Behavioral cohort analysis: compare actions over time

Behavioral cohort analysis compares later outcomes for customers grouped by earlier actions. Define the behavior, close its measurement window, and freeze membership before the outcome window begins. That order makes the comparison interpretable and keeps future activity out of the group definition.

Define the behavior before looking at retention

A behavior rule needs an action, a threshold, a time window, and a counting unit. “Business accounts completing at least three jobs during their first seven days” is usable; “engaged customers” leaves too much unstated. Amplitude's behavioral cohort documentation likewise defines these groups through observed actions within a specific time window.

Choose the rule before reviewing the later result. If you keep adjusting the threshold until a large retention gap appears, the report describes a rule chosen to fit those outcomes. Record the original rule and treat any revised rule as a separate comparison.

Keep a customer who later leaves in the original denominator. Dropping closed accounts would remove the very outcome the analysis needs to measure. For date-based grouping instead, use the method in building a cohort retention table from an invoice export.

Prepare one row per eligible account

Consider an illustrative US subscription software business. All accounts below started on January 1, 2026. The behavior window is January 1–7, inclusive; the outcome window is January 8–February 4, inclusive. The later window contains 28 days, and the report closes after February 4.

The business counts completed customer jobs, excluding canceled jobs and internal testing. Activity comes from a prepared account summary. A zero means the file is complete and no qualifying job occurred; missing coverage remains unresolved rather than becoming zero.

Put the following table in A1:E11. Columns B and C are the supplied counts; columns D and E are calculated. Every account has a complete behavior window and a complete outcome window.

In Excel, enter these formulas in D2 and E2, respectively, then fill down:

=IF(B2>=3,"3+ jobs","0-2 jobs")
=IF(C2>0,1,0)

These groups are mutually exclusive and exhaustive. An account completing more jobs during the outcome window keeps its original group. Membership describes what was known when the behavior window closed.

Calculate the rates and prove the totals

Subsequent activity rate = Members active during the later window / All eligible members of the frozen behavior group

Call this subsequent activity retention in the report, with the window beside the label. It means at least one completed job during the stated window; it does not mean uninterrupted daily use or a paid renewal.

Set H2 to 3+ jobs and H3 to 0-2 jobs. Put member count in I, later-active count in J, and retention in K. Enter these formulas in row 2 and fill down:

=COUNTIFS($D$2:$D$11,H2)
=COUNTIFS($D$2:$D$11,H2,$E$2:$E$11,1)
=J2/I2

Microsoft documents COUNTIFS as counting rows where all specified criteria are met. The one-account-per-row preparation is essential: counting event rows would count frequent users repeatedly.

Percentages are rounded to one decimal. The gap is 3 / 4 − 2 / 6 = 41.7 percentage points. The pooled result is 5 / 10 = 50.0%, calculated from counts rather than an unweighted average of group rates.

Prove the population first: 4 + 6 = 10 accounts. Then prove the outcome: 3 + 2 = 5 active accounts and 1 + 4 = 5 inactive accounts, with 5 + 5 = 10. Those totals must match the distinct accounts in the scoped export.

For a report with staggered entry dates, compare equal customer ages and exclude any member whose full later window is still open from both numerator and denominator. Explain that exclusion beside the rate. Reading a retention table covers the wider maturity pattern.

Use the difference as a question

The higher-activity group also has higher subsequent activity here. That warrants a review of onboarding and customer needs, with the account counts visible. It does not prove that completing extra jobs caused the difference: customers needing the product more may complete more jobs and return more often.

The operations owner should check whether the groups differ in plan, account size, or starting circumstances before changing a program. Keep additional measures separate; cohort metrics distinguishes retention from revenue and contribution. For signup-based activation, see app cohort analysis.

Where it goes wrong

• Outcome-based membership. Defining the group using jobs from the later window builds the answer into the definition.
• Moving membership. Reclassifying accounts after the freeze changes a historical comparison.
• Unequal exposure. A partial outcome window provides less opportunity to return and must remain incomplete.
• A different counting unit. Combining account denominators with individual-user activity can overstate retention.
• Causal language. An observed group difference alone cannot establish the effect of an intervention.

Covirage is not affiliated with Microsoft or Amplitude.

Compare behavior groups on your own file

Bring a sample of your own account summary with frozen behavior counts, later activity, and observation dates. Covirage maps the columns, its tools compute and check the comparisons, and the external AI model explains the results with every figure traced to its rows. Create a free Covirage account and bring the file; creating the account takes about a minute.

See self-service analytics for the supporting business reporting approach.