What Is Cohort Analysis? Definition, Examples, and How to Run One (2026)
Cohort analysis groups people by when they joined, then tracks each group separately over time. Instead of one blended number for all users, you get one line per starting month. That separation is the whole point, because it shows whether the product is getting better or worse for people who arrive now.
This guide covers the definition, why blended averages mislead, and how to read the triangular table. It also covers the two main cohort types, and the mistakes that make a cohort report misleading.
What is cohort analysis?
A cohort is a set of people who share a starting event in a shared time window. Everyone who signed up in March is the March cohort. Everyone who made a first purchase in Q2 is the Q2 cohort.
Cohort analysis measures each of those groups on the same clock. Not "March 2026 versus April 2026" on the calendar, but "month one versus month two" from each group's own start date. That shift is what makes the comparison fair.
The output is usually a triangular table. Rows are cohorts in the order they started. Columns are age since start. The triangle shape appears because the newest cohort has only reached month one, while the oldest has data all the way across.
Why an overall average hides the problem
Here is the situation cohort analysis exists to catch.
| Cohort | Users | Month-1 retention | Month-3 retention |
|---|---|---|---|
| January | 1,000 | 45% | 28% |
| February | 1,600 | 41% | 25% |
| March | 2,400 | 36% | 21% |
| April | 3,600 | 30% | — |
Total active users grew every single month in this example. A dashboard tracking monthly actives would look excellent. Growth is covering for a product that is retaining each new group worse than the last.
The blended retention number hides it for the same reason. Each month's average is dominated by whichever cohort is largest, and the largest cohort is always the newest one. You cannot see a trend in a number that keeps changing its own composition.
Cohorts fix this by refusing to blend. January's month-1 retention stays 45% forever. When April's lands at 30%, the decline is unmistakable.
How to read a cohort table
Rows are cohorts, columns are age
Each row is a group defined by when it started. Each column is time elapsed since that start, not a calendar date. Month 0 is the starting period and is almost always 100%.
Read down a column to judge whether you are improving
This is the most useful direction and the one people skip. Fix the column at month 1, then read top to bottom. If the numbers fall as you move down, every new group is retaining worse than the one before it. That is a product or acquisition problem, and it is happening right now.
Read across a row to see how one group decays
A single row shows the shape of decay for one cohort. Most products drop steeply in the first period, then flatten. The flattening point matters more than the drop, because it approximates your long-term retained share.
The diagonal is the calendar
Everything on one diagonal happened in the same calendar month. A shock that hits one diagonal, such as an outage or a pricing change, affects all cohorts at once regardless of age. If a bad number runs along a diagonal instead of a column, the cause was an event, not a cohort quality difference.
Acquisition cohorts and behavioral cohorts
Acquisition cohorts group by when someone joined. Sign-up month is the default. These answer whether the people arriving now are better or worse than the people who arrived before.
Behavioral cohorts group by what someone did. Users who connected an integration in their first week, versus users who did not. These answer which early action predicts staying.
The two do different jobs. Acquisition cohorts diagnose direction over time. Behavioral cohorts point at what to change. Teams that only build the first kind can see decline without knowing what to do about it.
A third variation, size or plan cohorts, groups by account tier. It is useful when enterprise and self-serve customers behave nothing alike and a blended curve describes neither.
Building one by hand
The manual route works and is worth understanding once, even if you automate it later.
Get one row per user event. You need a user identifier, the signup or first-purchase date, and the dates of subsequent activity. A summary table cannot be turned into cohorts, because the grouping happens at the row level.
Add a cohort column. Truncate the signup date to a month. Every user now carries the label of the group they belong to.
Add a period-index column. For each activity row, calculate the number of months between the signup month and the activity month. This is the column position in the final table.
Pivot it. Cohort month down the side, period index across the top, distinct user count in the values. Microsoft's create a PivotTable page covers the mechanics.
Convert counts to percentages. Divide each cell by the month-0 count for that row. Absolute counts are unreadable across cohorts of different sizes.
None of these steps is hard. All five repeat every time new data arrives, and the period-index calculation breaks whenever the date format changes.
How to build one with Powerdrill Bloom
If the user data already sits in an export, the pivot assembly is skippable.
Step 1: Upload your user data
Drop in the Excel, CSV, or TSV export holding your signups and activity events. Multiple files load together, so a users file and an events file can be joined in one pass. Column detection and date parsing happen on upload, which is where the manual route usually breaks.
Step 2: Describe the cohort in natural language
Say what you want grouped and measured: "group users by signup month, then show monthly retention for the first twelve months as a percentage." The agent builds the cohort column, calculates the period index, and produces the table. Ask for a behavioral cohort instead, or a weekly grain, and it rebuilds rather than making you restructure the source.
Step 3: Export the chart, report, or deck
Take it out as a chart, fold it into a written summary, or convert the canvas into slides for the review. Professional, Business, and Fancy styles export to PowerPoint or Notion. For the visual specifically, see how to create a cohort retention chart, and the AI graph maker for other chart types from the same upload.
Common mistakes
Choosing the wrong grain. Monthly cohorts on a product people use daily hide everything interesting in the first week. Weekly or daily cohorts on a product used quarterly produce noise. Match the grain to the natural usage rhythm.
Reading rows and ignoring columns. The row tells you how one group decayed. The column tells you whether you are improving. The column is the actionable one.
Comparing incomplete cohorts. The newest cohort has not lived long enough to have a month-6 number. Leaving a partial cell in the table invites someone to read a low value as a collapse.
Cohorts too small to mean anything. A cohort of forty users produces retention percentages that swing wildly on a few individuals. Widen the window or accept that the row is directional only.
Reporting cohorts without their sizes. A percentage with no denominator invites the same misreading every time. Put the month-0 count in the first column.
Defining retention loosely. "Active" has to mean one specific thing. Logging in and completing a core action are very different definitions, and mixing them across periods makes the table meaningless.
The short version
Cohort analysis separates people by when they started so that growth cannot disguise decay. Read down the columns to see whether each new group does better or worse than the last. Read across the rows to see how a single group decays and where it flattens.
The build is five mechanical steps that repeat with every refresh. Try Powerdrill Bloom free — upload your user export, describe the cohort in natural language, and export the table straight to a deck. For a tool comparison, see AI tools for cohort analysis.
Frequently asked questions
What is cohort analysis used for?
It shows whether a product is improving for people who join now compared with people who joined earlier. Common uses are retention tracking, churn diagnosis, measuring the effect of an onboarding change, and comparing acquisition channels on quality rather than volume.
What is the difference between a cohort and a segment?
A cohort is defined by a shared starting time, and its membership never changes. A segment is defined by an attribute such as country or plan, and someone can enter or leave it. Cohorts track change over time, and segments describe a population right now.
How do I choose the cohort time period?
Match it to how often people use the product. Daily or weekly cohorts fit high-frequency products, and monthly cohorts fit subscription or purchase cycles. If most of the drop-off happens before your first column, the grain is too wide.
Can I do cohort analysis in Excel?
Yes. Add a cohort month column and a period-index column to your event rows. Pivot cohort against period index with a distinct user count, then divide each cell by its row's starting count. The calculation is straightforward, and the repetition is what makes it tedious.
What is a good retention rate?
It varies far too much by category for a single benchmark to help. Compare your own cohorts against each other instead. A curve that flattens rather than continuing to fall is the signal that matters most.