Per Oct 2013 Internet Retailer report, 76.4% of the ecommerce sites in the Internet Retailer Top 500 now feature product recommendations but only 38.8% say they offer both product recommendations and site personalization. As end users, we have come to expect awesome tailored content, similar to our very own Facebook and Twitter newsfeeds, to reduce the time to discovery and to help us achieve our goals faster, while making us feel confident of our decisions. In this blog, I have attempted to compile types of recommendation engines that exist today, and examples where each is successful.
To begin with, there are two fundamental user intentions: search & browsing. Of these, browsing (exploratory intention) holds the excellent opportunity for a recommendation system versus search (specific intention). Recently mentioned by Carlos Gomez-Uribe, the VP of Product Innovation at Netflix ,
“the company’s 75% of the viewer activity is driven by recommendations. Our [Netflix’s] search feature is what people do when we are not able to show them what to watch. In contrast, 90% of what people buy on eBay comes from search.”
On one-hand, the likes of Netflix & Amazon have built & refined their homegrown recommendation algorithms over past few decades. Amazon’s item-to-item collaborative filtering algorithm is built upon existing purchases/ratings of other users who have similar purchases/ratings with the active user. Netflix’s recommendation engine goes a step further and captures viewer’s behavior such as videos played, searched, rated by the user, their scrolling/pausing behavior as well as the time, date when the video was viewed. On the other hand, there are reputed third-party providers such as Richrelevance, Bloomreach, Certona that enable such engines for other online retailers. The site objects for these recommendations could include the homepage, product pages, category pages, shopping cart, marketing campaigns, promotional offers, or search results.
Third-party providers classify their algorithms into one of the item-based, category-based, behavior-based, or popularity-based categories and these are mostly driven by the Wisdom of Crowd. However, this does not always capture some of the emerging categories based on Wisdom of Experts and Wisdom of Friends, terms that described in the OCDQ blog. Note that, as we move from item-based to profile-based category, we are moving closer to personalization.
Here is my attempt at drawing the range of user experiences that start at item-based recommendation and move towards profile-driven personalization:
The goal of targeted content delivery is to drive more sales and the cost of implementing such engines is not trivial. Pandora’s music genome project took 30 experts over 5 years to complete it. Hence, is important to evaluate the business case for whether the company needs one, and if yes, prioritize the options that can deliver most relevant results. High quality recommendations drive conversion rates, while poor quality recommendations ruin user experience.
Two scenarios where recommendations can prove most useful are:
- Company has behemoth amount of products/SKUs that are not being ‘searched’ for currently
- Consumer’s browsing behavior or market basket analysis, indicates a mood/taste pattern, or things frequently viewed/purchased together