From June 2020 to June 2021, YouTube paid more than $4 billion to the music industry, the company announced this month — a much-increased sum from the world’s biggest video platform where product reviews, how-to videos, and vlogs share the spotlight with music. Rights holders have historically complained that YouTube hasn’t been paying enough for its offering, but the 2020 payout was a notable jump from 2019, when YouTube paid over $3 billion to music rights holders, thanks in part to its powerful algorithms that surfaces the music each user desires.
Unlike other streaming services, YouTube Music relies entirely on algorithms to program its over 10,000 playlists; so instead of staff members hand-selecting songs for listeners, the company is putting its editorial focus on better engineering those algorithms with parameters that are constantly developing to suit listeners’ needs. (YouTube’s Music and Premium services have more than 2 billion monthly active users and 30 million subscribers, respectively.) That “algotorial” approach — a combination of “algorithm” and “editorial” — has allowed it to maintain its position as the world’s most-used music platform, despite increasing competition from Spotify, Apple and Amazon.
These algorithms run on the “signals” music fans are sending the service when they’re listening to music. These can range from obvious actions like liking or disliking a track to skipping it to passive signals like adding a song to their library or playlist and listening time. The algorithm ingests this information and adds it up to formulate the individualized options for each user — preferred playlists, related music, and albums they may like. That basic process is standard across all music platforms, but the specifics of their formulas can result in big differences when it comes to how users engage with the various platforms — and how much revenue they yield for the music business. Now, for the first time, YouTube popping the hood on its secret recipe. Here, YouTube’s executives reveal their “algatorial” curation method exclusively to Billboard.
Here’s how it works:
The raw signals are the ones we know — thumbs up or thumbs down on YouTube — but all signals aren’t created equal on the platform. For example, skipping a song isn’t weighed as heavily as one might think, while other passive signals like saving a song to your library, repeating a song and how long you listen to each song are key factors helping the algorithm craft to what YouTube Music calls your “taste profile.”
“A taste profile is made up of the listener’s favorite artists, the artists that they most recently engaged with, and it goes all the way down into the long tail of artists that they’ve engaged with,” says T. Jay Fowler, head of music at YouTube.
Just like every music streaming service, YouTube Music begins to build its taste profile when users they initially sign up for the service and go through a quiz (referred to as a “taste builder”) that lets users pick artists they like. But this is a bit different for YouTube, where users watch over 1 billion hours of video every day, searching for everything from music videos to home improvement instructions. “Most users are not new to YouTube,” says Gregor Dodson, YouTube Music’s product manager. “We’re often seeing Rick Astley there, or Gangnam style,” Dodson says of users new to the Music app, noting that the company doesn’t give too much consideration to your consumption of music videos, which may not translate to what you want to listen to on YouTube Music.
Fowler says YouTube is also cautious with how it weighs the artists you pick in its taste builder, noting that people can get stressed out by the onboarding process. “They’re like, ‘If I click this and I say that I like Billy Joel, because I like that one song or Fleetwood Mac, because I like that “Dream” song, is the service now just going to think I’m a massive Fleetwood Mac fan and just dominate all of the recommendations they give to us?’” Fowler says. “So given that, we take it as a light touch and we really take more stock in their behavior once we start to package and recommend content for them, whether it be a mix, or a playlist, or individual song recommendations.”
In fact, YouTube says it never wants to assume it has listeners completely figured out, acknowledging, for example, that users who hear songs on the radio and search for them on YouTube may not want to explore any more of those artists’ catalogs.
“What you don’t want is, that experience where you select Fleetwood Mac, and the very next screen is, here’s a bunch of shelves of Fleetwood Mac records, mixes playlist, where they’re featured and you’re like, ‘This isn’t actually valuable, this is just it’s spinning what we told it back,’” Fowler says. “The idea is, couple that explicit input, that explicit selection of artists, as just a raw guide. And then continue to encourage people to step out of that concentric circle.”
