From segment to action, and proving it

Start from what the code says is missing. 111 High value is missing nothing, so the action is to protect: early access, a human touch at the moments that matter, and never a discount they did not need, because a discount to someone who was going to buy anyway is margin given away. 101 Key develop is recent and big but only once, so the whole job is the second order: a category cross-sell, a reason to return before the memory of the first order fades. 011 Key win-back is frequent and big and gone quiet, so win back with a relevant new arrival rather than a blanket coupon, and ask why they left, because the answer is worth more than the order.

001 Key potential, one big order long ago, gets one cheap reactivation attempt and then release; the corner is 12.1 percent of Olist customers and 26.6 percent of revenue, which is exactly why a blanket effort is tempting and exactly why it must be capped. 110 General potential, recent and frequent with small baskets, wants the basket raised at the checkout with bundles and a free-shipping threshold, not a message. 100 General develop, the biggest corner on Olist at 37.9 percent, is onboarding to the second order at low unit cost. 010 General retain gets content and no spend. 000 Low value is deprioritised, and "deprioritised" means the cap is zero, written down, so nobody adds them to a blast later.

Then extend every row with two more columns: the channel and the frequency cap. The channel follows the action (a human touch is a call or a handwritten note, onboarding is a short series, basket-raising is on-site). The cap is the promise you make to the customer's attention, and it is the column that turns an action table into an operating rule, because without it every corner drifts toward the email tool.

The eleven-group map that most tutorials copy comes with its own action column, and it is worth reading beside the eight-corner table above because it says the same things at a finer grain. Quoted from putler.com/rfm-analysis, in short form:

Notice what is missing from both tables: a channel and a cap. Vendor action columns tell you what to say and never how often, because the vendor is paid per message. Add the two columns yourself.

Klaviyo's published rule (help centre article 17797937793179): recency scores 3 for a last order within 180 days, 2 within 365 days, otherwise 1; frequency and monetary are split into thirds; six named groups are assembled from listed sets of three-digit codes. Putler's rule (putler.com/rfm-analysis): quintile scores 1 to 5 on each of R, F and M, frequency and monetary averaged, and eleven segments defined by published score ranges.

Run both on Olist's 93,358 customers and the top groups are large. Klaviyo's map calls 13.9 percent of customers Champions, and 91.0 percent of those Champions bought exactly once. Putler's calls 16.1 percent Champions, of whom 92.1 percent bought once. Neither rule is broken. On a base where 97 percent of customers bought once, a frequency column split into thirds or fifths has to put most of the identical one-time buyers somewhere, and the pandas rank(method="first") convention that every tutorial uses to get past the duplicate-edge error assigns them by row order. One replication note applies to both: neither product documents its tie handling for a column that is 97 percent ties, so these are properties of that convention on this base, not claims about either product's live output.

The professional norm that follows: before funding anything aimed at a tool's top group, compute the share of that group who bought exactly once on your own data. If it is nine in ten, the group is not "your best customers", it is "recent single big baskets with a row-order score", and the right action is Key develop's second-order push rather than High value's protect. The name on the group did not tell you that. The cut did.

Put the two vendor maps side by side and one structural agreement stands out: the recency score does most of the sorting. Klaviyo scores recency on calendar thresholds (180 and 365 days) and only then splits the other two columns into thirds, so a customer's recency band decides which third of the map she can land in at all. Putler averages F and M into a single second axis and reads the map as recency against that average, so recency is one of only two axes and the one not diluted by averaging. On Olist the effect is visible in the population shares: Klaviyo's "Recent" group is 21.6 percent, Putler's "Loyal" 31.6 percent and "Potential loyalist" 24.3 percent, all bands of recent buyers with modest F and M.

That is a defensible design choice for a message-sending tool, because recency is the one column that predicts whether a message will be opened. It is a poor design choice for deciding who is worth a discount, because a recent one-time buyer and a recent ten-time buyer land in the same band under both maps. The eight-corner taxonomy keeps frequency as its own digit for exactly this reason, and it is why the 101 and 111 corners get different actions even though most tools would put them in one group.

The procedure has four lines. One: take the audience the segment defines and mark a random tenth held out, with a seed, before anyone sees the list. Two: message the other nine tenths and never message the tenth, for the whole window. Three: after the window (say 90 days from the send), compute the reorder rate in each group. Four: the lift is the treated rate minus the held-out rate, and incremental orders are the lift times the treated count.

What the held-out rate exposes is the organic reorder: the share of this audience that buys again with no message at all. Every one of those, in the treated group, will show up in the platform's "attributed revenue" because the platform attributes any purchase after a send to the send. Revenue that would have happened anyway is the organic rate times the treated count, and it is not the campaign's. Vendor content occasionally concedes the point; one vendor-side blog puts the organic share of win-back revenue attributed under Klaviyo's methodology at 15 to 22 percent. That is a vendor or press claim, not checked at source here, and the honest reading is not the number but the admission: the attribution model itself says some of what it attributes was not caused.

The same sheet must carry contact cost and the frequency cap from Part 1. Net incremental is incremental revenue minus the cost of every message sent, including the ones to people who would have bought anyway, and the cap is what stops the cost line from being paid twice a week. A lift that does not clear the cost is a campaign that lost money while reporting a positive attributed number, and that is a common outcome, not a rare one.

Ten percent is a habit inherited from the direct-mail test cell, and it is a reasonable habit on a large audience. It is the wrong habit on a small one. On the session 3 bench, the mean-split win-back corner is 33 people; a tenth of that is 3, and three people cannot show a difference in reorder rate against thirty. The right question is: what is the smallest held-out group that could still show the lift I am hoping for, at the reorder rates I expect? That is a sample-size calculation, and it depends on the organic rate, the lift you would act on, and how sure you need to be.

Two consequences for RFM practice. First, the holdout share should be decided per audience, and for a small corner it may need to be a third or a half, or the corner may need to be pooled with a neighbour before it can be tested at all. Second, an audience too small to test is an audience too small to campaign on with any confidence about the result; the honest move is to act on it anyway if the action is cheap, and to say plainly that the outcome will be an anecdote. The maths of the calculation belongs to Experimentation, which owns power, minimum detectable effect and the sequencing of tests.

From the bench in session 3, on the stated response model (lapsed one-time buyers reorder on their own 1.0 percent of the time and a message lifts that by 15 percent; lapsed repeat buyers reorder 4.0 percent of the time and a message lifts that by 50 percent; a message costs R$ 0.30; a tenth of every audience is held out):

Two columns, two orderings. Per message, the mean split wins because every one of its 33 people is a proven repeat buyer. In total, the median split wins because it has 50 such people rather than 33, and the extra 17 are worth more than the small drop in yield each. Neither column is "the" answer. The per-message column is what a finance partner reads (what does each contact return?), the total column is what a growth lead reads (what does this campaign return?), and a spec that shows only one of them is hiding a decision.

What transfers and what does not. The reais are on a model and will not be your numbers. The ordering is not on the model: it follows from the measured repeat-buyer share of each audience, 100 percent under the two-way splits against 2.4 and 2.5 percent under the quintile and sum cuts, and that share is a fact about who the knife put in the group. Any response model in which repeat buyers reorder more readily than one-time buyers will reproduce the ordering. That is why "message everyone in win-back" costs what it costs: under a quintile or sum cut, "everyone" is 767 or 2,555 people of whom nineteen in twenty have never bought twice.