Customer Retention Cohorts: Read the Repeat Business

What does customer retention cohort analysis tell you?

Customer retention cohort analysis shows whether a defined group comes back after its starting period. Fix the entry rule, return event, customer identity and time window before comparing results. Keep original group sizes visible. Separate periods still being observed from completed periods with no returns. This is business education, not financial advice. Financial decisions belong with a licensed financial adviser.

More returning customers can look like progress. It may only mean you acquired a much larger group. The operating question is sharper. Are comparable customers coming back more reliably, and what would you change if they aren't?

Which return would mean the business worked?

Choose the event that answers your question. For a repeat-order service, we would start with another completed order. A login tells you something else. A support visit tells you something else again. Do not label all three as repeat business.

Write down what completion excludes. A cancelled request should not become a completed order because the payment system briefly authorised a charge. Use the same record that operations uses to confirm delivery. Keep reversals visible under a documented rule.

Google's cohort documentation separates the condition that puts someone in a group from the condition for returning. It also warns that its exploration uses device data and does not use User-ID to create cohorts. A device-based report is therefore not automatically a customer-account report.

Before calculating anything, decide whether you are following accounts, people or devices. Our example follows customer accounts. It uses an internal account record, not a claim about any analytics product's default settings.

What belongs in the measurement agreement?

Our proposed agreement fits above the table:

• Entry: the account's first completed order in the available, verified history.
• Return: at least one later completed order in the named period.
• Unit: one customer account, counted once per cell.
• Time: calendar months under one recorded reporting timezone.
• Coverage: the records and dates the extract actually includes.
• Owner: the person who resolves disputed records and signs off the calculation.

If earlier order history is missing, say so. You may have first observed orders rather than first-ever orders. Do not silently put longstanding customers into a new-customer group.

This is a specific application of the metric contract. The retention question earns its own definition before it earns a chart.

What does a complete cohort table look like?

The following business and all its counts are fictional. They illustrate a method, not client results, benchmarks or forecasts. Assume complete records through June 30 and no unresolved identity or order corrections. Every account enters once, in its first completed-order month.

For this example, Month 1 means the next calendar month. Month 2 means the following calendar month. An account counts in each month where it completes another order, regardless of whether it ordered in the preceding month.

March's Month 1 is April. April's Month 1 is May. May's Month 1 is June. Compare down that column to compare the same stage of the customer relationship.

The returning count rises from 60 to 100 to 120. The rate falls from 60% to 50% to 40%. Acquisition has grown faster than returns. Both facts matter. A chart of returning accounts alone hides the weaker rate.

March's Month 2 rate is 50 divided by its original 100, or 50%. It is not 50 divided by the 60 who returned in Month 1. That second calculation would answer a different question and would require knowing which accounts overlap.

This denominator choice matches the structure in Amplitude's specific-cohort Return On formula. Our account records, calendar-month grouping and figures remain our own fictional example.

The cells also cannot be added to obtain unique returning customers. One account can appear in several months. Keep the account-level record behind each cell so the arithmetic can be checked.

Are you measuring a period or eventual return?

Amplitude distinguishes Return On from Return On or After. The former measures return in a specified interval. The latter includes return at that interval or later within the observed data. They answer different questions.

Our table uses individual periods. An account can skip May and order in June. It belongs in June's relevant cell even though it missed May. The table does not establish continuous activity or permanent loss.

Write the calculation type in the export name and review note. Do not compare a period-specific result with a later-return result and describe the difference as a product improvement.

The same discipline applies to time. Amplitude's time documentation distinguishes elapsed windows from calendar periods. Our example deliberately uses calendar months. Someone ordering near the end of March has less time before April starts than someone ordering near the beginning.

For a question about equal elapsed time after a first order, build that separate view. Record exact interval boundaries and the timezone. Do not rename this calendar table a 30-day retention table.

What should an unfinished period show?

April's Month 3 is July. At the June 30 cutoff, it has not happened. A zero would say nobody returned during an observed period. That would be false.

Our reporting rule is to leave that cell marked not yet observed. For a partly completed period, label it incomplete and keep it outside the completed-period comparison. Record the cutoff with every export.

Amplitude's FAQ explains its own incomplete-period markers and exclusion from overall calculations. Check the corresponding rules in whichever system you use. A blank, an asterisk and a zero can mean very different things.

A failed order feed needs a missing-data label. Fix or disclose it before assessing the group.

Is acquisition mix moving the average?

Before blaming the offer, split comparable groups by a recorded acquisition route. Keep an unknown category for accounts whose source cannot be established.

Here is a second, independent fictional example. Both groups have complete Month 1 observation. Referral accounts return at 60% in each group. Paid-channel accounts return at 30% in each group.

The combined rate falls by 18 percentage points. Neither channel's rate changes. The group contains a different mix of customers.

That does not prove the channel caused any customer's behaviour. It tells the operator where to investigate. Check the actual offer, customer need and entry conditions before changing spending or the product.

Keep raw counts beside each split. Cutting a small group into many segments can leave very little evidence in each cell. Do not award a channel a permanent verdict from a handful of accounts.

What does the table leave unanswered?

A returned account is not automatically profitable. Keep order amounts, refunds and observed service costs in a separate analysis. Do not turn this account-retention percentage into revenue retention, lifetime value or a return projection. Our marketplace economics guide explains why activity and contribution need separate records.

The table also does not explain why someone stayed away. They may not need another order yet. They may have had a poor experience. The record needs investigation, not a convenient story.

Choose the next evidence request from the decision. Review failed deliveries if the question concerns service quality. Compare offers if acquisition conditions changed. Ask about the buying cycle if the chosen interval may be too short. Keep personal records within authorised access and route privacy or contractual questions to counsel.

What decision should leave the review?

End with one testable statement. For example: "The combined rate fell while the two channel rates stayed unchanged. We will check what changed the acquisition mix before attributing the fall to delivery quality."

Name an owner, the records needed and the next review date. Preserve the original definitions so the next result is comparable. If a definition must change, show both versions and explain why.

That is the point of our working method. Make the assumption visible, gather the evidence and carry the decision through. A retention chart is useful when it changes the next investigation. The colour of the cells is secondary.

Sources

• Google Analytics: Cohort exploration
• Amplitude: Interpret your retention analysis
• Amplitude: How time works in a retention analysis
• Amplitude: Retention Analysis FAQ
• Amplitude: Retention calculations