A returning customer might love the business, live nearby, or simply want to finish a stamp card. All three produce another transaction, but they do not mean the same thing or justify the same incentive.

Retention measures defined behaviors within a stated period; loyalty initiatives aim to build a valuable relationship that helps explain why those behaviors matter. Amplitude and Mixpanel show why cohort analysis needs an entry event, a return action, and explicit observation windows [R01][R02]. The underlying method applies to cafés, barbershops, B2B contracts, and SaaS; it does not require an app.

1. Distinguish repeat purchases, retention, frequency, recurrence, and churn

A repeat purchase is a later transaction by an identifiable customer. Frequency counts purchases per customer during a period. Recurrence describes their regularity. Retention requires a rule: eligible people completing a valuable action within a specified window. Churn identifies canceled or inactive relationships under a previously agreed criterion; it does not mean the same thing for a monthly subscription and a customer who orders a celebration cake twice a year.

Activation is the first action demonstrating meaningful use, not mere registration. Stated loyalty is what people say they feel or prefer, not what they actually do. Recurring revenue comes from ongoing arrangements; do not label occasional retail sales as MRR. To separate feelings from observed failures, use satisfaction and service quality measurement without confusing CSAT with retention.

2. Choose the event and time window to fit the business

Business Entry event Meaningful return Illustrative horizon, not a rule
Café with voluntarily identified customers First observed purchase Another paid purchase Next month or six weeks, based on habits
Hair salon or barbershop Completed service Another completed booking Window compatible with service frequency
Occasion-based bakery Celebration cake purchase Another relevant purchase Quarter or year, not necessarily a week
B2B supply agreement Relationship or contract start Valid renewal or repeat order Contract and buying cycle
Subscription SaaS Activation or paid start Valuable use or renewal Weeks/months for comparable plans

Do not redefine “return” after seeing results just to improve the metric. An unpaid reservation is not the same as a completed service. Opening an app is weaker evidence of value retention than accomplishing the task for which it exists. Event design is connected to integrating data from multiple business sources.

3. Build cohorts without contaminating the denominator

An acquisition cohort groups customers by their first eligible purchase. Define period-N cohort retention = cohort members purchasing in period N / original cohort members × 100. Count each person once per period even if they buy three times. Amplitude distinguishes return exactly in N from return on or after N [R01]: the numerators differ and must not be mixed.

Assume this fully hypothetical café example uses a voluntary customer register; every month below means the customer made at least one purchase in that exact calendar month:

First-purchase cohort People Return in following month Return two months later What remains unknown
January 80 February: 30 (37.5%) March: 24 (30%) Whether the 24 also bought in February
February 100 March: 42 (42%) April: 34 (34%) Whether additional purchases were anonymous
March 120 April: 54 (45%) Not yet observed This cohort's future behavior

Do not add February's new customers to January's denominator of 80. And do not interpret “30% in March” as continuous survival: some customers might have skipped February. If a cohort has not had the full chance to complete its interval, mark the cell not yet observable, not zero. Mixpanel notes that overall retention averages can conceal material cohort differences [R02].

For a subscription, if you begin a month with 80 active subscriptions and 10 cancel during it, starting-customer churn = 10/80 = 12.5%, assuming cancellations are consistently defined and new customers are excluded from that denominator. Do not apply this calculation directly to occasional café purchases.

4. Investigate why customers return or leave without inventing motives

Returning may reflect geographic convenience, price, routine, lack of alternatives, quality, a reward, or trust. No transaction alone establishes a customer's emotional state. Compare patterns with complaints, product availability, opening-hour changes, and voluntary interviews. Segment cohorts by channel and product only with adequate sample sizes; a three-person difference should not dictate a major policy.

Distinguish real churn from low buying frequency, identity errors, or an incomplete observation window. A physical retailer with anonymous sales sees receipts, not people: it cannot calculate individual return rates from transaction counts without a legitimate identification method. The store may instead track the proportion of stamp cards presented, rewards redeemed, and total sales, clearly identifying what each metric measures.

