As consumers, we are all familiar with website recommendations. They have become table stakes for most e-commerce sites, we rely on them for finding new content on streaming video sites, and most online publications offer recommended news articles to their readers, just to name a few cases.
But are these recommendations any good? How many times have you seen a recommended item and thought “I would never buy that,” “I already have one of those,” or “I was never interested in that product”?
Last week, personalization experts Greg Hinkle and T.J. Prebil (CTO and Product Marketing Manager, respectively) presented the webinar The Future of Recommendations, providing Evergage’s perspectives on how online recommendations will evolve going forward.
Greg and T.J. focused on eight key areas of advancement in recommendations technology. Here’s an overview of one of them.
Fewer, More Complex Algorithms Create More Relevant Recommendations
Let’s say that two shoppers with different preferences viewed the same black dress on a retail site but left without making a purchase. Upon returning the retailer's homepage, they are each presented with recommendations related to the black dress.
To promote products that are frequently purchased with the dress, as well as items that are uniquely relevant for each shopper, many traditional recommendations engines will use a technique called “slotting.”
Here, one algorithm will show the items that are typically purchased with the dress (regardless of how relevant those products are to each shopper) and another algorithm will show the items that are similar to other products the shopper has viewed before (regardless of whether they are relevant to the product viewed in the current session).
It may look something like this image below – where the first set of products are identical for each shopper and the last two are customized for the shopper, but aren’t related to the dress.
A better approach would be to combine multiple algorithms into a single strategy (or recommendation “recipe”) that produces a more cohesive set of recommendations that work for each individual. With a single “recipe,” marketers can recommend items that are both typically purchased together (with the dress) but also appeal to individual visitors (based on their preferred price points, brands, styles, etc.).
This more sophisticated recommendation logic may take into account, for example, that one visitor is from California and prefers a higher price point while one is from New York and prefers a lower price point.
Ultimately, the products recommended are much more relevant to each visitor. The shopper in California who prefers a higher price point is shown more expensive pumps to pair with the dress, while the shopper in New York is shown less expensive heeled boots and booties.
The Future of Recommendations
In the webinar, Greg and T.J. explored how recommendations algorithms have evolved, the importance of using real-time data, the need for marketers to test and control their own algorithms, and more. To learn more about the future of recommendations, watch the webinar replay now.