Building behavioral cohorts in Amplitude

What it is

Where a static list groups users by a fixed property (plan tier, signup date), a behavioral cohort groups them by activity over time - did an action a certain number of times, within a certain window, optionally excluding people who also did some other disqualifying action. Amplitude evaluates cohort membership on an ongoing basis, so it updates automatically as user behavior changes, rather than needing to be manually rebuilt.

Because cohorts are the unit that syncs into other tools (a Braze audience, a HubSpot list, an ad-platform retargeting segment), the precision of the cohort definition inside Amplitude is what determines whether every downstream activation actually targets the right people - a loosely-defined cohort produces loosely-targeted campaigns no matter how well the destination side is configured.

When you need it

• You need to analyze or target a group defined by real behavior - "activated but hasn't returned in two weeks," "used feature X three or more times" - that a simple property filter can't express.
• A downstream tool (a messaging platform, an ad platform, a CRM) needs an audience that stays current automatically as behavior changes, rather than a one-time list export that goes stale.
• You're comparing two groups' outcomes (a retention curve, a conversion rate) and need the comparison groups defined precisely enough that the comparison actually isolates the behavior in question.

Implementation, done properly

1. Start from the specific question the cohort needs to answer, not from the tool's filter options - "who's at risk of churning" is a hypothesis, not a definition; translate it into an explicit behavioral rule ("performed the core action fewer than N times in the last 14 days, having done it at least once before that window") before opening the cohort builder.
2. Choose the time window deliberately based on your product's actual usage rhythm - a daily-habit product and a monthly-habit product need very different windows for the same underlying concept of "still engaged," and a default or copied window from an unrelated product will misclassify a large share of real users.
3. Decide whether the cohort should be a snapshot (membership fixed at creation) or continuously re-evaluated (membership updates as behavior changes) - most activation use cases (messaging, ad targeting) need the continuously-updating version, while some analysis use cases (comparing a specific launch cohort's long-term behavior) deliberately want a fixed snapshot.
4. Use exclusion criteria as deliberately as inclusion criteria - a cohort meant to target at-risk users that doesn't exclude users who already churned entirely, or already converted through a different path, will waste downstream targeting on people the campaign can't meaningfully affect.
5. Before syncing the cohort anywhere, check its size and a sample of its actual members inside Amplitude - a cohort that's an order of magnitude larger or smaller than intuition suggests is worth investigating before it becomes the audience for a live campaign or a published analysis.

How to verify it worked

1. Pull up a handful of individual users inside the cohort in User Look-Up and confirm their actual event history matches the definition you intended - this catches definition mistakes (an OR that should have been an AND, a window boundary that's off by a day) that are invisible from the aggregate cohort size alone.
2. Check a user you're confident should NOT be in the cohort and confirm they're correctly excluded - false inclusion is often harder to spot than a size that looks obviously wrong, since an inflated cohort can still look plausible at a glance.
3. If the cohort is continuously updating, revisit it after enough time has passed for membership to plausibly change, and confirm it actually has - a cohort that never changes size over a period where real behavior clearly changed suggests a filter or refresh-cadence problem, not a genuinely stable population.
4. Where the cohort feeds a downstream sync (to an ad platform, a CRM, a messaging tool), reconcile its Amplitude-side count against the destination-side audience size as a standing check, not a one-time setup step - drift between the two over time usually signals an identity-matching issue worth investigating on its own.

Related: Sync Amplitude cohorts to Braze · Cohort analysis, defined · Use case: know which feature drives retention