What is cohort analysis?
Cohort analysis groups people or accounts by a shared starting condition, then follows what happens as each group ages. A signup cohort might contain accounts that started in a given week. A first-paid cohort uses subscription start instead. Choose the starting event that matches the question; changing it changes who is included and when the clock begins.
Stripe’s SaaS overview describes cohort views as a way to inspect adoption, retention, and revenue over time. A chart of total active accounts can grow because more people arrive, even while fewer of each new group return. Aligning cohorts by age helps reveal that difference without treating growth in acquisition as proof of improved retention.
This guide uses product-return retention for a fictional BriefBoard workflow: an account returns when it completes another project brief during a defined interval. It does not count logins, subscription status, or revenue as interchangeable evidence. The customer-segmentation guide groups customers by job; this exercise follows cohorts through time.
Define the cohort, return event, and clock
Write a measurement contract before filling the table. In our original synthetic workbook the unit is one account, the start is signup, and each account belongs to one signup cohort. A qualifying return is at least one completed brief in the specified window. Multiple briefs from the same account still count once within that window.
We use elapsed seven-day intervals in UTC. Week 0 is days 0 through less than 7 after signup. The displayed Week 1 is days 7 through less than 14, Week 2 is days 14 through less than 21, and Week 3 is days 21 through less than 28. These are explicit example definitions, not a claim that every analytics tool names weeks this way.
Amplitude documents both rolling time windows and strict calendar dates, as well as different return definitions. Those choices can produce different results. Record the mode, timezone, and qualifying event next to an exported table so someone can reproduce it rather than guessing from a column labeled Week 1.
Read the original SaaS retention cohort table
The workbook contains 18 invented accounts in three single-day signup cohorts. To keep the exercise inspectable, every account in each cohort starts at 00:00 UTC on its listed date. The reporting cutoff is October 27, 2026 at 00:00 UTC. No real customer records or observed company retention results are included.
The September 29 cohort has eight accounts and complete Week 1, 2, and 3 intervals. Five return in Week 1, four in Week 2, and four in Week 3. The October 6 cohort has six accounts with only the first two return intervals complete. The October 13 cohort has four accounts with only Week 1 complete.
A dash means that interval is not complete by the cutoff. It does not mean zero retained accounts. One September account returns in Weeks 1 and 3 but not Week 2, illustrating that return-in-window retention can rise again. This is different from a rule that treats every missed interval as permanent churn.
| Signup cohort | Starting accounts | Week 1: days 7–13 | Week 2: days 14–20 | Week 3: days 21–27 |
|---|---|---|---|---|
| September 29 | 8 | 5 / 8 = 62.5% | 4 / 8 = 50% | 4 / 8 = 50% |
| October 6 | 6 | 4 / 6 = 66.67% | 3 / 6 = 50% | — incomplete |
| October 13 | 4 | 3 / 4 = 75% | — incomplete | — incomplete |
Calculate retention without counting the future as zero
For each complete cell, retention is the number of original cohort accounts completing the return event in that interval divided by the original cohort size. Keep accounts that did not return in the denominator. Counting only accounts active today would remove the very departures you are investigating.
To combine the complete Week 1 cells here, add retained accounts and eligible starting accounts: (5 + 4 + 3) ÷ (8 + 6 + 4) = 12 ÷ 18 = 66.67%. The simple mean of the three row percentages is about 68.06%, which gives the four-account cohort the same weight as the eight-account cohort. Those are different summaries.
For Week 3, only September 29 has a complete interval, so the aggregate is 4 ÷ 8 = 50%. Dividing by all 18 accounts would produce 22.22% by treating ten accounts without a complete observation opportunity as failures. Amplitude’s retention API explicitly identifies incomplete intervals; retain equivalent information in your own analysis.
Choose product, customer, or revenue retention deliberately
A completed brief, a paid subscription, and retained recurring revenue answer different questions. An annual subscriber may remain paid while using the product infrequently. An account may return for a task while downgrading its plan. Name the outcome rather than placing several measures under one retention label.
Mixpanel’s retention discussion emphasizes choosing an action and a cadence that fit the product. A monthly reporting workflow should not automatically inherit a daily-return target. If a person’s task is naturally one-off, lack of a second use may not indicate a broken experience. Establish the expected repeat opportunity before interpreting an absence.
Use the churn-rate guide for opening subscription cohorts and cancellations, and the LTV guide for observed value over a completed horizon. Preserve the original acquired-account denominator when comparing cohort value at the same age.
Investigate a pattern without claiming a cause
The synthetic October 13 cohort has 75% Week 1 retention versus 62.5% for September 29. That does not demonstrate an improvement. The groups contain four and eight accounts, customer mix may differ, and a single October account moves its percentage by 25 points. There is no randomized change or real customer evidence in this dataset.
In a real review, record what changed around acquisition, eligibility, product behavior, pricing, and instrumentation. Inspect representative completed and failed tasks before proposing an explanation. A missing event can resemble lost retention even when the person finished the task through another route.
End with one bounded question, an owner, and the next evidence to collect. For example: can accounts from the same job segment complete a second brief when they next need one? That question is more actionable than deciding that every downward cell needs another reminder email.
Download the cohort data and analysis worksheet
The original JSON workbook includes all 18 synthetic accounts, signup dates, qualifying return intervals, expected results, and rules for incomplete cells. The editable worksheet walks through definitions, weighted arithmetic, data checks, and an investigation decision. Sources were checked on October 27, 2026.