Recommendations have been all the rage for the last couple of years now, and since I’ve been working for Hunch (now eBay), I’ve been thinking about why they’ve become necessary and the value good recommendations bring to the piles of content on the Internet. Searching for information we’re looking for is basically a solved problem, and it’s relatively easy to carve up existing datasets such that they’re easily consumed. Surfacing interesting content proves to be a problem, though. Take Yelp, for example. If I want to find out when Wildwood BBQ (a fine watering hole, if I do say so myself) is open, I search for Wildwood BBQ, and it’s the first result. I learn not only when they’re open, but what kind of parking is available, whether they are wheelchair accessible, the nearest subways, and all sorts of wonderful things. A+. If I didn’t know about Wildwood BBQ but instead were interested in all of the places near Union Square in the $$ price range that served beer and had a TV, Wildwood BBQ would show up. But here comes the kicker. This is New York. There are a ton of relatively cheap places to drink and watch TV.

Now the problem becomes deciding among the available options, and for the most part, this task is up to me. Yelp is a very popular service, but unfortunately, this means that all kinds of people leave reviews, many of whom I probably disagree with. So, the average Yelp rating isn’t all that helpful. At this point, if I really care about where I go, I have to dig through the reviews for each place, deciding, based on the user’s photo, name, and review text, whether I trust that person and his/her rating for each venue. It’s rather tiring and has greatly reduced the value of Yelp for me. In reality, most of these places would serve my needs just fine, so I probably could safely choose a bar at random, but on the off-chance that it doesn’t work out, I have only myself to blame. I don’t mean to pick on Yelp, as they see this issue and have the beginnings of a decision-making feature, but it’s a pretty good illustration of the need for recommendations.

There are basically two ways to surface interesting content for users, by subscription and algorithmically.

Subscription-based content has been around for awhile, and there are basically two ways to go about it. Social services like Facebook and Twitter allow me to subscribe to content posted by certain people. The Yelp analogy would be subscribing to restaurant reviews by a particular person, which exists in the form of user profiles. This assumes that I trust everything a particular user likes. To continue the Yelp analogy, while I may enjoy the sushi recommendations for a particular user, I may not be interested in their Indian food adventures in Murray Hill or their recommendations outside of New York City. Such is the story of my Twitter feed.

The other subscription-based approach is what sites like Reddit utilize. Rather than subscribing to people, I subscribe to particular topics. When articles are posted to the “programming” subreddit, fellow users curate them by voting them up and down, and the “good” articles bubble up to the top. This puts the burden on the user to specify what they like. An analogy on Yelp would be subscribing to reviews for all sushi restaurants in the New York City area. This is a good start, but we now have a problem of specificity. The more specific the interest, the better the recommendations are, but the fewer people there are that have that particular interest. For example, if I’m interested only in sushi restaurants that serve Sapporo on draft in the West Village, either I find myself in a very small group of people with no recommendations, or I find myself in a more general group, filtering out all the restaurants I’m not interested in by hand. Still, not ideal.

The new hotness in the last few years has been algorithmic recommendations, as evidenced by the number of companies actively implementing them. This approach uses implicitly-expressed preferences (in the form of article views, item purchases, reviews written, etc) to predict my specific interests. Amazon, for example, will recommend similar or complimentary items based on other users’ purchases. The users themselves don’t explicitly tell Amazon what they like. Instead, their actions dictate what Amazon thinks about them. Consider that the next time you purchase the new Miley Cyrus album “for your niece.” Amazon isn’t alone here. Nearly every company with data is trying to do this with their data to surface good products, articles, advertisements, etc.

Even with algorithmic recommendations, there are two ways to go about it. Amazon, it appears, focuses primarily on the item similarities. That is, if I’m looking at a camera, they might recommend lenses because the people who’ve bought that camera tend to buy lenses. Users who are new to Amazon get the same value out of these recommendations as those who’ve actively used the service for years. The alternative is to use personalized recommendations, which I don’t believe Amazon focuses on to the same extent. For example, if I’m looking at a particular camera, Amazon may show me red cases with little pictures of guitars on the sides because I buy red things and guitar accessories.

To me, the biggest value-add of algorithmic recommendations is that they minimize the amount of work that I have to do to make a decision. Deciding among bars in Union Square is something that Yelp compels me to do, but it isn’t necessarily the best use of my time or brain space, particularly for such a low-importance adventure. Similarly, I don’t necessarily care what brand or thickness of socks that I wear, so I’m not likely to second-guess a recommendation given to me. It’s not always the “best for me,” but sometimes, not having to think makes it all worth it.

The flipside of algorithmic recommendations is that it’s often hard to explain why certain things are recommended. For example, Amazon may recommend a certain pair of headphones over another because people from New York who also browse Amazon on Thursday evenings and have similar click-streams to me have bought them, but that’s particularly difficult to explain, and depending on the algorithm Amazon is using, it may not even be able to do so. In contrast, if I pick whose recommendations I follow or the interest groups to which I’ve subscribed, why I see a particular recommendation is obvious.

So, as with many things in life, there are a ton of options when it comes to recommendations, and each has its own advantages and disadvantages. But the value of recommendations is very clear, given how much companies are invested in producing quality recommendations. Where search defined Web 1.0 and social has defined Web 2.0, discovery is likely to be remembered as the next focus of the Internet. I’d call it Web 3.0, but that’s hackneyed, to say the least.