• by Steve Smith , Staff Writer @popeyesm, October 15, 2010

Predicting what site visitors might want to read next most often is done by semantic or contextual analysis of the content they are reading now. A recommendation engine can pull from a page its topic and related categories, or even use the aggregated predilections of previous users who read that or similar articles.

When it comes to video recommendations, things get a little trickier. Not only is video content harder to categorize, but the reasons people pick a video to watch and then like or dislike are based on a number of variables that the content of the current video or other text on the page cannot determine.

"Recommending video is a problem because we don't have a lot of information about what users want to watch next, who the user is and what they want to do," says Adam Singolda, CEO and founder of Taboola. His video recommendation technology is used at CNN, Bloomberg, and in a new deal at NYTImes.com. The company name Taboola is a reference to the Latin "tabula rasa" or blank slate. Typically a video viewer or potential viewer comes onto a page giving the provider little clue as to what video they might want to see. Taboola tries to fill the slate.

When the company started developing its technology four years ago, semantic and contextual analysis of the page were the baseline methods of determining recommendations. But "we found that many people were disappointed," Singolda says. If you just build an algorithm around the meaning of a page, then you don't capture whether a user wants a longer or shorter video, whether they will care about its quality, etc. You need to know how people watch video, not just consume text and images.

"The way to create uplift in engagement is if you are able to study the way people behave and interact with videos," he says. And so Taboola tags and tracks its users anonymously across the video sites on which it works. A user coming to the NYTimes site may have looked at CNN or Demand Media videos. These cookied users have already given a video profile to Taboola that doesn't just include favorite content topics, but also video habits.

In order to determine the right videos to recommend to a user on the page, Taboola will study time spent with previous videos. Via Javascript, the company can cluster clickstreams on the page to see what the mouse was doing while users watched a video, whether they really were paying attention or clicking onto another tab. All of these elements can go into determining not only the content but the video type to recommend.

Singolda says that Taboola reaches 51 million users a month, but even when a visitor comes in truly '"tabula rasa" (cookie-less), the system makes fast work of determining recommendations on the fly. Where users are coming from, on the Web and in the world, what time of day they access the site, what first choice they make when presented with five video recommendations, all go into the mix on top of contextual and semantic analysis of the page. "Users make real-time decisions," he says. "If they choose that third video out of five, it transforms the journey. If he comes to the site from Tel Aviv at a certain time of day, we know that people coming from Tel Aviv at this time of day tend to be concerned with certain topics." Each "journey" or video playlist of recommendations on the page is customized to that visitor and sometimes divorced from the specific content of a page.

The basic metrics Taboola uses to prove value and success to partners like Bloomberg and NYTimes are the video viewed per user session, as well as the number of clicks on the recommendation window on the side of the page. There are two Taboola products: one that is designed to move readers of text pages into video experiences, and another to keep video viewers watching more videos. The partners who use both products see a video per session lift of 70% to 300%, Singolda claims. "We usually double or triple video views."

But proving value can be tricky in this environment, Singolda argues. A number of variables contribute to changes in metrics, from site redesign to external marketing pushes. Publishers can claim that their own internal activities affected video performance. It creates a problem for a vendor like Taboola, he says. "'How can I isolate value for your business?'" a partner might tell Singolda. "'How do I know it is you delivering the value?'"  

His answer is A/B testing on partner sites. About 1% to 5% of visitors to the site will be served baseline video recommendations based mainly on context and without the benefit of the Taboola algorithms. On a day to day basis, then, he can show partners how the videos viewed per session by the baseline group compare to the 95% or more who were getting recommendations powered by behavioral and other algorithms. "It is a problem for companies like ours, so we quantify the value."