RFM segmentation and cohort analysis at a glance
The RFM segmentation framework is a sorting step for the stage after the first sale. It ranks customers on recency, frequency and monetary value within a fixed window, collapses the combinations into a handful of named segments such as champions, at risk and lapsed, and maps each segment to a treatment, a channel plan and a cost limit. It is a form of SegmentationDividing a market or customer list into groups that need different offers or messages. of the customer base rather than the market, and its strength is that it makes movement visible: a customer who used to buy often and has gone quiet stops looking like a loyal one. Its limits, as the framework page sets out, are that it looks only backwards, it is blind to margin, and its bands are arbitrary.
The cohort analysis tutorial builds a different object. Each CohortA group of customers who share a starting event and period, tracked together over time. is the group of customers whose first qualifying event fell in the same period, and the table follows each group through the periods after, recording how many are still active or how much they have spent at each age. Reading down a column compares cohorts at the same age; reading along a row shows one cohort's curve; reading a diagonal shows what one calendar period did to every cohort at once. Its purpose, in the tutorial's words, is to replace the blended retention rate, which flatters whenever the base is growing, with what customers of the same age actually do.
So the unit is different. RFM gives every customer a label today. A cohort table gives every acquisition period a curve. One tells the email team whom to send what; the other tells the growth team whether the customers it is buying now are as good as the ones it bought before.
Compare table
RFM segmentation vs Cohort analysis
| Row label | RFM segmentation | Cohort analysis |
|---|---|---|
| Unit | The individual customer, given a score and a segment | The group of customers who started in the same period |
| Time axis | One point in time, refreshed on a schedule; customers move between segments | Age since acquisition along the columns, acquisition period down the rows |
| Question answered | Who is buying now, who is drifting and who has gone, so each can be treated differently | Whether customers acquired at different times, channels or offers keep buying as they age |
| Data needed | An order table with customer, date and value, over a fixed window | The same order or activity data, plus enough history for cohorts to age, and attributes recorded at acquisition |
| Output | A list of customers with named segments, sized and valued | A triangle of retention, revenue or margin per original customer, and curves plotted from it |
| Who acts on it | Email, paid media and service teams running lifecycle treatments | Growth, finance and acquisition teams judging channels, offers and lifetime value |
| How often it is refreshed | On the buying cycle, monthly for most retailers, recording each customer's previous segment | Monthly, adding a row and extending every existing row by a column |
| What it hides | Why customers behave as they do, margin, and what a customer will do next | Who, individually, needs attention now |
| How it misleads | Equal bands keep filling evenly when every customer buys less; seasonality moves people between segments | Comparing cohorts at different ages, or including an incomplete current period that looks like a collapse |
| Typical misuse | Reading the share of champions as a retention trend | Building the table once as a report, with nobody acting on it |
The rows that decide most cases are the unit and the question. If the next step is to send different messages to different customers, the work needs a label on each customer, which RFM supplies. If the next step is to judge a channel, an offer or a change to onboarding, the work needs customers of the same age compared, which only a cohort table supplies.
When to choose RFM segmentation
Choose RFM when the job is treatment: deciding which customers get recognition, which get a reminder before the habit breaks, which get a time-limited win-backMessages aimed at customers who have stopped buying or engaging, sent once they pass their usual gap between purchases. attempt, and which should be left alone or suppressed. The framework page's segment table sets out what each group usually needs and what to avoid, such as discounting champions who would have bought anyway. The email segmentation guide places purchase segments like these beside engagement tiers built from opens and clicks, and the repeat purchase and loyalty guide covers the replenishment, cross-category and loyalty programmes the segments feed.
RFM suits businesses where customers buy repeatedly at varying intervals and the base is large enough that one message for everyone wastes money. It needs little more than an order table, it can run in a spreadsheet, and it shows each customer's Purchase FrequencyAverage number of orders each customer placed in the period. and recency in a form a team can act on the same week.
It is the weaker choice when the question is about trend or cause. Because the bands usually adjust as the base changes, RFM can show a healthy spread of segments while every customer is buying less often than customers did a year ago. It is also a poor fit where purchases are rare by nature or where everyone pays on the same subscription schedule, as the framework page notes.
When to choose cohort analysis
Choose cohorts when the question is whether things are getting better or worse for customers of the same age: is month-three retention for this spring's customers higher than last spring's, did the customers a promotion brought in lapse faster, does one channel's cohort flatten higher than another's? The cohort tutorial's segment step splits the table by channel first, because a channel's acquisition cost can only be judged against the retention of the customers it brought, and by any attribute recorded at acquisition, such as first-order value or product line.
Cohorts are the input to lifetime value and payback. The lifetime value and acquisition cost guide builds LTVTotal value a customer brings over time. from margin per original customer, and the retention and payback guide reads the curve shapes and the point where cumulative margin passes acquisition cost. That makes cohort analysis the tool for decisions about acquisition spend; the wider question of how to split budget between winning and keeping customers is set out in the comparison of customer acquisition and retention.
