RFM Segmentation
Date: 2026-08-16
Score customers on recency, frequency and monetary value, then treat the resulting groups differently. It’s decades old, needs no modelling, and outperforms most machine-learning segmentations because the three inputs are the ones that actually predict behaviour.
What it is
RFM scores each customer on three dimensions:
| Measures | Predicts | |
|---|---|---|
| Recency | Days since last order | The strongest single predictor of buying again |
| Frequency | Order count in the window | Habit and commitment |
| Monetary | Total or average spend | Value if they do return |
Recency dominates. Someone who bought last week is far more likely to buy again than someone who bought twenty times but not for two years. If you only use one dimension, use that one.
Building it
Score each customer 1–5 on each dimension, using quintiles of your own base:
customer days ago orders spend R F M
Alex 12 8 £420 5 5 5 champion
Sam 240 6 £310 1 4 4 at risk
Jo 18 1 £45 5 1 1 new
Chris 410 2 £90 1 1 2 lapsed
Use quintiles of your own distribution, not fixed thresholds — a 90-day recency is excellent for furniture and terrible for consumables. Recompute monthly; scores drift as the base ages.
The segments that matter
You don’t need all 125 combinations. Five or six groups carry the value:
| Segment | Pattern | Action |
|---|---|---|
| Champions | High R, F, M | Protect. Early access, no discounting needed |
| Loyal | High F, mid R | Increase frequency — Purchase Frequency |
| New | High R, low F | The critical window. Drive the second order — Repeat Purchase Rate |
| At risk | Low R, high F, high M | The highest-value intervention. Valuable customers slipping away |
| Lapsed | Low R, low F | Winback, or accept the loss — Winback Campaigns |
| Low value | Low across | Don’t spend here |
“At risk” is where the money is. These are proven repeat buyers who’ve stopped, and reaching them before the habit breaks costs a fraction of acquiring their replacement.
Why it beats more sophisticated approaches
- It’s interpretable. Anyone can see why a customer is in a segment, which means the marketing team can act on it without a data scientist
- It needs no modelling. Three fields from the order table
- It’s stable. Recompute monthly and segments move sensibly
- The inputs are causal-ish. Recency genuinely reflects engagement rather than correlating with it
A propensity model may predict marginally better and costs enormously more to build, explain and maintain. RFM first; model later if RFM’s ceiling is genuinely the constraint.
Using it
- Differentiate the message, not just the discount. Champions don’t need money off — offering it trains a discount habit in your best customers and costs contribution for nothing — Discount Impact on Margin
- Set intervention timing from Time Between Orders, not from the recency quintile alone. “At risk” means different elapsed times in different categories
- Feed it into Lifecycle Messaging as the audience layer
- Test the interventions. RFM tells you who to talk to, not what works — and a winback discount to “at risk” customers who’d have returned anyway is pure cannibalisation. Hold out a slice — Incrementality Testing
Limits
- It’s backward-looking. It describes what someone did, not what changed. A customer whose circumstances shifted looks identical to one who drifted
- Monetary is skewed, so the M quintile is dominated by a few large spenders — Skewed and Heavy-Tailed Distributions
- It needs identity. Guest checkout across orders splits one customer into several, each scoring low — Identity Stitching
- It says nothing about why, which is where qualitative work earns its place