Quick answer
A retention curve plots the share or count of a defined starting cohort that performs a meaningful return behavior at successive elapsed intervals. Acquisition cohorts group customers by a shared starting period, while behavioral cohorts compare people who did or did not perform a specified action. A valid analysis defines the unit, eligibility, start event, return event, interval, retention mode and observation window; excludes immature intervals; and keeps cohort denominators stable. Curves reveal early drop, long-term level, frequency and cohort change, but behavioral differences are associations. Use experiments or appropriate causal designs before claiming that a feature caused retention.
What are retention curves and cohorts?
A retention curve shows how many members of a starting population continue to perform a defined value behavior as lifecycle time passes. The horizontal axis is elapsed time since the cohort event; the vertical axis is a retained count or proportion. The cohort denominator remains identifiable.
A cohort is a group sharing a start period or characteristic. Acquisition cohorts align people by first experience, purchase or activation. Behavioral cohorts group people by actions, while account, plan or source cohorts can test operational questions. The grouping must precede or clearly relate to the outcome being interpreted.
Why snapshots mislead retention decisions
Monthly active users combine new, long-tenured, resurrected and soon-to-churn users. Growth in acquisition can raise the total even when each new cohort retains worse. A retention table or curve aligns customers at the same age and shows whether repeated value improves.
Cohort analysis also prevents incomplete follow-up from becoming false decline. A cohort that started last week cannot have a twelve-week outcome. Mature cells, cohort sizes and definition versions should be visible so readers understand which comparisons are supported.
How to construct a retention curve
Choose the unit, eligible start event, fixed cohort window and return event. Decide whether retention means return on a specific interval, on or after an interval, or at least once within a bracket. Each answers a different question and can create a different curve shape.
Match interval length to natural use. Daily retention is unhelpful for a monthly accounting workflow. Define timezone, identity resolution, grace periods, multiple devices and account membership. Store the original cohort membership so later profile changes do not rewrite history unintentionally.
Start
Define the qualified event and population that enter the cohort.
- When does lifecycle time begin?
- Who had a real opportunity to return?
Return
Choose a repeated behavior that represents continued customer value.
- What counts as meaningful return?
- Is the unit a user or account?
Interval
Match elapsed periods and retention mode to the natural usage cycle.
- Daily, weekly or monthly?
- On, on-or-after or within a bracket?
Compare
Read curve shape and compare mature, decision-relevant cohorts.
- Where does loss occur?
- Are newer cohorts improving at equal age?
Explain
Investigate mechanisms without confusing association with cause.
- What differs between cohorts?
- Which intervention can test the explanation?
Read curve shape and cohort tables
A steep early drop often points toward expectation, activation or first-use quality. A continued decline may indicate insufficient recurring value, replacement or poor fit. A plateau can suggest a durable core, but it may also reflect a small specialized group or the chosen retention definition.
Read a cohort table left to right to follow one cohort across age and top to bottom to compare cohorts at the same age. Examine absolute counts with percentages, because small surviving groups can produce unstable rates. Add uncertainty when decisions depend on modest differences.
How to run retention analysis
Begin with the customer's recurring job and the cadence on which value should repeat. Audit events, billing, identity and account logic. Build the simplest acquisition cohort table, inspect curve shape and verify several records manually before adding segments.
Generate explanations from interviews, support cases and product paths. Create behavioral cohorts only for specific hypotheses, check eligibility and time ordering, and test interventions where feasible. Review new cohorts after enough time has elapsed and record changes in definition or population.
- Unit of analysis correct
- Qualified start event defined
- Return event represents value
- Interval matches use cycle
- Retention mode named
- Timezone and identity governed
- Cohort denominator fixed
- Immature cells masked
- Counts shown with rates
- Reactivation handled explicitly
- Segments answer decisions
- Causal claims tested separately
Retention cohort example
KinshipCare's hypothetical analysis changes both unit and event. A household workspace captures shared value better than an individual login, and a completed care plan is more meaningful than opening the app. This definition can reveal genuine coordination patterns.
The caregiver-invitation cohort is still observational. More organized families may be both more likely to invite and more likely to retain. A targeted experiment on role setup can estimate the effect of that intervention without claiming that every invitation creates retention.
