A Sonic Boom
On Monday, Spotify publicly announced the launch of Discover Weekly, taking out of private beta their signature effort in personalized playlists. After a few days of listening to my playlist, I have to say it’s outstanding. The thirty song playlist, updated every week, is exactly what I most want and stress over in music exploration – artists and tracks that I will like, ideally that I haven’t come across yet, that match a particular moment in my musical taste.
Spotify drew on a methodical human-machine blend of content and user analysis to achieve what feels like a massive leap in recommendation engine efficacy. The platform’s internal Truffle Pig database and search engine slices and dices every song into each machine-tagged data point or human-input characteristic possible (i.e. ‘1970s Jam Band’ or ‘turns your knees to jelly’). With this degree of specificity, machines and humans then monitor user interaction and listening behavior - from all 75 million users, within groups of users like yourself, and down to your own track by track habits – and creates a playlist not just of music that matches what you’ve listened to but more appropriately what you would want to hear next or more of. The analysis goes as far as eliminating serious outliers to your known tastes (for me: Katy Perry, Dark Horse). Basically the recommendation gets smarter every month. Wired published a great piece exploring Spotify’s perfect playlist here.
Music may well be the cultural cog and unit of Internet consumption that we are the most picky about. We want what we think we want, although sometimes we don’t know what that is, but we always want it right away. Usage analysis + machine learning = dynamic recommendation has long been the equation in sales pitches and marketing splash pages for streaming services as well as many other online publishing and commerce companies. Until Monday, none of them were sophisticated enough to capture the fine line of accurate individual recommendation consistently.
With the exception of the person who creates it, everyone discovers anything new from some direct or indirect recommendation. A frustrating state of things online is that as users of every platform on the web, we’ve been putting ourselves – what we like, what we don’t - out there in a clear, measurable way for the better part of a decade. Yet most recommendation engines, which should lead each of us down Internet Interstate ‘Me’, don’t go much further then matching against a set of recurring keywords or purchase history.
A few services that would benefit from exploring Spotify’s recommendation strategy:
• Any RSS feed aggregator and just about every major media publisher, old and new.
• Other streaming media services – despite detailed data, Netflix reccs are too rigid.
• Retail / Apparel – especially any large marketplace or e-comm. brand with more than 100 SKUs. Granted not as much data exists per customer. Services like Scratch, a NextView portfolio company, are initiating with a 100% human recommendation to jumpstart user interest data in gifts, home goods and apparel.
• Twitter – the service recommends followers when a new account is created based on user inputs during sign-up but the service should have troves of information on who users elect to follow thereafter, types of accounts they interact with etc.
• Facebook – the newsfeed could reflect a much more personalized stream of things I genuinely want to know, see or read as opposed to mindless click-bait.
• Local Places, Events – Songkick, Yelp, Foursquare et al. could have extremely accurate ‘recommended’ lists available to all regular users that push to lock screens when someone is near a recommended place. Past 4Sq attempts at this have been piecemeal and spotty.
The problem with lookalike recommendations is that we don’t often want the next closest thing to what we’ve just read or recently purchased. With Discover Weekly, Spotify is approaching a sophistication that’s anticipating and encouraging the expansion of one’s taste. One foot out the door of our consumptive comfort zone: similar but not the same.