I’ve long been interested in Spotify’s podcast recommendation system, mostly because it uses a different strategy than its competitors.

Unlike Apple Podcasts, which recommends shows you might like:

Spotify recommends episodes you might like:

This approach – episode-to-episode recommendations rather than show-to-show recommendations – makes sense given Spotify’s roots in music. Episodes are to podcasts what songs are to albums.

So if the atomic unit is the episode, and if gateway episodes are a thing, then Spotify’s episode-to-episode approach has the potential to yield more useful cross-pollination of audiences than show-to-show recommendations.

Why does this matter? Because in-app podcast discovery matters. According to the most recent Canadian Podcast Listener report, in-app discovery is on the rise, with 5% of monthly podcast listeners first discovering a new show when “it was recommended to me based on other podcasts I listen to.”

I wanted to better understand Spotify’s recommendation engine. So I decided to draw a map.

(Larger version, interactive version)

I started with Spotify’s podcast charts. In the US, Spotify maintains a “Top Podcasts” list of 100 shows (more are available via API), plus 17 category-specific charts with 50 shows apiece. All together, accounting for duplication across charts, I found 890 distinct shows on Spotify’s US podcast charts in mid-August 2022.

But of course, Spotify doesn’t recommend shows. They recommend episodes of shows. So for each of the 890 charting shows I found, I made a list of episodes. All together, the 890 shows on Spotify’s charts had a total of 223,554 episodes between them.

Then, for each episode in my list, I checked for a “You might also like” section. If one existed, I looked at the first 6 recommended episodes, matching the number of recommendations that appear in the Spotify desktop experience.

Then, recursion. I took all the recommended episodes and used them to look for morerecommendations. And so on, and so on, and so on… until I couldn’t find any more recommended episodes.

Finally, I discarded episodes without recommendations, leaving me with 90,235 episodes and 279,818 episode-to-episode recommendations connecting them.

Once I’d collected the data, I drew a map. Each dot represents an episode, and each line represents an episode-to-episode recommendation.

But things got really interesting once I started to analyze the data.

14% of recommended episodes are Spotify originals and exclusives

Spotify has made a number of content-focused acquisitions in recent years, including Gimlet, The Ringer, and Parcast. Many shows from these companies have been branded as Spotify Originals and Exclusives. It’s easy to identify these shows through Spotify’s API.

Of the 90,235 recommended episodes I found:

7.98% are Spotify Originals (e.g. The Journal, Today in True Crime, Haunted Places)
4.14% are Spotify Exclusives (e.g. Call Her Daddy, Armchair Expert with Dax Shepard, Unlocking Us with Brené Brown)
1.45% are Spotify Adaptations (e.g. Asesinos Seriales, Crímenes Pasionales, Misterios Inexplicables)

Together, episodes from these shows represent 14% of all the recommended episodes I could find in Spotify’s system. Looking just at Spotify Exclusives and Originals, I found 12,253 episodes connected by 17,497 recommendations:

Perhaps unsurprisingly, Spotify’s owned and operated shows seem to be connected through a robust network of recommendations, and their system seems to do a good job of surfacing their own original and exclusive content.

Spotify’s algorithm wants me to relax

With a dataset of 279,818 recommendations, but only 90,235 episodes, it stands to reason that some podcast episodes are more frequently recommended than others. So I sorted my list of episodes by degree, and was surprised by what I found at the top of the list.

Among the top 25 most frequently recommended episodes:

Spotify’s recommendation system may be a contributing factor to the success of so-called “white noise podcasters,” a trend identified by Ashley Carman in June 2022.

Of course, Spotify has plenty of white noise audio available outside the Podcasts section. They have an entire White Noise genre that contains Spotify-curated playlists like Sleep Noise, and Baby Sleep Aid: White Noise.

Given the royalty fees associated with licensing music, Spotify may have a financial incentive for listeners to use “podcast white noise” instead of “music white noise.” Buy why might users choose podcasts? This episode description sells one benefit:

DID YOU KNOW: By using a PODCAST instead of a white noise SONG on Spotify to go to sleep, it will not affect your listening history which means your Spotify Wrapped won’t be bombarded with white noise and instead contain the songs you actually listen to?

The separation of music from podcasts in Spotify has some really interesting knock-on effects.

Where’s Rogan?

One of the most surprising things I found in my dataset was what didn’t appear: episodes of The Joe Rogan Experience.

Rogan’s show is very regularly at the top of Spotify’s own “Top Podcasts” chart, and his episodes routinely appear on their “Top Episodes” list.

Though I did find a few Joe Rogan guest appearances in my list of episode recommendations – including appearances on This Past Weekend and Hotboxin With Mike Tyson – I couldn’t find any evidence that Spotify is displaying episodes of The Joe Rogan Experience in their “You might also like” sections.

Perhaps Spotify has decided against using their recommendation algorithm to actively promote Joe Rogan. Or perhaps his show’s gravity is so strong he doesn’t need support from a recommendation engine.

Episodes of a feather flock together

I applied a community detection algorithm to group recommended episodes, and found that they tend to cluster based on a few attributes including language.

Here, we can see clearly defined groups of Spanish-language episodes (pink), German-language episodes (blue), and Portuguese-language episodes (orange):

Explore for yourself

If you’ve read this far, I suspect you might like to explore the dataset for yourself.

Good news: In addition to the static images in this post, you can also explore an interactive version of my Spotify recommendation map.

Launch interactive map

NB: To keep file sizes managable, the interactive version contains a subset of my overall dataset, prioritized by degree. You’ll be able to explore 51,835 episodes and 186,721 connections between them.

Big thanks to the team behind Gephi, and to the Oxford Internet Institute for their very useful sigma.js export plugin for the tools I used to make this happen.

Remember

Spotify recommends episodes, not shows
About 14% of the episodes Spotify recommends are its own exclusives and originals
Many of the most frequently-recommended podcast episodes in Spotify don’t include human voices at all
Recommendation algorithms matter because in-app discovery matters. Research suggests an increasing number of podcast listeners find new things to listen to through in-app recommendations.

If, like me, you’re attending Podcast Movement 2022 in Dallas and enjoy podcast-related data visualization, I hope you’ll check out my session Podcast Neighborhoods: How to Find Niches and Reach Audiences You Didn’t Even Know Existed.

I made a map of Spotify podcast recommendations. Here's what I learned.