I have recently been listening to Spotify Daily Podcasts, an algorithmically generated podcast playlist. It meets my bare minimum requirements for use, in that it has broadly reasonable choices of podcast for me to listen to and I don’t have to make decisions. However, it does not meet any of the other requirements for a podcast. Indeed, I am concerned that the designers of Daily Podcasts have in fact never listened to a podcast, never talked to anyone who listened to a podcast, and possibly live on a mountain in Tibet from which hard drives full of code are trekked by yaks.
In this post, I outline a few suggestions for improvement from a person who has listened to as many as several podcasts.
Generating The Playlist
First of all, podcasts I finished yesterday should not be on today’s playlist. Conversely, podcasts I am currently in the middle of should not randomly disappear from the playlist. This is the behavior of Daily Podcasts that makes me most worried that no one involved has ever listened to a podcast.
Contrary to the name, Daily Podcasts does not refresh every day the way that their Daily Mixes do. Instead, “the more you listen, the more frequently it updates.” Ugh. Fine. I guess.
I do think that generating a new playlist every day is not really the desired behavior. Daily Podcasts is typically seven or so hours long; most people are not going to listen to the whole playlist in one day. If you recommend a good podcast episode, people will want it to stick around on their playlist and not randomly disappear. Our first step is to rename Daily Podcasts to Your Podcasts.
Instead of refreshing every day or on an inscrutable and confusing schedule, Your Podcasts should never be shorter than seven hours long. Sometimes it will be seven hours and ten minutes long; sometimes it will be eight hours long; if you’re a fan of 80,000 Hours, it might be ten hours long. When a listener finishes a podcast, new podcasts are added to the playlist until it is the appropriate length. That way, the listener has options to choose from and can look forward to podcasts they’re going to listen to later.
Your Podcasts should also have a “remove episode” button. If you press this button, the episodes is removed from your playlist, and the algorithm updates in the direction that you dislike podcasts like this one. I should not have to repeatedly skip past the true crime podcast I do not want to listen to. I should not have to worry that if I am zoned out and not really listening so I don’t skip fast enough, then Spotify will conclude I want to listen to more true crime. I should have some way other than refusing to listen to convey that I don’t like true crime. This is not acceptable behavior for a recommended podcast playlist.
Podcast Selection
If I am following a podcast, episodes of it should appear in Your Podcasts on a fairly regular basis. Following a podcast means I want to listen to it often.
If you play me the trailer for a show you think I like, you should next play episodes of it. You shouldn’t play more trailers for different shows. Spotify’s programmers seriously overestimate how much I want to listen to podcast trailers.
Types of Podcast
There are, essentially, three types of podcast behavior you want to incorporate in your recommendation algorithm.
Evergreen podcasts are podcasts that you can listen to in any order whenever you want: for example, interview podcasts, podcasts where someone is reading you a short story, and most nonfiction podcasts that wrap up each topic in a single episode.
News podcasts are podcasts you only want to listen to shortly after it comes out. As you might guess from the name, these are mostly going to be about the news.
Series podcasts are podcasts that have an order. Most fiction and actual play podcasts are series podcasts. Some nonfiction podcasts are also ordered: for example, history podcasts that go through a topic starting at the beginning.
I don’t think there’s any way to separate the categories without having podcasters specifically indicate which category they’re in, although for already existing podcasts you could probably hire a bunch of data entry people to look for key phrases like “Part One.”
Ideally, you’d want to do it by episode rather than by series. Some news podcasts might have occasional evergreen episodes, or vice versa. Many nonfiction podcasts have multiple subseries: you can listen to the English Civil War and American Revolution subseries in any order, but you should listen to English Civil War Part One before English Civil War Part Two. You would want to allow podcasters to indicate this.
By default, the algorithm would probably favor evergreen podcasts over news podcasts or series podcasts: you can listen to any of a hundred episodes of Conversations with Tyler, but you can only listen to the next episode of the Magnus Archives. It would probably be a good idea to weight the algorithm based on how many news/series podcasts the person follows and how often they have listened to podcasts of that sort in the past.
Relistening
People often want to relisten to podcast episodes. The algorithm should track how often people relisten to a particular podcast and recommend episodes a person has already listened to, after a suitable time interval, if they are commonly relistened to by other listeners.
The algorithm should also observe which podcasts the individual relistens to and, if their behavior is unusual, weight the individual’s behavior more highly. If someone comfort-listens to old episodes of Fivethirtyeight, Your Podcasts should accommodate this preference.
End Of The Episode
If I stop listening to a podcast within ninety seconds of the end, you should assume the remainder is credits and mark it as something I’ve listened to. (If you want to improve the listening experience while somewhat increasing the burden on podcasters, you can also have them indicate when the credits start.)
Music
For reasons best known to Spotify, you can get a playlist of podcasts interspersed with music, but only if you specifically want to listen to podcasts about sports, news, music, or meditation. If you want to listen to podcasts about your actual interests, they aren’t interspersed with songs.
Your Podcasts should have a sister playlist, Your Listening, which is half podcasts and half music.
I'm not familiar with how Spotify works with podcasts, but there is a fourth type of episode, which are meta episodes that you probably never want to hear in a daily shuffle. For instance, I listen to a table-top role-playing game podcast, which has a meta episode to tell new listeners which season they might want to start out with depending on their interests.