A visualization of every song on every Spotify Release Radar playlist over the course of a year. Filled in squares equal songs that were listened to.

I Decoded the Spotify Recommendation Algorithm. Here’s What I Found.

A print version of this analysis can be found here.

On April 7, 2018 at 12:00 AM, Spotify placed Mythological Beauty by Big Thief on that week’s Release Radar playlist. It was the first time I was recommended a song from Capacity, the album that would go on to become my favorite of the year. The last time I took a deep look into music discovery was in 2011. That year my favorite album of the year was recommended to me by the clerk at Slowtrain Records, which is now closed down.

Back then, I knew music consumption and discovery was changing, but I never would have guess that in less than 7 years it would change so dramatically. The amount of new music I listen to higher now and the sources of discovery have been overshadowed by algorithmic recommendations. Even music blogs, like the excellent Fuel/Friends, also no longer operating, have been displaced. This time by social media.

Man vs Media vs Machine

I assumed that the prominence of recommendations built into Spotify would dominate my discovery. But to really see if that was true, I needed to dig into the data. Through the course of the year, I tracked ever single music recommendation I received — from friends and colleagues to Spotify algorithms to social media. Side note: In 2011, my data collection was all manual, with pen and a notebook. In 2017, only 8% of the data I collected was done manually. That’s the byproduct of music’s migration to software platforms.

I combed through the data to measure the efficacy of the different sources. I cross-referenced the data with my streaming history to identify additional patterns and correlations. I categorized the recommendations I received throughout the year into three groups — people I knew personally, media sources like music sites or podcasts, and algorithm-driven playlists in Spotify — to see if once type had unique success.

One of the clearest differences between the different types of sources is the rhythm in which I receive them. There is a daily onslaught of recommendations coming from media sources in large part thanks to my social media follows and habits. Recommendations coming from machines have a steadier and predictable cadence. People have an asymmetric flow of recommendations to me but their presence remains constant enough to be influential throughout the year.

Dates align with each side of the triangle, counterclockwise from corner to corner — January to December

Machine driven recommendations came with reliable consistency. But the fact that there was little friction between receiving the recommendation and listening to it (all I had to do was press play), led me to believe that Spotify’s algorithmic recommendations would be the most dominant. They did dominate in terms of quantity, but quality was a harder metric to measure. I received the most recommendations for my favorite albums of the year through media sources. By percentage, people were the most successful with 10 of the 117 recommendations given for top ten albums.

For me, recommendations are also about discovering things I’ve never heard before. There were 162 new artists that I listened to through recommendations. By quantity, more came from media and machine sources. By percentage, human sources were more successful at introducing me to things I wouldn’t have otherwise heard. The best example was Vince Staples’ album Big Fish Theory. I had no record of listening to Vince Staples prior, but Big Fish Theory became one of my favorite albums of the year after being recommended by a friend.

This Machine Makes Playlists, Playlists Make Me Listen

After years of believing that I was an album listener, in 2015 I discovered that playlist listening had an equal presence in my listening behavior. We are squarely in the era now of the playlist, thanks in large part to Spotify. I wanted to see if the ease of listening to recommendation playlists somehow gave them an advantage. Those songs were only a click away whereas other recommendations required a little more work.

134 songs were listened to the same day they were recommended. About half of those were from playlists. In looking at the recommendations I listened to versus those that went unheard, there was no noticeable discrimination between types of sources. It seems as if playlists didn’t have too much of a built in advantage after all.

Most recommendations come in at midnight (from playlists) or in the morning (from social media) and are then listened to throughout the day.

Product vs Magic

Many tout the machine learning voodoo at the core of the Spotify recommendation algorithm. The product rules appear to be just as important. They take the magic and make it useful. I was able to to infer a few product rules from analyzing the Release Radar playlist:

  • Post to the playlist at midnight, in time for the next day’s listening.
  • Keep the song on the playlist for up to 4 weeks if it hasn’t been listened to.
  • Favor artists that I have listened to before.
  • Use remixes and Spotify live recordings when new music runs thin.
  • Try all types of songs to keep the playlist fresh and exciting.
There’s really no strong pattern in the song attributes of machine-based recommendations. They will try anything.

The Missing Piece

Algorithmic recommendations in Spotify are easy to get to and pretty much in line with my current tastes. As I looked deeper into the data, I saw that they didn’t do much to get me out of my comfort zone. I realize that if the playlists were full of artists I didn’t know or didn’t normally listen to, I would probably ignore them. So it takes more than just exposure to drive discovery.

When people would recommend music, they had a chance to say why they thought I would like it. They could qualify the recommendation, which is something missing from playlists today. Those explanations also induced a type of social pressure. I felt like I had to listen and I had to pay attention so I could talk about the music later.

The Next Seven Years

Music discovery will continue to change. Adaptation is already happening. Historically influential sources of music discovery, like Pitchfork and other music blogs, have adapted to the era of playlists. They now have their own playlists that they update weekly. Evolution is beginning with the integration of Spotify in Instagram Stories. Music podcasts, the saturated music festival landscape, and smart speakers are providing additional avenues for finding new music.

As long as music is a social currency, we will want to find the best new music. If the last seven years are any indicator of how fast things can change, by the middle of next decade music discovery will look a lot different than today.


There is a 12 page book summarizing my findings and process in tracking my year in music. It is a limited edition of 200 copies with some of the charts above and more. It’s printed in a really amazing fluorescent orange ink that you have to see in person. Order a copy of the book here.

Previous year’s projects are compiled on my website: www.ericboam.com