Cohorts are the weaker choice when the team needs to know whom to contact this week. A cohort row says that a share of March's customers have lapsed; it does not say which ones, or what each should receive. And where the question is where in a journey people drop out rather than whether they stay, the comparison of cohort and funnel analysis sets out that choice.
Decision table
| Situation | Choose | Why |
|---|---|---|
| Every customer receives the same monthly email and offer | RFM segmentation | It sorts the base into groups that each deserve a different treatment |
| A promotion brought in many new customers and the team wants to know if they were worth it | Cohort analysis | Only a comparison at the same age shows whether that cohort lapses faster |
| The share of customers in the best segment is rising and leaders call it better retention | Cohort analysis before believing it | Recent customers are recent by definition; the cohort table removes that effect |
| Customers who used to buy often have gone quiet and nobody has contacted them | RFM segmentation | Recency falling against high frequency is exactly what it is built to show |
| Two acquisition channels cost the same per customer | Cohort analysis split by channel | The channel whose customers keep buying is worth more |
| A lifetime value figure is needed for budget setting | Cohort analysis on margin | Lifetime value comes from margin per original customer over time |
| A win-back offer needs a list and a hold-back group | RFM segmentation, with a hold-back slice of each segment | Treatment works on individual customers, and the hold-back shows what it added |
Using both: reading RFM segments by cohort
The two methods meet when RFM segments are read by cohort. Record each customer's acquisition period, and alongside their current segment keep the segment they held last time. Then, for each cohort, count how many customers sit in each segment at the same age. A cohort that has the usual share of new customers turning into regulars at a given age is healthy; one whose customers move straight from new to lapsed is not, whatever the overall segment sizes say. The framework page notes that movement between segments connects naturally to the cohort view, and this is the table that shows it.
Read the result in both directions. From the cohort side, it explains a weak curve: the cohort is not lapsing evenly, it is failing to produce regulars, so onboarding and the second order are the place to look, and a shop's Repeat Purchase RateShare of customers who placed more than one order in the period. for that cohort will show the same gap. From the RFM side, it shows that a segment is not one population: the customers who bought once and went quiet may be mostly one promotion's customers, whose drift is a fact about the offer rather than about service. That changes the treatment, since a reminder may bring back a customer who used to buy at full price and do nothing for one who only ever bought at a discount.
Keep the ownership clear. The framework page owns the scoring and naming; the cohort tutorial owns the table; the email segmentation guide owns how segments reach the list. The retention section brings the store-side pieces together. Use the cohort reading to set targets and judge changes, and RFM to carry those changes to individual customers.
A bakeware shop sorts its customers and then checks its newer cohorts
Fogliump Bakeware, an invented online shop selling tins, proving baskets and cake decorating kits, has 9,200 customers with an order in the last 18 months. It scores them on recency, frequency and spend in five bands each and collapses the grid into six segments. Champions are 9% of customers, at risk 14% and lapsed 33%. The shop stops sending its monthly discount to champions, sends at-risk customers new recipes for the tins they bought, and gives lapsed customers a three-email win-back sequence before suppressing non-openers. A held-back tenth of the at-risk segment orders at 11% over the next two months against 18% for those who received the recipes.
The segment sizes barely move over the following quarter, so the team builds monthly cohorts on first paid order. The March cohort, 2,400 customers won by a launch promotion at 40% off, is nearly three times the size of its neighbours. At age three, 14% of March's customers have ordered again, against 29% to 31% for the January, February, April and May cohorts at the same age. Reading segments by cohort, 3% of March's customers have become loyal regulars by age three, against 9% to 10% for the other spring cohorts, and March customers make up 58% of everyone who bought once and has not returned. The team excludes March from its lifetime value projection, changes the launch offer for the next range to free delivery on a second order, and checks that cohort's age-three cell before calling the change a success.
Common mistakes
- Reading RFM segment sizes as a retention trend. Equal bands and new customers keep the segments looking stable; compare cohorts at the same age instead.
- Building a cohort table and never acting on it. The table exists to move budget, onboarding and lifetime value inputs.
- Treating a cohort as a target list. A cohort says what share lapsed, not who needs which message.
- Scoring RFM on revenue without refunds. Heavy returners can top the monetary ranking, and the same error inflates cohort revenue.
- Changing the RFM window or the cohort definition between runs. Customers then move because the rules changed.
- Ignoring one promotion's cohort. A discount-heavy cohort can fill a whole RFM segment and distort both views.
- Judging RFM treatments without a hold-back. Without one, a segment's normal behaviour is credited to the campaign.
The comparison library lists the other pairs that are easy to confuse.
