RFM Segmentation for Email: Recency, Frequency, Monetary
The scoring model, how to set the bands for your data, and the eleven segments it produces.
On this page
The short version
- Three numbers — recency, frequency, monetary value — scored into bands and combined. The model is simple; setting the bands from your own data is the part that matters.
- Eleven named segments come out of it, and only five of them usually justify their own messaging.
- It needs purchase data. Without it the model does not apply, and adapting it to engagement metrics produces something that looks like RFM and is not.
RFM is a scoring model borrowed from direct mail, where it has been in use for decades. Each customer gets three scores: how recently they bought, how often, and how much they have spent. Combine the scores and you get segments with distinct behaviour.
Its virtue is that it requires no modelling expertise and no tooling beyond a spreadsheet, and it produces groups that behave differently enough to be worth mailing differently.
Setting the bands from your own data
The standard approach scores each dimension one to five by quintile: sort all customers by recency, the most recent fifth score 5, the next fifth 4, and so on. Repeat for frequency and monetary value.
Quintiles rather than fixed thresholds is the important part. A threshold like "bought within 30 days" means something completely different for a monthly consumable than for an annual purchase, whereas quintiles adapt to your own distribution automatically.
One adjustment is usually needed. Frequency is heavily skewed on most lists — a majority have bought once — so the quintiles collapse. Where that happens, score frequency by actual count bands instead: 1, 2, 3-4, 5-9, 10+.
The eleven segments
| Segment | Roughly | What they need |
|---|---|---|
| Champions | R5 F5 M5 | Early access, referral asks, nothing pushy |
| Loyal customers | R4-5 F4-5 M3-5 | Recognition, and the next tier up |
| Potential loyalists | R4-5 F2-3 M2-3 | A reason to buy a third time |
| New customers | R5 F1 M1-2 | Onboarding, not a second offer yet |
| Promising | R4 F1 M1 | Help getting value from the first purchase |
| Need attention | R3 F3 M3 | A reactivation offer while they are still warm |
| About to sleep | R2-3 F1-2 M1-2 | A reminder of why they bought |
| At risk | R2 F3-5 M3-5 | The most valuable to save — act here first |
| Cannot lose them | R1 F4-5 M4-5 | Personal contact, not a campaign |
| Hibernating | R1-2 F1-2 M1-2 | One win-back, then sunset |
| Lost | R1 F1 M1 | Sunset — the economics do not work |
At risk is where the model earns its cost. These are people who bought often and recently stopped, and they are both the most valuable to recover and the easiest to miss in an aggregate view.
Where monetary value misleads
Monetary value is the weakest of the three for many businesses, because it correlates strongly with frequency — someone who bought ten times has spent more, and scoring both counts the same behaviour twice.
For a business with one price point it carries almost no information at all, and the model reduces to RF. That is fine; a two-dimensional version is still useful and is more honest than pretending the third dimension is doing work.
Where monetary value does matter is a wide price range, where a customer who bought once at a high value is a genuinely different proposition from one who bought five times at the lowest tier.
Recalculating, and how often
Scores are relative to the current population, so they drift as the customer base grows. A recency score of 5 in a young business may correspond to a different absolute window a year later.
Monthly recalculation is adequate for most, weekly for high-frequency businesses. Quarterly is too slow — an at-risk customer identified three months late is usually a lost one.
Keep the previous scores when you recalculate. Movement between segments is more informative than the current position: a customer moving from champion to at risk is a signal that a static snapshot cannot show.
RFM check
- Bands set by quintile on your own data, not by fixed thresholds
- Frequency scored by count bands if the distribution is skewed
- Monetary dropped if there is effectively one price point
- Five segments chosen for their own messaging, not all eleven
- At-risk and cannot-lose-them handled first
- Recalculated at least monthly
- Previous scores retained so movement is visible
Frequently asked questions
Can RFM be adapted for a list with no purchases?
Substituting engagement metrics produces something with a similar shape and different meaning, and calling it RFM invites people to apply conclusions from the purchase version. Engagement tiers do that job honestly and with less machinery.
How many customers before it is worth doing?
Enough for quintiles to be meaningful, so several hundred at minimum. Below that the segments contain single figures and the model is describing noise.
Should scores drive automation or reporting?
Both, and start with reporting. Watching how the segments move for a month before automating anything tells you whether your bands are set sensibly.