Merchandising rules vs AI picks: who chooses the hero product
What rules are good at
Rules are good at intent. A merchandiser who pins a new launch to the top of the collection, boosts a high-margin product during a sale, or hides a product that is about to be discontinued is expressing the business's priorities, and no model knows those priorities unless someone tells it. The rule is the business talking, in a language the storefront understands.
Rules are also good at the edge cases the model gets wrong. A product that is new has no purchase history, so a model that ranks by past sales will bury it. A product that is seasonal has a history that is wrong for this month. A product that is being cleared needs to be seen, and the model, which optimises for conversion, will hide it, because clearance products convert worse than full-price ones. The rule corrects for all of that, because the rule knows what the model does not.
And rules are auditable. When the hero product is wrong, the merchandiser can say which rule put it there, and change the rule. When a model puts the wrong product first, the explanation is a score, and the score is not a reason. A business that cannot explain why a product is first cannot fix it when it is wrong.
What the model is good at
The model is good at the shopper. A rule that shows the same hero product to everyone shows the wrong product to most people. A model that ranks by what this shopper looked at, what people with a similar basket bought, and what fits the cart they already have, shows a different hero product to each person, and the difference is the conversion.
The model is also good at scale. A catalog with ten thousand products cannot be merchandised by hand. The rules cover the top hundred, and the model covers the rest, ranking the long tail for each shopper without a human touching it. A store that relies on rules alone leaves the long tail unmerchandised, which is most of the catalog.
And the model is good at the pattern no human sees. A product that sells with another product, a combination that converts at twice the rate of either product alone, a segment that buys a product the merchandiser would never have pinned: the model finds these because it counts, and the merchandiser misses them because they cannot count ten thousand combinations by hand.
How to run both, without letting either run the store
Let the business set the candidates and the model rank them. The merchandiser defines the pool: the launch, the high-margin products, the seasonal push, the clearance that needs to move. The model ranks that pool for each shopper. The business decides what is eligible. The model decides what is first. Neither decides alone.
Measure the split, not the blend. Track conversion by source, by device, and by new versus returning, and look at what the rules sold versus what the model sold. A blended conversion rate hides the truth: if the rules are selling the high-margin products and the model is selling the long tail, the blend looks fine while one of them is failing. The split is the number that tells you which one to fix.
And review what actually sold, every week, with the rule that put it there. A product that the model ranked first and that sold well is a model win. A product that the rules pinned and that did not sell is a rule to change. The review is the control, and the control is what keeps the hero product honest. A store where nobody reviews what sold is a store where the merchandising runs itself, and it will run itself into the ground.
Related reading on this site: Your return rate is a margin problem: the numbers that decide the policy, Checkout extensibility: what Shopify Plus can change now that checkout.liquid is dead, and What should a store do with a product page that gets traffic but no sales?.
Frequently asked questions
Should AI choose my hero products?
Not alone. AI is good at ranking products for each shopper, but it does not know the business priorities: which launch matters, which margin matters, which product needs to move. Let the business set the candidates and the model rank them for the shopper.
Why do new products get buried by recommendation models?
Because models rank by past sales, and a new product has no history. A manual rule that pins a launch to the top of the collection corrects for that. The rule knows what the model cannot: that the product is new and needs to be seen.
How do I know if my merchandising rules are working?
Measure the split, not the blend. Track conversion by source, device, and new versus returning, and look at what the rules sold versus what the model sold. A blended conversion rate hides which one is failing.
What is the risk of letting a model run merchandising alone?
It optimises for conversion, so it hides the products that convert worse, including clearance and new launches, and it cannot explain why a product is first. A store that cannot explain its hero product cannot fix it when it is wrong.