Expert analysis · reviewed 6 Oct 2026
RFM segmentation: prioritising customers without mistaking a good score for a good target.
Key distinctions
Three conditions before calculating.
A relative score, not a value
RFM ranks customers against one another, by quantiles of a given base and period. A score of 5 cannot be compared from one base to another and measures no margin.
The past, not the response to action
RFM describes past behaviour. It does not say who will change behaviour thanks to an incentive: the best customers often buy without one.
Monetary value as margin
An amount measured in revenue favours customers who buy on promotion or return a lot. Margin after discounts and returns gives the true ranking.
Method
Building an actionable RFM segmentation in four steps
- 01Set the window and the unit
Choose the observation period, the customer identifier and the reference date; a window that is too short flattens frequency, while one that is too long mixes lapsed and active customers.
- 02Score by quantiles
Split recency, frequency and monetary value into quintiles on the same base, documenting how ties and single-purchase customers are handled.
- 03Group into decision segments
Merge the 125 combinations into a few segments, each linked to a possible action; a segment without an action is a statistic, not a decision.
- 04Test the action against a control
Hold out a random control group in each targeted segment, then measure the incremental margin net of the cost of the action.
The RFM matrix, from segment to testable action
Four typical segments, their reading and the action to test. Each action remains a hypothesis until a control group has measured its effect.
| Segment | Recency (R) | Frequency (F) | Monetary (M) | Reading | Action to test |
|---|---|---|---|---|---|
| Champions | 5 | 4–5 | 4–5 | Already buy without an incentive | Non-monetary benefit, control group mandatory |
| Declining loyals | 2–3 | 4–5 | 3–5 | High past value, slowing activity | Targeted reactivation, capped cost |
| New customers | 5 | 1 | 1–3 | Future value still unknown | Path to the second purchase |
| Lost customers | 1 | 1–2 | 1–2 | Low probability of return | No spend, observation only |
The thresholds are the quintiles of the base studied: a score of 5 cannot be compared from one base to another.
The best RFM score is not the best target
Illustrative example: a €10 voucher is offered to 2,000 champions and to 1,200 declining loyal customers, with a margin of €60 per additional purchase and a control group in each segment.
Champions: cost = 2,000 × 40% × €10 = €8,000; incremental margin = 2,000 × 3% × €60 = €3,600. Declining loyals: cost = 1,200 × 15% × €10 = €1,800; incremental margin = 1,200 × 8% × €60 = €5,760.
The best-scored segment loses €4,400: most of its customers would have bought without the voucher. The declining segment gains €3,960, because the offer genuinely changes behaviour there.
Acceptance conditions
What must be true to act.
- 01
Calculate monetary value on margin after discounts and returns, never on revenue alone.
- 02
Document the window, the reference date and the quantile thresholds of each score.
- 03
Hold out a random control group in every segment that receives an incentive.
- 04
Keep a segment only if its incremental margin exceeds the cost of the action.
Limits
What this analysis does not prove.
- RFM ignores future margin, the product life cycle and purchases made outside the observed base.
- Quantiles depend on the base: a change of scope shifts every score.
- Single-purchase customers concentrate in the low frequency scores, whatever their potential value.
- An RFM segment does not predict the response to an offer; only an experiment measures it.
References