KinshipCare is a hypothetical scheduling app for families coordinating elder care. The team reports monthly active users, but an app open may reflect confusion and households begin at different points in the month.
The analysis uses a family workspace as the primary unit because coordination value is shared. A workspace enters after completing its first agreed care plan, not when one person creates an account.
Retention means the family completes or confirms another shared care plan in the relevant weekly interval. Message views and logins remain diagnostics because they can occur without successful coordination.
Weekly start cohorts are followed across equal elapsed weeks. Intervals that have not matured are masked rather than treated as zero. The team reports cohort size and the distribution of active caregivers.
A behavioral cohort compares families that activate an invited caregiver early with those that do not. The difference suggests an onboarding question but does not prove that inviting causes retention.
KinshipCare tests clearer caregiver role setup for eligible families and monitors completed plans, messaging errors, unwanted invitations, support burden and longer-term retention.
KinshipCare and all outcomes are hypothetical. Caregiving data can be sensitive and requires strong privacy, consent and access controls.
Select retention modes and companion metrics
N-period retention asks whether the return event occurs in a particular interval. Unbounded retention asks whether it occurs on or after that interval and therefore cannot fall in the same way. Bracket retention gives a flexible window. State the method on every chart.
Pair retention with frequency, depth, quality, revenue retention, reactivation and time between value events. For subscriptions, payment retention and product-use retention answer different questions. For team products, track retained accounts and active contributors without letting seat inflation masquerade as value.
Use behavioral and acquisition cohorts carefully
Segment by source, plan, use case, geography, device or behavior when a decision could change. Freeze attributes at a stated time or treat them as time-varying deliberately. Comparing current plan labels can move upgraded users between historical groups and distort interpretation.
Avoid searching many behaviors until one correlates with retention and presenting it as discovery. Reserve data for validation, adjust interpretation for multiple exploration and replicate important patterns. Qualitative evidence can explain why the same event has different meaning across segments.
Govern retention data and interventions
Publish a semantic contract for start, return, window, identity, exclusions and latency. Version material changes and backfill only with documented methods. Restrict sensitive segment access, apply data-retention rules and prevent small cohorts from exposing individual behavior.
Retention work should improve customer value, not make departure difficult. Guard against excessive notifications, obstructive cancellation and features designed only to consume attention. Monitor complaints, accessibility, support burden and unintended effects when changing recurring workflows.
Limitations and common retention mistakes
A curve records observable behavior, not satisfaction or welfare. Customers can achieve value and leave because the job is complete, or remain because switching is costly. Offline outcomes, seasonality and multi-product use may not appear in one event stream.
Common mistakes include mixing cohort ages, using login as value, changing denominators, reading immature cells, ignoring reactivation and claiming a behavior caused retention. Use curves with customer research, churn analysis, experiments and economics. The objective is repeated customer value, not retention at any cost.
Retention curves become trustworthy when the cohort, clock and return event describe the same customer-value question.
Frequently asked questions
What is a retention curve?
A plot of the share or number of a fixed starting cohort that performs a defined return behavior at successive elapsed time intervals.
What is the difference between a cohort and a segment?
A cohort is usually anchored to a shared time or behavior and followed longitudinally. A segment can be any grouped population viewed at a point or period.
What does a flattening retention curve mean?
It may indicate a stable retained core under the chosen definition. Check cohort size, duration, seasonality and whether the event truly represents value before concluding product-market fit.
Should retention be measured daily, weekly or monthly?
Use an interval that matches how often customers naturally receive value. Report the selected cadence and test sensitivity when the choice is uncertain.
Do behavioral cohorts prove what causes retention?
No. They reveal associations and generate hypotheses. Selection and common causes can explain differences, so test specific interventions causally.
Sources and further reading
- Amplitude: Complete Guide to Cohort Analysis ↗Current publisher guide to acquisition and behavioral cohorts and chart reading
- Amplitude: Retention Analysis Calculation ↗Technical definitions for return, rolling and bracket-style retention calculations
- Amplitude: Cohort Retention Analysis ↗Workflow for cohort retention with explicit correlation and causation caveats
- Mixpanel: Cohort Analysis ↗Practitioner guidance on time-bound and behavioral cohorts for retention decisions