Discovering The Recommendation Network

“Discovery” is the new buzzword in the digital lexicon, as everyone looks for more efficient ways to push relevant content and offers to users. Surprising people with the offers and articles that they didn’t realize they wanted is the essence of discovery, according to Paul Martino, CEO of Aggregate Knowledge. The company,  which recently closed a $20 million investment round, uses a super-fast recommendation engine of content, offers, and services across 25 sites, including WashingtonPost.com and Overstock.com. Currently in beta, the commercial network launches in Q4.

Behavioral Insider: How is ‘discovery’ different from other ways of accessing information online?

Paul Martino: Discovery is how you find stuff when you don’t know what you are looking for. That is the difference between discovery and search. It is that serendipitous, unexpected, non-keyword-driven way to find things.  You know it when you see it. It can appear in any number of ways. It can appear as ‘people who read this also read that,’ or ‘people who read this article eventually purchased this product.’

Behavioral Insider: How does your technology work, especially the behavioral engine behind it?

Martino: We built a high volume, high performance supercomputing system for behavioral computation. We can literally deal with hundreds of millions of data points of behavior to be able to identify hot movers, to [show that the] people who bought this, bought that. On a site like Overstock, we are able to compute the personalized recommendations in real time. Discovery powered by us on sites like the Post are no more than a few seconds out of date. Discovery is very much about, I am currently engaged right now. Show me what it is I need to know.

Behavioral Insider: Is speed the key differentiator in this category?

Martino: There are a few areas. Being truly real-time is extremely important. Also, there is the customer facing use case. First, we don’t necessarily put an advertisement in front of the user. We put another piece of content, whether it is an offer, a product, or a service. We are not a traditional ad network. We are a content, product, and service syndication network.

Second, we work heterogeneously. We will be able to put heterogeneous products services next to others. It is not just books for books. It might be,  people who read this news article ended up buying this CD. Or people on the social network who viewed this profile also found this discussion forum to be of interest. That is where we have insight; this heterogeneous approach to discovery. I was one of the founders of Tribe.net and we learned [not just to] show people more profiles if they were browsing profiles. The key was to get the heterogeneous types of data in front of them.  

Behavioral Insider: But you are also aggregating data across platforms.

Martino: I can recommend certain products to you better if I know about your news browsing behavior than if I know what your recent product history is. I can predict more accurately in some cases, based on a discussion forum you joined, what restaurant you might want to go to. Those data sets are in disparate sites across multiple providers. That is what we allow.  

Our system is completely anonymized and in aggregate. We look at a user identifier which is randomly assigned. We are using a third-party cookie that ages out quickly to take a look at where a user ID appears multiple times in the network. We are not a traditional personalization approach. A traditional approach will actually say, this is your user ID; you’re 24 from San Francisco and you bought these twelve items. We, on the other hand, know some 24-year-old users at some point in time looked at this product but ended up buying that other product, so there is a good chance this other person will be one you will want to buy as well.

Behavioral Insider: Are you selling placements in this network in segments like a traditional behavioral network?

Martino: There is no notion of a behavioral segment in our system. There is no notion that this 24-year-old would be interested in Xbox games because of this particular purchase history. We actually know for every single item in the system, every SKU or article or item, what the other affinities are across all other items in the system, whether on the site you are on other sites. We do that dynamically in real time. There is no notion here of buying a behavioral segment, it is more the notion of automatically merchandising the right products and offers right next to the right content product. You don’t come in with the traditional ad network kind of buy. You come in and say, I am willing to spend X dollars to place my product at the right spot -- go to town.

Behavioral Insider: Text or display?

Martino: We can do display ads as well and syndicate out actual product images. But we found that the actual explanatory text on the discovery window is very important. You wan to explain to the users there is a reason this thing is here; this is not just a paid placement. We’ve analyzed hundreds of millions of people who have gone before you. We will do you a favor and put in front of you another place you will be interested in. That text at the top of the box is important to create that trust with the user.

Behavioral Insider: But there is only so much room on the page for so many links. How do you get in?

Martino: Every one is fighting an effective CPM per pixel fight with one another in the truly long run. Our value proposition is different from strict monetization. We actually help make real estate better in terms of navigation. On some sites we have click ranges of 18% to 22% on our box. We are primary navigation. We say, not only will we make more money, but we actually make your user like your site more, and give them more page views per session.