Merchandising Rules Best Practices

Overview

Boosting, filtering, or hard-coding products in recommendations can have a material impact on incremental revenue. For example, we have seen significant gains in Apparel by applying a boost to ensure that recommended similar items are at least 85% the price of the seed product. We have not logged sufficient precedent to confidently suggest applying these optimizations out-the-gate. Rather, they should be validated against a control environment through an AB test to ensure that they work for the specific retailer. Following the guidelines set forth in the pages that follow will help to safeguard performance of Recommend on a retailer's site.

All sites are different, and what works best for one is not guaranteed to translate into gains for another. As such, treat this guidance as general best practices, a foundation for optimization from which site-specific tuning should occur. Understand the logic behind our optimization methodology, the questions to ask, so that you can adapt and apply as appropriate.

Merchandising Rules

In general, we endorse merchandising by exception or to satisfy very specific use cases that cannot be met with pure behavioral recommendations. For example, on a home improvement site, we might boost items above $5 to ensure that common, inexpensive purchases (for example, nails) do not frequently show up in recommendations. Anything broader should be tested as the client can very easily create a rule that hamstrings our system, limiting our ability to drive incremental revenue.

For configuring Boosts and Recommendation Restrictions, we have the following best practices:

• Monitor our models to ensure that inexpensive items do not commonly surface in top-seller recommendations—so that we can use the real estate for higher ticket buys (for example, Office Depot likely does not need our help selling paper clips but would appreciate if we could increase incremental sales of monitors). If you see this occurring, consider boosting items above a specific price threshold if appropriate for the retailer.
• Whenever possible, use a Boost over a Recommendation Restriction as this will prevent a scenario whereby all products are filtered from display. For example, suppose you want to upsell on the item page and recommend products priced higher than the seed. If the seed is of the highest price percentile, a Recommendation Restriction will filter out all products since there are no high-priced options remaining; a Boost would ensure that recommendations are still available.
• If you choose to configure a Recommendation Restriction, strive to set up an “Only Recommend” inclusion rule rather than a “Do Not Recommend” exclusion rule. If the parameters that we are targeting—for example, category or a specific attribute value—a product can have multiple values, a “Do Not Recommend” rule could filter out an unintended number of items. For example, a Do not recommend women’s clothes in men’s rule will filter out unisex products, because the system sees that they are categorized in “Men’s” and therefore discards them. Instead, configuring this as only recommend women’s clothes in women’s will retain these products as the system only requires these products are categorized in “Women’s” regardless of what other categories the product might belong to.