How to rank recommendations when margins differ wildly
When one product makes 60 percent margin and the next makes 8 percent, a relevance-only ranking will happily fill the recommendation rail with the 8 percent item. The business wants margin in the ranking, but shoppers punish recommendations that feel like ads. The answer is a blended score with guardrails, so margin influences the order without ever overriding relevance.
The margin-relevance tension
Every recommendation engine optimizes something, and the default something is predicted engagement: clicks, views, add-to-carts. That objective is shopper-aligned but business-blind. A rail that perfectly predicts what the shopper will click can still be a bad rail for the business if everything it predicts is low margin.
The naive fix, sorting by margin, fails immediately. Shoppers spot a rail of high-margin items that do not match their intent, trust drops, and the rail's click-through collapses. The business gains margin per click and loses clicks, which is usually a net loss. The tension is real and both sides have a point.
Blending scores the honest way
The workable approach is a weighted blend: final score equals relevance score times a margin factor, where the margin factor is bounded. A common shape is relevance multiplied by one plus a capped margin bonus, so margin can promote but never rescue an irrelevant product. The cap is the whole game: it encodes how much relevance you refuse to sacrifice.
Calibrate the cap with data, not opinion. Run the blended ranking against pure relevance in an A/B test and watch revenue per session and rail click-through together. Increase the margin weight until click-through starts to decay, then back off. The right weight is the one just before shoppers notice, and it differs by category.
Guardrails that keep trust
Hard guardrails beat clever weights. Never recommend a product below a relevance floor no matter the margin. Never let margin reorder the top slot, which shoppers read as the single best pick. And never apply margin blending to categories where trust is the product, like health or safety, where a mercenary-looking recommendation does lasting damage.
Transparency helps at the edges. When the business runs a deliberate margin push, label it as a promotion rather than laundering it through the recommendation engine. Shoppers accept sponsored content they can identify; they resent recommendations that pretend to be pure relevance while serving margin.
When to ignore margin entirely
Some placements should stay pure. The cart-page cross-sell, where the shopper has already committed, converts best on genuine relevance, and margin games there feel extractive. New-customer experiences should optimize for trust and second purchase, not first-order margin. And any rail explicitly labeled as personalized or recommended for you is a promise that relevance comes first.
The rule of thumb: blend margin where the shopper expects merchandising, like category pages and homepage rails, and keep it pure where the shopper expects advice. Shoppers have a fine sense for which is which, even if they never articulate it.
Will margin blending hurt click-through rates?
Slightly, at any positive weight. The question is whether the revenue gain outweighs it, which is exactly what the A/B test measures.
Should margin apply to search results too?
Search is a promise of relevance for a stated query, so keep it pure or nearly so. Margin belongs in browse and discovery surfaces.
How do you explain the blend to merchandisers?
Show them the cap and the guardrails, not the math. They need to know margin matters but cannot override relevance, which is a policy conversation, not an algorithm one.
Reviewed
Published Oct 1, 2026.