Quick answer
Cohort analysis groups customers or users by a shared event, time or behaviour and compares their outcomes over time. Acquisition cohorts align people by start date; behavioural cohorts group people who performed a defined action. A retention matrix usually places cohorts in rows, age since entry in columns and the chosen metric in cells. Good analysis defines inclusion and return events, uses a natural interval, marks incomplete periods, reports denominators and treats behavioural associations as hypotheses rather than proof of causality.
What is cohort analysis?
Cohort analysis compares groups of people or accounts that share a defined starting event, time or behaviour. It follows their outcomes over time so changes in customer age do not disappear inside one blended metric.
An acquisition cohort might contain customers whose first purchase occurred in the same month. A behavioural cohort might contain users who completed a particular action during onboarding. Cohorts can also use product, market or experiment assignment when the definition supports a question.
A cohort is related to a segment but adds time or event alignment. A segment can describe customers at a moment; a cohort usually has an entry condition and a clock.
Why aggregate metrics can mislead
An overall retention or conversion rate mixes new and mature customers, different acquisition sources and changing product experiences. The average may remain stable because one group improved while another deteriorated.
Calendar reporting also confounds tenure. Customers in January may be six months old while July customers are new. Comparing them in the same monthly total does not answer whether the new-customer experience improved.
Cohorts create fairer comparisons by aligning age since a meaningful event. They also expose the period where drop-off, repeat purchase or value accumulation changes most.
Acquisition, behavioural and predictive cohorts
Acquisition cohorts group customers by when they began, such as signup, first purchase or activation. They are useful for retention, repeat contribution and the impact of calendar changes on later cohorts.
Behavioural cohorts group people by an action or sequence, such as completing setup or using a repair service. They can reveal strong associations with later outcomes, but people select into behaviours. The behaviour may be a marker of motivation rather than the cause of retention.
Predictive cohorts group people by modelled likelihood. They can support targeting or risk management, but their definition depends on model quality and may drift. Keep predicted groups distinct from observed behaviour and test fairness and intervention value.
How a cohort matrix works
A typical matrix places cohorts in rows and age intervals in columns. Each cell reports the share, count or value from that cohort meeting the return criterion at that age. The first column often represents cohort size or age zero.
Read across a row to follow one cohort as it ages. Read down a column to compare different cohorts at the same age. A heatmap can make patterns visible, but always show the underlying percentage and denominator when decisions matter.
The lower-right diagonal is incomplete because recent cohorts have not reached later ages. Blank or marked cells are correct. Treating them as zero would falsely imply churn.
Question
State the decision and outcome the cohort analysis should clarify.
- Which aggregate hides variation?
- What decision will change?
Inclusion
Define the event, date or characteristic that assigns members to a cohort.
- When does cohort age begin?
- Can a person enter more than once?
Return or value
Choose the later event or amount measured for cohort members.
- What demonstrates continuing value?
- Is the metric binary, count or currency?
Time and comparison
Select cohort granularity and compare groups at equal age with complete observation.
- What interval matches natural use?
- Which cells are not mature?
Explain and test
Investigate patterns, control obvious mix differences and test causal hypotheses.
- What changed for this cohort?
- Would the outcome change without the suspected action?
Set up a cohort analysis
Begin with one question and choose the economic or product unit. Define inclusion, return or value event, cohort granularity, age interval and observation horizon. State whether users can enter several cohorts and how reactivation is handled.
Use intervals that match expected behaviour. Daily periods may fit a high-frequency product, while monthly or quarterly ages fit durable goods or B2B renewals. Excessive granularity creates small noisy cells; broad periods can hide the moment that matters.
Validate identities, event time zones, refunds, duplicate accounts and late-arriving data. Freeze definitions or version them so a tracking change does not masquerade as a behavioural improvement.
- Decision question written
- Customer or user unit explicit
- Inclusion event and date validated
- Return or value event meaningful
- Age interval matches natural use
- Observation maturity marked
- Counts shown beside rates
- Mix differences reviewed
- Causal claims reserved for experiments
- Metric logic versioned
Cohort analysis example
The backpack analysis uses first delivery as age zero because that begins ownership. It avoids monthly-login logic and measures repair, repeat and referral outcomes on a slower durable-goods cycle.
A repairable-backpack company sees stable annual repeat revenue. It uses cohorts to learn whether the ownership and repair system is improving later customer value.
