How should recommendations change during a sale or promotion?
During a sale, your recommendation engine is working with distorted signals: prices drop, bargain hunters flood in, and margin profiles invert. If the engine keeps optimizing for its normal objectives, it will push the deepest discounts to everyone, train shoppers to wait for sales, and quietly destroy margin. Sale-period recommendations need different ranking weights, explicit margin guardrails, and separate measurement from business as usual.
The short answer
Treat sale periods as a different operating mode, not a busier normal. Shift ranking weight from pure conversion probability toward margin-aware objectives, suppress recommendations that only exist because of the discount, and keep a separate dashboard so sale performance does not contaminate your baseline metrics.
The biggest mistake is letting the sale train your models. If discounted purchases flow into the same training data as full-price ones, the engine learns that cheap products are what everyone wants. Quarantine sale-period interactions in training, or at minimum down-weight them.
Why promotions break recommendation logic
Recommendation engines learn from behavior, and sale behavior is not representative. Shoppers buy categories they never touch at full price, average order values spike on discount depth rather than preference, and the bargain-hunter segment, which barely exists in your normal data, dominates the traffic. An engine trained on normal times misreads all of this as genuine preference.
The margin distortion is the expensive part. A product with 60 percent off converts beautifully and the engine promotes it harder, but the contribution margin may be near zero or negative. Without margin in the objective function, the engine optimizes revenue while the business bleeds. Sale mode must price in profitability, not just clicks.
Ranking adjustments for sale mode
First, blend margin into the score. The standard approach is a weighted objective: predicted conversion times expected margin, rather than conversion alone. During sales, increase the margin weight, because the discount has already done the conversion work and the remaining question is which discounted products are worth selling.
Second, handle the discount itself as a feature, not a surprise. The engine should know which products are discounted and by how much, so it can distinguish genuine affinity from price-chasing. A shopper who buys a discounted item they also browsed at full price is a real signal; one who buys only the deepest discounts is a bargain hunter who needs a different experience.
Guardrails that protect the brand
Sales tempt engines into behaviors that hurt long-term: recommending loss leaders to full-price-intent shoppers, pushing clearance items that damage brand perception, and creating urgency messaging that trains discount dependence. Set explicit guardrails: never recommend a product below a margin floor to a non-sale segment, never let clearance inventory dominate the homepage, and cap the share of recommendations that are discount-driven.
Protect your full-price shoppers actively. Identify visitors whose history shows full-price buying and shield them from the sale experience where it would erode their willingness to pay. Showing a loyal full-price buyer a 40 percent off banner is not personalization; it is margin destruction with extra steps.
Measuring sale recommendations separately
Never blend sale-period metrics into your baseline. Conversion rates, click-through, and revenue per session all inflate during promotions, and an engine that looks brilliant in November may be coasting on discounts. Maintain parallel dashboards: sale-mode performance judged on margin-aware objectives, normal-mode performance on the standard ones.
After the sale, watch the hangover metrics: return rates on discounted purchases, repeat purchase rates of sale-acquired customers, and the time it takes for full-price conversion to recover. The true cost of a sale is not the discount; it is the behavior change. Recommendations that minimize that hangover are worth more than ones that maximize sale-week revenue.