“But what I do have are a very particular set of skills; skills I have acquired over a very long career.”
Was tweeting the other day about how Twitter’s “Who to follow” recommendations lacked context and speculated the inclusion of more information might increase conversion. Here’s what the web version looks like today:
While the suggested accounts might be personalized by Twitter’s understanding of my interests, and in shel’s case it provides a bit of data (someone I follow follows him), on the whole it’s pretty basic. I feel no excitement, no engagement and have largely gone blind to the unit. Twitter is full of smart people so I don’t doubt whatever they have is optimized but perhaps to a local maxima (or at least the effort<>reward of this growth feature). What would be more interesting to me? An interesting exchange between someone I follow and someone I don’t with a recommendation to follow the latter. A single account with their last – or most popular – tweet displayed alongside so I’m enticed to follow SoftTechVC or Bitcoin. Some variation in the display that causes me to check it out each time.
Besides gut instinct why do I think Twitter might want to explore these ideas? Well, it would be that aforementioned ‘very particular set of skills.’ During my years at YouTube I had the opportunity to observe many of our experiments in providing context. One that feels particularly relevant was our own launch of recommended videos based on your viewing history. While it performed fairly well at launch there was one tweak made early on which increased clickthrough and perceived relevance: we told you why we were recommending a specific video. Although the UX has since changed to remove this additional information, here’s an example from several years back:
You can see the “because you watched: <title>” displayed (although this isn’t a great screenshot since it cuts off most of titles, but you get the general idea). The exact numbers are hazy but I recall when we launched this additional information, two things happened:
- Clickthrough on the unit went up ~20% over previous results (and page CTR increased as well so it was additional clicks, not pure cannibalization – although there was some cannibalization)
- When we recommended a video someone didn’t like, the user blamed themselves and not us!
The first benefit is self-explanatory but the second one was unexpected, amusing and deserves a bit of elaboration. Let’s say you come to your YouTube homepage one day and we’re recommending you watch a Justin Bieber video. “No way” you might say, “I’m NOT a Bieber fan. YouTube’s personalized recommendations suxxxx.” When we showed you that same recommendation but told you we were recommending it because you had watched Miley Cyrus “Party in the USA,” all of a sudden the user was all “um yeah, I guess I did. Heh heh.” Now of course our goal was to provide the right video, something they’d enjoy and click on, but it was interesting to see the change in perception this context provided (an impact we deemed as positive).
So I’m generally a fan of experimenting with additional context when it provides a look inside your algorithmic black box. I think it makes your product seem more human to its users and feels like your algorithms are working on behalf of the person and their interests rather than just treating them as a row in a user database.
Would be very interested to hear about other people’s experiences that had similar or different outcomes with regards to context.