Did the new ownership guide and repair onboarding improve service use and later contribution compared with earlier customers?
Assign purchasers to the month of first delivered order. Exclude replacement shipments and link later orders to the original customer.
At age-aligned quarters, measure completed repair or part order, repeat purchase, verified referral and cumulative contribution. Keep these outcomes separate.
Compare each cohort across age and later cohorts at the same age. Mark quarters that recent cohorts have not yet reached.
If post-change cohorts improve, check acquisition and product mix, then run a randomized or phased onboarding test to estimate whether the guide caused the difference.
Cohorts organize evidence; they do not automatically create a counterfactual. Calendar changes and customer mix can still explain differences.
Patterns to look for in cohort tables
A steep early decline can signal acquisition mismatch, weak activation or instrumentation that defines activity too narrowly. A later drop can reflect renewal, product exhaustion, competitive change or a service failure at a predictable lifecycle moment.
Improving later rows at the same age may correspond to product, onboarding, pricing or channel changes. Annotate the calendar of changes and inspect whether acquisition mix or eligibility shifted at the same time.
A flattening retention curve can indicate a stable retained group in some products, but do not assume permanence. Longer observation and external changes can still alter behaviour.
Move from pattern to explanation
Drill into the cohort with customer research, journey evidence, support, quality, billing and acquisition data. Compare customer characteristics and starting conditions before attributing a difference to one product change.
Create behavioural subcohorts to explore mechanisms, but watch for reverse causality. Customers may complete a feature because they already intend to stay. Propensity adjustment can reduce observed differences but does not replace randomized evidence.
Translate the pattern into a falsifiable hypothesis and intervention. Define the expected change, eligible population, guardrails and observation period before testing.
Use cohorts beyond retention
Cohorts can track conversion, cumulative contribution, order frequency, service cost, feature adoption, referral and complaint recurrence. The method is a comparison structure, not a retention-only chart.
Marketing teams can compare acquisition cohorts on downstream value rather than first conversion. Product teams can see whether later release cohorts reach value sooner. CRM teams can compare service interventions on repeat behaviour and cost.
Use one primary metric per analysis and add diagnostic measures only when they explain it. A matrix crowded with unrelated outcomes becomes difficult to interpret.
Assess uncertainty and data maturity
Report cohort size and confidence, especially in later ages where fewer members remain. A high percentage from a tiny cohort should not drive a large decision without replication or pooling.
Account for censoring: recent customers have less time to produce outcomes. Compare complete ages or use survival methods where appropriate. Avoid selecting only mature successful cohorts after seeing the outcome.
Backfill can change historical cells. Establish a data-latency window and label whether a dashboard is provisional or final for a period.
Limitations and common mistakes
Cohort analysis is descriptive unless design supports causality. Calendar events, acquisition changes and customer mix can explain apparent improvements. A before-and-after cohort is not automatically an experiment.
Definitions can also manufacture the desired story. Choosing an easy return event inflates retention without proving value. Define events from customer and business outcomes before looking for favourable patterns.
Finally, too many cohort cuts create false discoveries. Start from decisions and hypotheses, correct for repeated exploration where needed and seek replication.
A cohort chart improves comparison by aligning age, but explanation still requires evidence about what changed and why.
Frequently asked questions
What is a cohort in analytics?
A group of users, customers or accounts sharing a defined event, start time, behaviour or characteristic that can be followed over time.
How do you read a cohort table?
Read across a row to follow one cohort with age and down a column to compare different cohorts at the same age.
What is the difference between a cohort and a segment?
A segment groups similar entities, while a cohort typically adds a shared entry event or time and an age-based observation window.
Does a behavioural cohort prove a feature causes retention?
No. People who use the feature may already differ. Use the association to form a hypothesis and test an intervention with a credible counterfactual.
Why are some cohort cells blank?
Recent cohorts have not yet reached later age intervals. Blank or marked cells represent incomplete observation, not zero retention.
Sources and further reading
- Google Analytics Help: Cohort Exploration ↗Official definitions for inclusion, return criteria, granularity and cohort tables
- Amplitude: Complete Guide to Cohort Analysis ↗Types, setup, retention matrices and common interpretation mistakes
- Amplitude Docs: Build a Retention Analysis ↗Official event-based retention analysis workflow
- Fader and Hardie: Customer-Base Valuation and Heterogeneity ↗Research context for cohort retention dynamics and customer heterogeneity