Why contextual analysis kicks behavioral’s butt on content

For the past year, we at Sphere have been waging a polite and respectful war on our cousins in the space, powering related content using behavioral analysis.  We’ve believed and maintained that behavioral analysis (ie: forming relationships based on repeat user connections) cannot compare to contextual analysis of the content, mainly because readers don’t always group their reading so systematically by topic.  I blogged about this last year and our opinions/learnings haven’t much changed.  We think behavioral matching is a cool technology and extremely valuable in other applications – such as commerce.  I love the feature on Amazon.com where it tells me that 64% of users bought the item I’m looking at, 24% bought another one and still, 14% bought a third.  It makes my shopping experience more informed and leaves me feeling better about my purchases.  Likewise, it’s helpful, when buying an iPhone, for instance, to know that most previous purchasers also bought the charger set and plastic case.  Content, though, is a different story.

I don’t consume content in the same organized way that I buy stuff.  My topics of particular interest right now, in no particular order, are: politics, economics, the stock market, technology, digital media, Pitt basketball, Duke basketball, convergence, Apple products, Israel, venture capital, Brooklyn and indie music.  On any given day, I’ll peruse my rss feeds in Google reader and surf from site to site gathering news on these topics of interest.  I might read an article about Duke’s upcoming matchup against UNC, followed by an article on the stimulus bill.  Knowing my friends and their interests, there’s a good chance that many of them follow the same patterns.  The problem, for behavioral, is that these two topics have no obvious relationship to one another, yet there is no way to decipher it.  This was illustrated beautifully today in an article that appeared on our partner Time.com.  In this article about the Israeli election, like all Time articles, our contextually related links show up side-by-side with those powered by Loomia, who use behavioral analysis.  The first time I read this article, this is what I saw.  Beneath door #2, Sphere is powering related stories based on contextual analysis of this article versus Time’s archive.  All three stories seem to fit nicely into the sphere (lower case) of Israeli politics…way to go team!  Behind door #3, Loomia is powering results based on behavioral analysis (ie: users who read this also read….).   While all three stories are interesting – particularly the one about Israeli model Bar Refaeli – none of them have any real connection to Israeli politics or specifically, the Israeli election.  What’s more, when I returned to the article later in the day, I saw different behavioral results.  Since it’s based on surfing and reading habits, not on contextual analysis of the content, behavioral is especially likely to return articles that are popular that moment, but may not be later in the day after their newness has worn off and a new flavor of the moment has surfaced.  Here is what I saw the second time around.  Sphere’s results under door #2 are identical as before, as you’d expect unless a new article was published on very similar topics bumping one of these three out.  Behind door #3, there are three completely different stories, arguably even less relevant than the previous results, though Ron Jeremy can usually be counted on for a few clicks.

While this is only one example, it’s quite illustrative of our view in that it provides an opportunity to compare behavioral and contextual performance on content side-by-side.  If you’re a publisher considering the pro’s and con’s of behavioral versus contextual content recommendations, please get in touch and we’ll help you explore the the options and achieve the best performance for your site.