Downplaying Skipping Songs
Skipping a song also doesn’t hold much weight for YouTube Music, which executives say can be hard to decipher, as the motivation behind the skip may have little to do with a user disliking the song.
“One behavior that we talk about a lot is rage skipping, which is where you see someone just skip, skip, skip, skip, skip,” Dodson says. “So you think, ‘Oh, they’re annoyed. They’re clearly unhappy if they’re doing that.’ Well, turns out, in talking to users, sometimes people do that because they want to hear what’s on the playlist. If you were to take that as a negative signal, you might actually end up delivering something or believing something that was actually wrong.”
Dodson says skips are expected in things like YouTube’s discovery mix, which is designed to get people to listen to songs they aren’t familiar with. “A skip could be not now, or not ever.” Fowler says. “There is a general rule of thumb that you can take with a skip, which is, if you assume that it’s a not now, you can just delay the frequency at which you resurface it. We want to be relatively light touch with the not-ever, because as a music fan these new songs take a while to get to know and take a while to become your new favorite. So we never take a skip as being an explicit, don’t play this ever again.”
Managing Unexpected Events and Genre Lines With a Human Touch
“It’s really a mix of human and machine because even the algorithms are written by music fans,” says Doug Ford, director of music and product programming at YouTube, noting that human interaction is still crucial to building a successful algorithm for music. “An algorithm can’t cold start a track or a new artist or a trend, so we’re there along with other music fans and music experts in the industry, getting the information from partners, from artists, from management, from signals we’re seeing, from search signals that are coming from our parent company and from YouTube consumption data.”
Fowler points to users searching for viral hits from TV shows and the borders between different genres — the lines of which get blurrier with every passing year — as spaces where algorithms need additional guidance from the YouTube Music team. Unconventional super groups will also throw a wrench into an algorithms success rate, Fowler says, as Labyrinth, Sia and Diplo’s group LSD did when it released its debut album in April 2019. “Those three artists have incredibly sizeable followings on YouTube, but the algorithm, when it saw that content come in, it didn’t know what to do with it,” says Fowler. So, contextual awareness around the groups following needed to be added by hand. “Often you need to have a human that can actually help provide input,” he says. “Let’s actually provide context to the system that’s important.”
Genres can provide a challenge to YouTube’s algorithm as well, and the company’s solution to the problem has been the cavalcade of sub-genres you’ll see in year-end wrap ups with names like “pop rap” or “deep pop emo.” “We use humans to train things like classifiers by creating groups of like music,” says Fowler. “Genre is incredibly subjective, and things that are labeled as Brazilian music necessarily aren’t palatable if you can lump it all together. There’s Brazilian funk, and there’s classic Brazilian music. We think of it as human and machine, the marriage of it, but we allow the algorithms to speak to the user once we’ve given the algorithms the full context.”
To determine a playlist’s success, Fowler says, “The first metric we look at is the number of impressions and the number of click-throughs. Did we put the right playlist in front of you that drew your attention?” That analysis comes down to taste profile (is it the right playlist for the listener), if the user is able to understand what the playlist offers through the artwork and title, and how the user consumes the playlist.
“Once you click through, we look at what we call super consumption,” Fowler continues. “Did you spend time with it, and have a long listening session? Skips are fine. We don’t penalize you for skips, but did you listen to a bunch of songs and are you getting value out of it?”
YouTube also considers “tertiary metrics,” including giving a thumbs up to songs, and repeated listening or returning to the playlist, which YouTube calls re-consumption. “That’s how we really know that we’ve got a good property,” Fowler says.
There are still improvements the YouTube Music team would like to make, most notably to develop their algorithm to better understand and predict how a user’s taste in music changes over time.
“We will probably struggle with this to the end of time, especially as this next generation is truly genre-spanning, truly era-spanning,” says Fowler. “They have their TikTok songs that they know all the words to, and then they’ve got the real songs that they like, and there’s going to be an intersection of those two. It’s going to be an ongoing, evolving problem, but it’s a fun one.”