Upcoming Forums:

BirdsEye View

 Daniel Ek originally conceived of Spotify as ‘the future of the record store’.

People already knew what they wanted to hear; Spotify’s job was to help them find it. But in the early 2010s, the authors of Spotify Teardown write, as the likes of Apple and Amazon launched streaming services with equally large libraries, Spotify ‘began to transform itself’. It would no longer be ‘a simple distributor of music’. It was now ‘the producer of a unique service’: recommending music. Ek had come to believe that ‘the biggest unsolved question for most users is, how can you help me figure out what I’m going to listen to?’ So much music was coming out that it was hard not to feel overwhelmed. In the early 1970s, five thousand albums were released in the US each year; in 2013, almost 130,000 were released. Today the equivalent of three million albums’ worth of music is added to Spotify every year. (There are more than a hundred million songs on Spotify, millions of which have never been played; a website called Forgotify lets you stream them randomly.)

Over the last decade, Spotify has developed a range of methods to help listeners decide what they want to hear. When you open the app, the home page recommends a number of things – artists, albums, playlists – based on what you’ve listened to in the past. Many of these playlists are collections of songs put together by Spotify employees. They can be immensely popular: one, RapCaviar, has 15 million followers, giving it more influence in determining which rap breaks through than any radio station or TV channel….

When Spotify began introducing these features, it compared its role to that of a ‘new best friend’, offering ‘hand-picked recommendations’. This implies agency: you’re given a recommendation, and it’s up to you how you act on it. But that’s not how Spotify’s Autoplay function works. Once you get to the end of an album or selection of songs, Autoplay kicks in; the songs that follow are meant to be similar in some way to whatever you’ve just been listening to. (Some songs are more likely to follow than others: Spotify recently launched a service called Discovery Mode, which allows artists and labels to increase the likelihood of a song coming up on Autoplay in exchange for a reduced royalty rate.) 

Recommendations now ‘drive close to half of all users’ streams’, according to Spotify’s co-president Gustav Söderström. In Computing Taste, an ethnography of the data scientists and product managers working in ‘the world of music recommendation’, Nick Seaver gives an account of the way this sort of technology operates. The job of his interviewees, who tend to work for private companies hired by streaming services, is to help their clients ‘answer an apparently simple question: what’s next?’ There’s a tendency, when talking about a platform like Spotify or Netflix, to refer to ‘the algorithm’, as if a single formula determines what each person is recommended. But at Willow, one of the companies Seaver writes about (he refers to the companies and their employees by pseudonyms), there are ‘dozens and dozens’ of algorithms, tracking many different things: ‘What does a song sound like? What device is a user listening on? What has a user listened to in the past?’ These different bits of information are then ‘orchestrated together’ by another algorithm, which establishes how a user goes about listening to music: whether they like being introduced to new things, say, or prefer sticking to what they already know. ‘Every recommended song is a little test, a probe meant to fill out a picture of what a given user likes.’ (You can read more at lrb.co.uk)

The insights above are many, but a few stand out to me.

• Data for the most sophisticated algorithms and models can be collected sometimes by people, surveys and other analog-era data collection methods
  

• A single algorithm is often insufficient to generate fully-customized incremental value

• The loyalty generated from value-added information is enormous (note RapCaviar)
  
• Banks look to models and Martech to generate leads and improve marketing results and targeting.  The Spotify model yields informational value, reduces search time for the user, leads rather than follows and only monetizes the value after it is created

Consider how your own data scientists, whether you are a $500mm community bank or the nation’s largest, can build such loyalty through creating information value that is relevant at the user’s level to our customers.  This is achievable without a huge investment through shifting the mindset from targeting the customers likely to purchase a specific product to generating information on our customers based upon past behavior that will help you predict their areas of interest, questions about the financial services business they are thinking about, and more attractive product features, functionality, applicability and packaging.

One example that might stimulate your thinking: studying past behavior of CD buyers can tell you, at the individual customer level, what are their typical maturities or what premium over market will get them to move.  This knowledge can help your customer continue to choose you through private offers that respond directly to their preferences and concerns.

Spotify continues to compete effectively against industry giants with far greater resources by shifting the conversation and the value created.  You can do the same.