5. Choose a loyalty method that fits the business and its economics

Methods range widely in cost: consistent service and effective issue resolution, requested appointment or replenishment reminders, repeat-purchase benefits, referrals, physical stamp cards, membership, and subscriptions. Consider both customer value and the incremental cost of the benefit, not the sophistication of the technology. To design the customer relationship and offer itself, consult the customer journey and relationship guide rather than reteaching the business model here.

A physical stamp card need not store identity, but it then measures stamps and redemptions per card. A person can hold multiple cards, lose one, or share it. A subscription generates identifiers and different commercial obligations: activation, renewal, cancellations, actual service use, and margin after benefits become more relevant. High enrollment in a points club does not establish customer loyalty.

A customer who shared an email address for a receipt has not necessarily authorized promotions or marketing reminders. The UK Information Commissioner's Office (ICO) explains planning direct marketing, assessing lawful bases, and respecting expectations and opt-out rights under relevant UK rules [R03]. That is contextual methodological guidance, not automatically applicable legal advice for Peru or every jurisdiction. Verify local requirements before sending messages, linking purchases to identities, or reusing customer data for another purpose.

Minimize data: a voluntary random ID when sufficient, purchase date, permission for the stated purpose, and permitted channel. Do not buy personal profiles or reconstruct the identities behind anonymous purchases to improve a KPI. Offer a clear way to stop marketing communications. Retention measurement should serve customers, not turn them into surveillance targets.

7. Worked example: stamp cards versus a subscription

Teaching scenario, not an experiment that actually occurred. A café trials a stamp card with 60 voluntary customers, compared with a comparable group of 60 without the benefit, over four weeks. In the stamp-card group 24 return for one eligible purchase; in the other group 18 return. Observed rates are 24/60 = 40% and 18/60 = 30%, a difference of 10 percentage points, or six additional observed returns under a group-comparability assumption. The difference does not prove causation without suitable assignment and control of context.

Suppose each eligible return contributes S/40 price − S/26 variable cost = S/14 before the reward. The program gives an incremental S/5 reward to each of the 24 returning customers and requires S/45 in setup spending. Under the strong assumption that all six additional returns were caused by the program and that each person made just one eligible purchase, additional contribution would be 6 × S/14 = S/84; rewards and setup would cost 24 × S/5 + S/45 = S/165. Simplified incremental outcome: −S/81. Even an apparent improvement in repeat purchasing does not necessarily justify the promotion. Check whether customers merely shifted purchases from other days or received rewards for purchases they would have made anyway.

A separate subscription offer provides a contracted weekly purchase. It requires tracking signups, renewals, cancellations, actual usage, margin after benefits, and the obligations of the plan. Contractual retention can be calculated monthly, but it cannot be directly compared with the 40% of customers who voluntarily accepted stamp cards. Comparing different commercial mechanisms without segmenting them confuses the lesson.

8. Test an incentive with explicit limits and economic controls

Define in advance: eligibility, comparison method, horizon, return action, reward cost, communication cost, natural repeat purchasing, satisfaction, complaints, consent, and stopping conditions. A small trial may use comparable groups and a spreadsheet, but selection and calendar limitations still matter. The methods for defining experimental groups and guardrails appear in operational improvement experiments.

Do not treat “points issued” as success if customers do not return sustainably. To assess the program's economic effect—including acquisition, maintenance and discounts—use the improvement profitability evaluation. Assumptions about lifetime customer value can be explored, but they are not guaranteed future revenue.

9. Deliverable: cohort matrix and program card

The cohort sheet should record cohort ID, first purchase, eligible customers, interval N, returning buyers, denominator, return definition, rate, incomplete observations, and identification limitations. The program card documents benefit and conditions, eligible audience, redemption cost, setup, frequency, margin, test design, privacy, owner and review date. Leave room for qualitative comments rather than forcing every finding into one rate.

Editable CSVs accompany this article for cohorts, loyalty-method comparison, and a trial's economic assessment. For a small enterprise a spreadsheet and paper card may suffice; a CRM is justified only when it proportionately solves a real continuity and data-protection need.

What you should have by the end

A meaningful return definition, comparable cohorts, explicit identity limitations, and a calculation showing whether incentives produce incremental contribution, not merely more transactions or registrations.

References