"Should we sort with AI or with rules?" is usually asked as if one had to win. In practice they answer different questions, and a store of any size has collections for both.
What a sorting rule is good at
A sorting rule is a formula. It takes signals, such as weekly sales, conversion rate, stock and days on market, and turns them into a score. You, or someone on your behalf, decide which signals count and how much.
Its strengths follow from that:
- It is predictable. You can read the rule and know why a product is where it is.
- It carries policy. "Newest first in the new arrivals category." "In clearance, the deepest discount first." These are decisions, not discoveries, and a rule states them exactly.
- It works from day one. A rule needs today's data, not months of history.
- It is easy to defend. When a brand manager asks why a product moved, the answer is one sentence.
Its weakness is the other side of the same coin. A rule only knows what you put into it. If season matters more for coats than for socks, the rule does not find that out. Someone has to notice and write a second rule.
What a model is good at
A machine-learning model does not get the weights from you. It learns them from your data: how price, discount, stock, brand, category, season, color, freshness, views, add-to-carts and sales have related to demand in your store. Then it predicts a score for each product.
- It weighs many signals at once, including combinations nobody would write into a formula.
- It adapts. Retraining on new data picks up shifts in demand without anyone editing anything.
- It is specific to the store. In Merchandiser every store has its own model, trained only on that store's products and performance. What sells in a homeware shop does not leak into a fashion store's ranking.
The costs are real too. A model needs enough history to learn from, and "why is this product third?" has a less tidy answer than with a rule.
Where each one fits
| Collection | Better fit | Why |
|---|---|---|
| New arrivals | Rule | The policy is the point: newest first. |
| Clearance and outlet | Rule | You decide what to clear: discount, stock or age. |
| Campaign pages | Pins plus a rule | The campaign products are decided; the rest follows a simple order. |
| Large core categories | Model | Many products, many signals, plenty of data. |
| A new store, or a category with little traffic | Rule | Too little history for a model to learn from. |
| Brand pages | Either | A rule is often enough; a model helps when the brand is large. |
The pattern: use a rule where you know what you want, and a model where you want to find out what works.
What they share
Whichever you choose, a few things should hold.
Availability belongs in both. A product that cannot be bought should not be at the top, however good its numbers. A rule should multiply demand by size availability. Merchandiser AI does this by design: each prediction is weighted by the share of the product's sizes that are in stock. See Size-Broken Products for why this matters.
Pins sit above both. A pinned product keeps its position under a rule and under the model. Your decisions are not something the automatic order gets to argue with.
Your priorities can be added to a model. A model predicts demand; it does not know your margin targets or the season you want to push. In Merchandiser, products can carry a boost that is added to the model's estimate, which is how business priorities enter an AI-ranked collection.
Both should be previewed. Before switching a collection from one to the other, look at the resulting order. In Merchandiser the sort page shows the preview, and nothing changes on the store until you save.
How to start
- Put every collection on a sensible rule first. That alone removes manual sorting and makes sure availability counts.
- Let performance data build up.
- Move the large, busy categories to the model, one at a time, and compare the previews.
- Keep rules where the policy is explicit.
- Review the choice when the store changes: a new category starts on a rule, a grown one may be ready for the model.
How Merchandiser's model works, what it learns from and what it predicts is described in Merchandiser AI.