Should recommendations factor in margin, or only relevance?
Optimize for relevance first and let margin break ties. Rank candidates by predicted relevance, then prefer higher-margin items among near-equals. Guard the system with relevance floors so margin never promotes something the shopper would not want, and measure profit per session instead of click-through rate.
The click-versus-profit tension
A recommender trained only on clicks learns to show whatever gets tapped: often the cheapest items, the most sensational products, or the ones shoppers were going to buy anyway. Clicks feel like engagement, but they do not pay the bills. Two recommendation strategies can show identical click-through rates while producing wildly different gross profit, because what gets clicked and what makes money are different questions.
Margin-aware ranking closes the gap by treating the recommendation slot as inventory with an opportunity cost. Every carousel position shown to a shopper is a chance to sell something; the engine should prefer the options that are both wanted and profitable.
When margin-aware backfires
- Trust erosion. Shoppers notice when recommendations feel like upsells. One irrelevant premium push teaches the customer to ignore the carousel entirely, destroying future value.
- Short-termism. Maximizing margin per session can sacrifice the entry-level products that create customers in the first place. The first purchase is often low-margin by design.
- Category distortion. Heavy margin weighting can starve entire categories of exposure, shrinking assortment perception and long-term demand.
The tiebreaker model
The practical implementation: score every candidate on relevance first, keep the top tier, then re-rank that tier by margin. Only items the shopper would plausibly want compete on profitability. The margin signal never rescues an irrelevant product; it only chooses among relevant ones.
Tune the blend explicitly. A common starting point is full relevance ranking with margin as the secondary sort key, then gradually increasing margin weight while watching relevance metrics. Stop when relevance metrics move, not when profit peaks on paper.
Guardrails
- Relevance floors. Set a minimum predicted relevance for any recommendation. Nothing below the floor shows, regardless of margin.
- Diversity requirements. Cap how many slots margin optimization can allocate to one category or price band, so the carousel still looks like a store, not a clearance rack for high-margin goods.
- New-product protection. Exclude new and low-data products from margin weighting until they have enough history to be judged fairly on relevance.
How to measure
Replace click-through rate with profit per session as the primary metric, with relevance metrics as guardrails. Run the margin-aware variant against a pure-relevance control and require both: profit up, relevance flat or better. If profit rises while relevance falls, you are borrowing from the future. Report the tradeoff honestly so the business can price it.
Segment the readout by customer type. New visitors, one-time buyers, and loyal repeat customers respond differently to margin-weighted ranking, and the aggregate can hide a pattern where the strategy works for one group and backfires for another. The decision to keep margin in the ranking should survive this segmented view, not just the headline number.
Reviewed
Published Sep 28, 2026.