Beyond purchase

On a platform like this the letter that is empty is M. Most users have never spent anything, so a monetary column is zero for nearly everyone and the knife, wherever you put it, sorts almost the whole base onto one side. The tempting repair is to keep M and bolt an engagement axis on beside it. That gives you four letters, of which one is still empty, and sixteen corners of which eight hold almost nobody.

The cleaner repair is RFE: keep R and F, and put E where M was. R becomes days since the last action of any kind. F becomes the count of actions, or of sessions if the events carry a session id. E is an engagement score: a weighted count of what the person did, where a purchase is worth more than a page view because it costs the person more. One common choice of weights is pv 1, fav 2, cart 3, buy 5. Those numbers are a decision, not a measurement. Nobody has shown that a cart is worth exactly three page views; someone decided that ordering felt right and wrote it down. That is fine, as long as it is written down and can be argued with.

The test for "replace" is simple: if the letter you are about to add would be empty for most of the base, you are not adding, you are replacing, and you should say so. Part 3 shows the opposite case, where the old letters are all still full and a genuinely new behaviour earns a fourth column.

Replacing M with E does not mean forgetting purchase. It means purchase recency and action recency become two different columns, and they answer two different questions. R_action, days since anything happened, tells you whether the person is still around. R_buy, days since the last purchase and missing for anyone who never bought, tells you whether they have ever crossed the line and how long ago. A user with R_action 1 and R_buy missing is alive and has never converted: an onboarding problem. A user with R_action 40 and R_buy 41 bought once and left the same day: a different problem entirely.

Keep both. Segment on R_action, because it is defined for everyone, and carry R_buy as a column you read when you choose the action. The build-along computes both, and the missing values in the second are the point, not a data-quality problem to fill.

A brand room is not a shop with a video on it. It runs at a fixed hour, the same faces come back, and the thing a regular does most is watch. So the letter that breaks is not M this time, it is F. Counting orders makes the nightly regular who buys twice a year look like a two-time buyer, indistinguishable from someone who found the room through a shelf search and never came back. The repair is not to replace F but to redefine it: frequency becomes sessions attended in the room plus orders placed, so attendance counts as the loyalty it is.

Recency splits into two clocks. R_watch is days since the last session attended, defined for everyone who has ever been in the room. R_buy is days since the last order, and missing for anyone who never bought. Monetary stays what it always was, the person's GMV in the room, because on a live platform it is the one letter that still means what it says.

The segment this exposes is the one a 店播 (brand rooms) operator cares about most: high watch frequency, zero spend. In a purchase RFM that person does not have a low score, they have no row. Under the redefined F they are a named group, and session 5 gives them an action and a holdout.

The white paper's split between 店播 (brand rooms) and 达播 (influencer rooms) is not just a market-share table, it is a question about who owns the RFM row. In a 店播 (brand rooms) room the brand runs the broadcast, the regular's loyalty is to the brand, and every session she attends is a data point the brand may keep and act on. In a 达播 (influencer rooms) room the loyalty is to the host. The brand buys a slot, the sale is real, but the room regular is the host's customer, and a "frequency" built on attendance in a host's room is measuring a relationship the brand does not have. The RFM row belongs to the room owner, and only in 店播 (brand rooms) is that the brand.

This is why a brand that sells through both channels should build the redefined RFM on its own room only, and treat 达播 (influencer rooms) orders as acquisition events feeding the purchase side, not as frequency. Mixing them inflates F with sessions the brand never hosted.

On Pinduoduo every buyer has an order, so R, F and M are all defined and all mean what they say. What the order table misses is a fourth thing the platform is built on: the invitation. A customer who buys a small basket but forms three teams a month has brought the platform more new buyers than a customer with a large basket and no shares, and in a three-letter RFM the two are indistinguishable or, worse, the sharer scores lower.

So the move is RFM-I: add I, a count of invitations sent or teams initiated in the window. Now the customer with low M and high I has a name, acquisition channel, and an action that is nothing like the action for a low-M non-sharer. The distinction between this and Part 1 is the whole session: there, the letter you would add is empty for almost everyone, so you replace. Here, the letters you already have are full and the new one is too, so you add. Confusing the two either buries a real behaviour inside a score or multiplies your corners by two for nothing.

On disclosure: PDD Holdings' 20-F describes the mechanism in the words quoted above but no longer discloses an active-buyer count. The 882 million figure that circulates for 2023 is a secondary number and is not used here.

A fourth binary axis doubles the corners to sixteen, and sixteen groups is more than any team writes distinct actions for. The alternative that survives in practice is to keep the eight-corner code from session 1 exactly as it was and carry I as a flag beside it: a "sharer" column set when I clears its own knife. Every one of the eight groups then has a sharer and a non-sharer variant, and the operator chooses which groups the flag changes the action for (usually the low-M ones, where the referral is the value) and ignores it elsewhere.

The same logic applies to any new axis. A second dimension that changes the action in two groups and nothing in six is a flag, not a corner. Multiply corners only when the new letter changes the action nearly everywhere, and be honest that it rarely does.

There is no measured answer to how much recency should weigh against spend, and the literature does not pretend otherwise once you read it rather than quote it. Hughes weighted nothing. Stone is cited for weights that could not be verified at source here. Vendors propose weights that suit the product they sell. What all three have in common is that the weights encode a belief about which behaviour predicts the next purchase, and a belief about your customers is something you can test.

So the working rule: write the weights down as a dictionary with a one-line reason each, apply them, and hold out a slice of every segment when you act on it. Session 5 builds that holdout. If the recency-heavy weighting produces a segment that responds and the equal weighting does not, you have learned something about your base that no citation could have told you. If both respond the same, the weights did not matter and the simpler one wins.