{{ p.title }}

A ranking model optimizes what it is trained to optimize, usually some measure of how likely a shopper is to engage or buy. A business cares about more than that. It needs margin on what it sells, it has promotions it has committed to, and it has products it may not show to some shoppers or in some regions. A model that ignores those rules produces rankings that are good by its own measure and unusable in practice. The question is how to make the rules hold without undermining the ranking.

Four kinds of rule

Business rules for ranking fall into a few types, and each behaves differently:

• Exclusions remove products from consideration entirely: items out of stock, age-restricted products, items not licensed for a region. They are absolute.
• Floors and caps limit what can rank: a minimum margin, a maximum share of results from one brand, a price range for the context.
• Pins fix a product in a position, such as a launch the business has paid to promote or a contractual placement.
• Boosts nudge the ranking without forcing it, for example toward products with deeper stock or a seasonal range.

In our system, merchandisers configure these as guardrails, and the ranking engine applies them on every decision.

Where each rule belongs

Rules can be applied before ranking, during it or after it, and the choice changes the result.

Exclusions belong before ranking. Removing ineligible products from the candidate set is cheap, guarantees they never appear, and keeps the model from spending effort on products that cannot be shown. Deleting them from a finished ranking also works, but it leaves gaps that have to be backfilled.

Floors, caps and boosts belong inside ranking, as constraints and adjustments the scoring respects. Applied afterwards as a filter on a finished list, they can strip out the top results and leave an order the model never intended. Applied inside, the model ranks among products that already satisfy the rules, so the result is the best ranking that respects them.

Pins are the one rule that sensibly applies after ranking, because by definition they override it. The rest of the ranking then fills the remaining positions.

Rules and reasons

A reason describes why a product was chosen. When a product appears because of a pin, a fit reason would misdescribe the decision, so pinned and promoted positions should be labeled as such, for example as featured, instead of carrying a reason the scoring did not produce. Exclusions never reach the shopper, but they should be visible to the team. When a merchandiser asks why a product never appears, the decision record should show which rule removed it.

Rules and measurement

Rules change what a holdout measures. If the treated group receives model rankings with guardrails and the holdout keeps the old experience, the measured effect is the effect of the whole system, rules included. That is usually the right comparison, since the business will never run the model without its rules. It does mean a heavily constrained system can show a small effect for reasons that have nothing to do with the model, and the guardrails are the first place to look when the gap is smaller than expected.

Rules also accumulate. Each one is reasonable when it is added, and together they can leave the model with little room to rank at all. Reviewing guardrails periodically, and removing the ones that no longer serve a purpose, keeps the ranking from turning into a list of exceptions.

Trade-offs

Every rule costs some ranking quality by the model’s own measure. A margin floor removes products the shopper might have preferred, and a pin occupies a position a better match could have held. Those costs are often worth paying, and applying rules deliberately means paying them knowingly. Rules are also only as good as their data. An exclusion based on stale stock data will show out-of-stock products or hide available ones, and a margin floor is meaningless if cost data is incomplete.

There is a governance question as well. Merchandisers need enough control to keep the last word on what the storefront shows, and enough visibility to see what their rules are doing. A system that offers rules without showing their effect invites over-configuration.

The Decision Inspector shows eligible and blocked products for a worked example, with the rule that blocked each one.

Open the Decision Inspector