Product recommendations on websites came into the public consciousness because of their prevalent adoption and continued use by two high-profile sites – Netflix and Amazon. Amazon was first with a homegrown solution that offered recommendations based on what others browsed or bought that matched items that you bought. Netflix next offered suggestions for films you might be interested in watching, based on the ratings you gave to films previously viewed. They also considered the reviews of others. In 2009, Netflix raised the profile of recommendation algorithms by hosting the Netflix Prize. It was an open competition for the best collaborative filtering algorithm to predict user ratings to films.
Both Amazon and Netflix have continuously improved their recommendations but their approaches are vastly different than what we are doing at Baynote. The key difference is that Amazon and Netflix serve recommendations on their own sites. At Baynote, we serve recommendations on other people’s websites.
Serving recommendations on your own website makes data collection and recommendation display easy – you control the site, so you can implement whatever data collection methods you want as part of your web pages. If someone clicks on one of the “customers who bought this item also bought…” items on Amazon, the click-through data is in Amazon’s own server log. Amazon knows the sales data for every item on their site.
When serving recommendations on someone else’s site, Baynote technology collects the user data from our customers’ site, processes the data and then injects our recommendations back to the customer site. We do this in a way that takes as little effort as possible on the part of our customers. A lot of effort goes into providing a system that is as easy as possible to integrate into a client’s site. And it’s also really clever. The observer tags that our customers put on their site effectively read the web page, and extract the information that is useful to us. These tags monitor the page to observe signals of engagement with the page content and send the data back to our system through a web interface.
Once we receive the data, we use it to build the predictive models that make recommendations. By providing a simple web API our customers’ websites can easily inject recommendations directly into their pages. As a result, we close the loop between observing user behavior and determining the optimal products to recommend.