The RFM bench

Top group share and revenue share. How many people the cut calls its best, and how much of the money they hold. A top group that is 16 percent of customers and 27 percent of revenue looks healthy until you read the next dial.

Bought once, inside the top group. The share of the "best" customers who have exactly one delivered order. A Champion who bought once is a first-time buyer wearing a badge, and every action aimed at Champions, early access, a loyalty tier, a personal note, is being spent on someone the shop has met once. On a base where 97 percent buy once, this dial is where the cuts separate.

Win-back audience, and its repeat share. How many people the cut would put on the win-back list, and what fraction of them have ever bought twice. A win-back message assumes a relationship to win back. Sent to a one-time buyer it is an acquisition message with a discount attached.

Yield needs a response model, and no dataset carries one. The bench states five numbers and uses nothing else: a lapsed one-time buyer reorders on their own 1.0 percent of the time and a message lifts that by 15 percent; a lapsed repeat buyer reorders 4.0 percent of the time and a message lifts that by 50 percent; a message costs R$ 0.30; one in ten of every audience is held out and never messaged; the reactivation order is worth that customer's own average order from the data. Net yield is incremental revenue minus message cost, and per-message yield divides that by the messaged count.

Change any of the five and the yields change. What does not change is their ordering across cuts, because that follows from the measured repeat share of each audience: an audience that is 100 percent repeat buyers responds on the repeat rates, an audience that is 2.4 percent repeat buyers responds almost entirely on the one-time rates. The pages that quote a yield quote it as "on the stated model", and the only thing they claim from it is the order.

Two vendor defaults call 13 and 16 percent of a base where 97 percent bought once Champions, and nine in ten of those Champions bought exactly once. The three splits that put the knife somewhere the base can actually clear produce a top group of about 1 percent, none of whom bought once. Same people. A Champions tier funded on the first kind of cut spends most of its budget welcoming first-time buyers with a loyalty programme.

In the 4,000, frequency is 3,883 ones. Ranking them into fifths gives 800 identical one-time buyers a frequency score of 1 and 683 identical one-time buyers a score of 5, in row order. Averaged with the spend score, enough of them reach the Champions corner to fill it. This is not a flaw in Putler's map or Klaviyo's rule; it is what any equal-count bin does to a column that is almost all one value, and the "fix" for the duplicate-edge error does it silently. Session 2 has the population version: 18,672 against 15,871.

Thirty-three people under the mean split, all repeat buyers. Four hundred and twenty-two under the tutorial route, one in eight a repeat buyer. Two thousand five hundred and fifty-five under one total score, one in forty a repeat buyer. On the stated model the per-message yield runs from R$ 3.46 down to R$ 0.02, and that ordering is robust because it tracks the measured repeat share. But the total column ranks the cuts differently: the median split's 50-person audience returns R$ 128 against the mean split's R$ 103, because it is larger and still all repeat buyers, and the tutorial route's 422-person audience returns the biggest total of all, R$ 203, at R$ 0.53 a message, because one in eight of them is a repeat buyer and there are many of them. Neither column is "the" answer. Per message tells you how well the cut found the right people; total tells you how many of them there were to find.

Four of Putler's segments are empty on this base: New, Promising, Can't lose and Lost. The map reaches those corners through a combined frequency-and-monetary score of 1 or 4 to 5 at a given recency, and averaging two quintile scores on a base where the frequency score is arbitrary and the spend score is skewed never quite lands there. This is not a defect in the bench and it is not a defect in the map. It is a measurement: a taxonomy written for a repeat-purchase business has four rooms that a one-time-buyer marketplace never fills, and the shop that adopts it will carry four empty segments in its dashboard forever.

The three digits are not three measurements of one thing. Recency says whether the relationship is alive, frequency says whether it is a habit, monetary says what it is worth. Adding them treats a day of recency as interchangeable with an order of frequency, and the sum forgets which was which. A 5-1-1 and a 1-5-1 both score 7; the first wants an onboarding message and the second wants a win-back, and a campaign built on the band sends both the same thing.

The bench's count is the argument: 80 of the 125 codes fall into the middle band. Any audience defined on that band is a mixture whose composition nobody chose. The eight-group taxonomy of session 1 keeps three bits and eight names; it is already the simplification, and it is the smallest one that still tells the two people apart.

A single ranking earns its place when the action is the same for everyone and only the order matters: who to call first from a list where every call is the same call. RFM's whole purpose is that the action differs by corner, so a ranking that forgets the corner has removed the thing the model was for. If a dashboard needs one number, show the code and sort by recency within it; that keeps the corner visible and still gives a list.

The honest nuance: a weighted sum with recency weighted heavily is closer to defensible, because recency dominates response in most direct-marketing tests. Session 4 covers weighting. Even then, the weights are a business judgement to be tested, not a simplification to be assumed.