Deep learning product recommendations provide the most relevant product selection to each customer using real-time customer data to predict what every customer wants next. But ecommerce teams can’t always let the algorithm run wild, completely unrestricted.
Flexible business controls or ‘recommendation rules’ add the human touch to the machine, They provide a powerful way to finetune recommendation strategies so that the product selections are more closely aligned with campaigns, promotions, and wider business goals. These controls effectively act as ‘AI Overrides’ by incorporating a business logic layer overlaid on top of the algorithmic strategies
We’ll outline below when it’s best to incorporate business controls including promote, blacklist and conditional rules and why it’s always best to start with the algorithm as un-ruled as possible.
Why it’s best to start the algorithm as un-ruled as possible (and then, layer in rules over time).
Deep learning product recommendations are constantly learning in real time to each visitor’s journey. Applying lots of rules with complex attribute combinations will restrict the effectiveness of the learning process in generating the most relevant product selection for each visitor. Because of this, it’s best practice to carefully plan your recommendations strategy with only absolutely necessary rules to start. Once the algorithm has had enough learning time, rules should be added iteratively, determining the effects of those rules, before incorporating any additional rules.
When ecommerce teams should implement rules, no questions asked
When there is a contractual obligation.
It’s commonplace for multi-brand retailers, general merchandisers and department stores to have restrictions on which brands they can merchandise side-by-side and which they cannot –
-i.e. competitor brands that cannot be positioned against each other. In this instance, a conditional rule would be applied e.g. if the visitor is viewing x brand product, don’t show y brand products.
When the product catalog includes non-revenue items such as samples.
Some retailers have non-revenue items in their product catalog that are only to be shown on certain pages or under certain conditions. For example, beauty brands that leverage samples to get visitors deeper into their product catalog as a risk free trial. In this instance, sample products would be blacklisted from certain carousels or only restricted to the cart page.
When to test rules
When there are explicit product pairs
Certain products have defined product pairings or complimentary product offerings that should be shown simultaneously as means of cross-selling. An example is if a visitor is viewing a bike, then they should be shown a maintenance kit. In these instances, an override carousel slot would be used to force the maintenance kit to be shown when the visitor is viewing a bike.
When the business wants to promote certain categories.
To best align to business goals and targets such as the launch of a new or seasonal collection, the ‘promote’ rule can be leveraged. This increases the likelihood for products within that category to show up to visitors.
When the category shown should be restricted.
Certain categories are retailer’s bread and butter and if a visitor has landed on one of those pages and there is only one recommendation carousel, it’s best not to direct their on-site journey, elsewhere. For example, if a visitor is on a bags page of a luxury retailer, recommendations should be restricted to exclusively show other bags using the ‘only show’ rule.
When price banding should be implemented.
Price banding can be leveraged in a couple of ways as aligned with wider business goals including on the PDP and on the basket page. On the PDP, merchandisers can leverage rules to only recommend products that are within a certain price range of the product being viewed. On the basket page, rules can be implemented to recommend products to get visitors above a certain threshold and increase the value of their basket.
Business controls are powerful mechanisms for ecommerce teams to add an element of curation to product recommendations as aligned with business needs or goals. We always recommend to try to start with the algorithm “un-ruled” as possible, then layer on rules over time. Doing so sets the baseline and maximizes performance. The last thing ecommerce teams want is to be 6 months in leveraging deep learning recommendations and discover that a rule based on an early assumption was costing 0.5% RPV.
At Qubit we are data-driven, enabling us to be laser focused on achieving high success rates for our customers. Generic solutions that treat every customer the same breed generic results and that’s why we’re always at the forefront of new technologies that enable us to drive a truly personalized ecommerce experience. We’re here to make ecommerce better. If you think your product recommendations are underperforming or you’ve seen them plateau we invite you to sign up for a free recommendations health check.
Jessica Sta. MariaRead More