RFM Segmentation

RFM comes from catalogue direct marketing, where it was used to decide who was worth the cost of a printed mailing. It survives because it is cheap to compute and hard to beat for the effort. Retail analytics tools ship it as standard, and a restaurant CRM can produce it as soon as a guest record carries visit dates and spend from the reservation book and the point of sale. Nothing about the model is hospitality-specific. What is hospitality-specific is how easily the three inputs go missing.

How does RFM work for a restaurant?

Each guest gets a score on three axes, then the scores are combined into a cell. The translation to a dining room is direct, and each axis carries one decision.

• Recency. Days since the guest’s last matched visit. The most predictive of the three in most guest bases, and the one that changes daily.
• Frequency. Visits inside a chosen window. The window is the decision that matters: a lifetime count rewards tenure and keeps a guest who stopped coming in 2023 near the top, while a rolling twelve or twenty-four months describes current behaviour.
• Monetary. Either total spend or average spend per visit, and the two produce different top buckets. Total spend favours frequent guests who spend modestly. Average per visit favours the occasional big table.

Scoring is usually quintiles, so each axis is split into five bands of roughly equal size. Quintiles keep the model honest across venues, since a score is defined relative to the venue’s own guest base rather than to a fixed dollar figure that ages.

What segments does RFM actually produce?

The useful output is four or five cells rather than the full grid, and each one implies a different action.

• Recent and frequent. The regulars. These guests need recognition on the floor and almost never need a promotion. The main risk with this cell is over-mailing it.
• High monetary, slipping recency. The most valuable win-back audience a restaurant has, because the spend is proven and the absence is recent enough to still be reversible.
• Recent, single visit. First-timers still inside the window where a second visit is realistic.
• Low on everything, long gone. Worth leaving alone, and worth suppressing from most sends. Continuing to mail this cell costs deliverability and returns very little.

Naming a cell is not the same as having a plan for it, and the common failure is generating twenty-seven cells and acting on none. Two or three cells with an owner beat a full matrix on a slide.

Why does RFM mislead in a restaurant?

RFM assumes every customer is on a comparable clock, and restaurant guests are not. A recency score of one means something entirely different for a weekly bar regular than for a couple whose only visit each year is their anniversary. Applying one recency scale across both flags the anniversary couple as at-risk every autumn and misses the bar regular for months.

The monetary axis has a hospitality-specific distortion too. A restaurant’s revenue often arrives through more than one system, so a guest who books a private event, orders catering, or stays at a group’s hotel may have most of their spend recorded somewhere the scoring never reads. Any guest scored on dining checks alone is scored on a fraction of what they are worth.

Frequency has the most damaging failure, and it is silent. A guest scored one visit may have dined eight times, with seven of those visits unmatched or attached to a duplicate profile. RFM has no way to represent uncertainty, so it puts them in a low cell with the same confidence it puts anyone else. Two habits handle this: check the venue’s match rate before reading the output, and treat low-frequency cells as provisional rather than as findings.

Can I run RFM in a spreadsheet?

Yes, and the arithmetic is genuinely trivial. Given one row per guest with a last-visit date, a visit count, and a spend total, the ranking is a few formulas and an afternoon.

The reason RFM is treated as a platform feature is that assembling those rows is the whole job. A last-visit date has to account for visits taken through more than one reservation platform plus walk-ins that only ever appeared as checks. A visit count is only correct once duplicate profiles are resolved, so one person is not sitting in the file as three. A spend total requires point-of-sale checks joined to bookings, and at many venues that join is made on table and time because the POS passes no guest identity at all. Once a platform holds a resolved profile with matched visits and spend, RFM is close to free. Before that, a spreadsheet will produce a clean-looking model of the wrong guest base.