Browse the guide
The twelve signals, and what each one needs
What every personalization signal measures, which ones work on a first-time visitor, the two context signals that read your orders rather than the shopper, and the ones that quietly do nothing until something else is in place.
A shopper group’s priorities are a weighted mix of twelve signals. You choose which ones matter and how much — that mix is the rule, and it is yours to read and change. This page is what each signal actually measures, and what it needs before it does anything.
The twelve
| Signal | What it boosts |
|---|---|
| Recently viewed | Products the visitor browsed in this session or recent ones |
| Favorite category affinity | Products whose category or attributes match the visitor’s profile |
| Cart history | Products currently in the visitor’s cart |
| Repeat purchase | Products the visitor bought before — for reorder flows |
| Top sellers | Shop-wide bestsellers, inside the collection’s window |
| New arrivals | Products released in roughly the last 30 days |
| Session intent | Products similar to what the visitor clicked in this visit |
| Return rate penalty | Lowers products with a high historical return rate |
| Price affinity | Products near the price band this visitor browses or buys in, fading as the price moves away |
| Intent match | Products whose intent tags match what the shopper is doing right now — treating themselves, buying a gift, replacing something, browsing, or close to deciding |
| Bought together | Products your customers buy together with what is in the visitor’s cart or what they just viewed, more often than chance |
| Trending | Products selling faster over the last 30 days than over the 30 before |
Most of these work on a stranger
This is the part worth internalising, because it is the opposite of how recommendation engines usually work: ten of the twelve need no purchase history and no login.
Top sellers, New arrivals, Trending and Return rate penalty are facts about your catalogue — they apply to anyone. Recently viewed, Session intent, Cart history, Price affinity and Bought together are read from the visit happening right now, so they start working within a few pageviews. Favorite category affinity builds from browsing.
Only Repeat purchase genuinely requires a shopper who has bought before.
Two signals that read your orders, not the shopper
Bought together and Trending are context signals: they say something about your store rather than about the person. Bought together counts a pair once it has been bought together in at least two orders, and lifts those partners when one half is in the cart or was just viewed. Trending needs a product to have sold at least three units in the last 30 days before its rate of change counts. Neither is a model feature — they are arithmetic over your orders — so they start working as soon as the order import has finished, with no training.
The ones that need something first
Three signals will sit in your mix scoring nothing at all unless a condition is met, and none of them announces it:
Repeat purchase only applies to products marked replenishable — consumables you would expect somebody to buy again. That is deliberate: re-surfacing a mattress to the person who just bought one is noise. But the flag is a product metafield, and it is not filled in for you. Until something sets it, this signal returns nothing for every product.
Favorite category affinity can only describe products that carry taxonomy values. On a shop where Fill product values with AI (Settings → Product data & AI) has not run, most products have no category to match on and the signal stays near zero.
Intent match needs Intent AI enabled in Settings, and the plan that carries it — Growth applies it live; Pro reads in shadow and reports. Weighted without that, it contributes nothing.
Weighting them
Give more weight to what should decide more. There is no correct set — a fashion store leaning on New arrivals and a supplement store leaning on Repeat purchase are both right, because they sell differently.
Two habits that help: start with two or three signals rather than all twelve, and change one thing at a time so you can tell what moved. See Create a personalization rule for the workflow, and How segment matching works for how a shopper is routed to a group.