Recommendation Engines: 5 Lessons from Pandora, StumbleUpon and YouTube
I went to SXSW this year but happened to miss a really cool panel on Recommendation Engines. Luckily, it was recorded here. I figure that most of you have companies that allow for an endless number of combinations, so I think that some of the lessons from this panel might be interesting to you, too.
What I had in mind when I took notes from the recording is that MyCereal.com, before it shut down, didn’t ask consumers to pick the different ingredients for their cereal, but rather asked them questions such as “Do you like nuts?” or “Do you want to lose weight”, and then gave the users recommendations based on that. This is a new twist on the idea of a configurator: Rather than letting the consumer really design their own, you create a product for them that they can then alter. That is also one of the reasons why NikeiD is so successful: People start customizing when they find something in the standard catalog that they kind of like - with NikeiD they can make it truly perfect.
So, here are the lessons from the panel:
LESSON ONE: Make it simple
Pandora realized that the radio is an inherently simple experience: You press a button, and you get music. They had a lot of ideas for features and more control for the user over time, but found that it was important to keep the experience simple and accessible to all users. I think that translates into customization - don’t ask your consumers to do too much “work” before giving them anything (or at least try to make it really seamless and fun!)
LESSON TWO: Offer both broad and close matches
Not everyone knows yet what exactly they want. YouTube found that most of the searches on the site are broad, and so in their recommendations on what to “watch next”, they both give close matches to the content of the video as well as broad matches to the overall thematic cluster or the type of video. So as I’m thinking of cereal recommendations, I imagine a cereal that exactly matches your preferences, as well as one that has maybe been created by a nutritionist, that’s very delicious but maybe doesn’t have all the ingredients you indicated beforehand.
LESSON THREE: Provide users context why you recommend what
You could have a black box (as in: a complicated algorithm) that spits out recommendations that you want your users to eat without asking questions. Both Pandora and StumbleUpon however recommend to keep it comprehensible to the user why you recommend what - not only because users like that more, but also because they found it easier to innovate and keep making it better by keeping it traceable. All of the companies found a way to measure how successful their recommendations were.
LESSON FOUR: Make it social, but not too social
Social recommendations are for example when you connect to Facebook, and now the cereal mix that you see has been liked by a friend. Or the cereals mixed by friends just rank higher on the list of recommendations that the user gets. This is a great marketing tool: Not only does it build trust in the product (it’s easy to show data that people are more likely to click on something they friends “liked”), it’s also a starting point for a conversation with a friend, enforcing the connection to the brand or product. Also, in a weird way, influencers are turned on by the fact that no one has liked something yet - they feel that they discovered something valuable that they can share with their friends.
At the same time, the panelists stressed the importance of serendipity, and they suggested not getting all recommendations from a social graph - because then users won’t get the “really wacky things” as often.
LESSON FIVE: Tom Conrad is funny and insightful
Tom Conrad, the CTO of Pandora, is hilarious. He’s also extremely insightful, and one of the thought leaders in personalization. Follow him at @tconrad on Twitter and go to a panel if he’s talking - I first met him at the Mass Customization Conference at MIT in 2010, where he was a fellow presenter.
Want to get started?
All the panelist companies are building their own native recommendation engines, but if you want to plug into an existing taste graph, they recommend Hunch.