something about the cohorts themselves.

Retention curves and what "flattening" means

Plotting each row as a line gives you retention curves. A curve that keeps declining toward zero suggests that even engaged users eventually leave. A curve that flattens at a stable level means a core group keeps finding value. Many growth practitioners treat a flattening curve as one of the clearest early signals that a product meets a real need for a defined audience.

Choosing your definitions

• Cohort period: weekly for fast-moving products with enough volume; monthly otherwise.
• Activity definition: the meaningful action that indicates value, as above.
• Retention type: classic retention counts users active in a specific period; rolling or "unbounded" retention counts users active in that period or any later one. Classic is stricter and usually more informative for habit-forming products.
• Revenue versus users: user retention shows engagement; revenue retention shows financial health, including expansion and contraction.

Segmenting cohorts to find causes

The real insight comes from splitting cohorts. Compare cohorts by:

1. Acquisition channel: do users from organic search retain better than users from paid social? This complements attribution reporting, which only tells you who converted.
2. Activation status: do users who reached your activation milestone retain better? If not, your definition may need work.
3. Plan or company size: do teams retain better than individuals?
4. Onboarding variant: after shipping a new onboarding flow, compare cohorts before and after, or better, compare randomized groups.

Revenue cohorts for subscription businesses

For paid products, track net revenue retention by cohort: the revenue a cohort generates in a later month divided by its starting revenue. Upgrades, extra seats and usage growth push it up; downgrades and cancellations pull it down. A cohort can lose customers while revenue holds steady if the remaining customers expand, which is an important distinction for planning.

Common mistakes

Comparing incomplete cohorts

The newest cohorts have fewer data points. Do not compare a cohort's week-two number with another's week-eight number.

Small cohort sizes

A cohort of thirty users can swing wildly from one week to the next. Combine periods or wait for more data before drawing conclusions.

Mixing very different users

If one cohort includes a large batch of users from a one-off promotion, its retention may differ for reasons unrelated to the product. Note unusual events beside the table.

Building a cohort table yourself

You do not need a specialized tool to get started. With an export of users and their activity dates, a basic table takes four steps:

1. Assign each user a cohort based on the week or month of their starting event.
2. For every activity record, calculate the number of periods since that user's start.
3. Count distinct active users for each combination of cohort and period.
4. Divide each count by the cohort's starting size to get percentages.

A pivot table handles steps three and four in most spreadsheet tools. Once the logic is clear, it is straightforward to rebuild in SQL or a product analytics tool for regular reporting. Building it by hand once is worthwhile, because it forces you to make every definition explicit and reveals data quality issues, such as test accounts or duplicate users, that automated reports quietly include.

Turning cohorts into action

Use cohort analysis as a feedback loop. When you change onboarding, pricing or messaging, write a hypothesis about which cohort metric should move and by how much, then check the cohorts that follow. Over a few months, you will see which changes improve long-term retention rather than just short-term activity. Pair this with our activation metrics guide to connect early behavior with the curves you see here.