Cohort

Origin & evolution

Cohort analysis is borrowed from epidemiology, where a cohort study follows a group exposed to similar conditions over time and reads outcomes🏆OutcomeStrategyA desired business or user outcomeView reference → against that fixed longitudinal frame. The method entered business as companies looked for behavioural insight beyond aggregate metrics, because an aggregate retention number hides the thing that matters: whether the product is getting better at keeping each successive intake.

The move into software product analytics came with the subscription🔂SubscriptionSales & RevenueA recurring subscriptionView reference → and SaaS era, when retention, not a single purchase, decided whether a business compounded. Practitioners formalised the retention curve, the percentage of a cohort still active at day 1, day 7, day 30, and the comparison of curves across acquisition windows became a standard health check. Retention is calculated as active users from a cohort divided by the cohort's original size, read at each interval.

The refinement that growth teams now treat as doctrine is that a flattening retention curve, one that levels off above zero rather than decaying to it, is the real signal of product-market fit. A curve that keeps falling means the product never finds a durable habit, however good the top-of-funnel⬇️FunnelGrowthA conversion funnel tracking user progressionView reference → looks. Cohorts make that distinction legible in a way a blended retention figure never can.

How it works in practice

A meal-planning app ships a redesigned onboarding flow on 1 March. To judge it honestly, the team compares the February cohort against the March cohort rather than watching the blended number, which would mix old and new users and blur the effect.

February's cohort, 5,000 users, retains 41 percent at day 7 and 22 percent at day 30. March's cohort, 5,300 users, retains 47 percent at day 7 and 26 percent at day 30. The day-7 lift is real and it survives to day 30, so the onboarding change improved durable retention, not just first-session curiosity. Had day 7 risen while day 30 stayed flat, the read would be a novelty bump that fades, and the team would have shipped a worse product while congratulating itself.

Cohort vs. its neighbours

• Funnel measures conversion across stages at one point in time. A cohort tracks one intake of users forward through time. The funnel answers "where do people drop?"; the cohort answers "is this intake healthier than the last?".
• Behavioural segment groups users by what they do, such as power users or dormant accounts, regardless of when they joined. A cohort fixes the join window and follows the group. A segment can shift membership as behaviour changes; a cohort's membership is sealed at definition.
• Persona🎭PersonaUserAn archetype representing a user segmentView reference → is a qualitative, evidence🔎EvidenceValidationData supporting or refuting a hypothesisView reference →-based character standing in for a user type. A cohort is a quantitative, time-bounded population. A persona explains why a group behaves as it does; a cohort measures whether it stayed.

In the graph

In the Unified Product Graph, a cohort lives in the Growth domain within the Business, GTM and Growth region. It connects to the people it stands for through cohort_represents_personaCohortrepresentsPersonacross-domain, draws its definition from a cohort_defined_by_behavioral_segmentCohortdefined byBehavioral Segmentsemantic edge, and reads its numbers via cohort_measured_by_metricCohortmeasured byMetriccausal. The cohort_exposed_to_experiment_runCohortexposed toExperiment Runcross-domain edge ties a cohort to the experiment🧫ExperimentValidationA test designed to validate a hypothesisView reference → it tests, which is where retention analysis and experimentation meet. The schema also carries the region's anti-pattern guard against "growth without retention", because acquiring users who churn is expensive waste, and a cohort with no retention metric🔢MetricStrategyA unified metric that measures progress, health, or behaviour across the productView reference → leaves that waste invisible.