How Spotify’s recommendations UX might be pushing down per-stream royalty rates — even if total payouts increase

Streaming services are often discussed in a way that keeps their components — royalty rates, recommendation algorithms, new features, internal editorial policies and organizational processes, etc. — separated from each other. In reality, these aspects are all intertwined.

In this article in particular, I want to point out that better performance on metrics that users and product teams care about, means poorer performance on one of the metrics musicians care most about: The per-stream royalty rate.

I will approach this from the user perspective, because Spotify’s history is rooted in competing with the piracy and P2P file-sharing that was rampant in the first decade of the 2000s. It is in this environment that the Swedish company had to figure out how to build a service that gets people to pay for something they could also access for free elsewhere. Even if today’s environment differs a lot from those days, the company’s value proposition to the end user has remained unchanged.

Spotify’s most succinct description of their value proposition is listed in the About section of their website: “With Spotify, it’s easy to find the right music or podcast for every moment – on your phone, your computer, your tablet and more.” Let’s break that down, for now ignoring the podcasts part:

• Easy to find: This means a well-structured and complete catalogue that can be navigated in a variety of ways.
• The right music for every moment: This means the service has to deal with scenarios where people know exactly what they want to listen to, have no idea at all or fall somewhere in between those extremes.

On a surface level, one way to know that you’re succeeding at your value proposition is through growing subscriber numbers — but this can be influenced by things like discounts, promotions, ad campaigns and screen redesigns that increase conversion rates. So how do you know your product is actually doing well?

You might look at:

• Active users: The typical metric is monthly active users (MAU). This indicates growth and engagement, but obviously listeners using Spotify just once a month is not a good sign. What proportion of users are active daily and weekly, versus just monthly?
• Churn & retention: Churn refers to the proportion of users that drop off over time — for example, people who were an “MAU” but are now no longer active or have cancelled their subscription. The inverse is retention: How well do users stick to the service? It can range from “day 1” retention — “of people who tried the app on day 0, how many return the next day (day 1)?” — to much longer-term definitions.
• Listening hours & session length: How many hours per week do users listen to music? When a user starts listening, how long does their session last? These types of questions help product teams understand how much value users are getting out of the service.
• Sessions without playback: For a service like Spotify, which measures itself by making it easy for users to find music for every moment, it could be important to understand occurrences where people open the app, don’t find any music for the moment and leave the app without any playback. Session length could tell a similar story.

All of these can be broken down into more specific metrics, depending on what you’re interested in. The important thing to understand is that these metrics are derived from questions that are informed by the company’s value proposition, goals and understanding of their users.

For instance, you might describe the user’s problem and ask a “how might we” question to brainstorm solutions, like: “How might we make sure someone can quickly and easily play a specific album?” Such a problem would probably be solved by looking at the journey from opening the app, to using search, to actually playing the album. It could impact a metric such as session length or ratio of sessions without playback.

Now let’s look at another problem that’s more relevant to our discussion today:

“When a music fan doesn’t know what to listen to, how might we always offer something interesting to explore?”

It must have been a question like this that started off Spotify’s journey into becoming one of the world’s most-used music recommendation engines.

2013: Discover — Spotify’s early forays into personalization

One of Spotify’s earliest forays into recommendations was its Radio feature, which couldn’t compete with the more established music recommendation engines of Last.fm and Pandora until the company introduced improvements in 2011. Its Discover feed then launched in 2013. Whereas personalized radio stations offered departure points for a more passive form of music exploration by letting users generate stations from every piece of content on the platform, Discover brought actual personalization to the service as a whole.

One can assume that the metrics Spotify was trying to improve at the time focused on engagement: How many days of the week do users use the service? What’s the average session length for users?

It wasn’t until a few years later that Spotify’s algorithmic recommendation strategy really came into its own.

In early 2014, Spotify acquired The Echo Nest, a company specialized in creating and making sense of data about and around music. Whereas Spotify’s current mission is stated as giving a million artists the opportunity to live off their art, back then the focus was absolutely on the listener. Addressing the Echo Nest acquisition, Spotify’s founder & CEO Daniel Ek said at the time (emphasis added):

“At Spotify, we want to get people to listen to more music. We are hyper focused on creating the best user experience and it starts with building the best music intelligence platform on the planet.“

2015: Discover Weekly

A year later, in July 2015, Spotify launched its now-famous Discover Weekly playlist. To do so, the company leveraged a project from The Echo Nest called Taste Profiles, which aimed to represent a nuanced understanding of a user’s music taste and got integrated into Spotify post-acquisition. The result: More data for the algorithms to use, and thus even better recommendations allowing the service to cater well to long tail tastes.

Immediately, Discover Weekly presented a clear user experience: Every Monday, you’ll get a fresh selection of 30 tracks personalized for you. Some may be familiar, while others may be new gems. The playlist gave people something to dig through and new ways to explore Spotify’s catalog by clicking through to artist profiles and albums. Besides introducing a new playlist into users’ libraries, it also created a ritual, with users checking in on Mondays to review the new music and on Sundays to save songs before Discover Weekly would get refreshed. Whether the songs were just listened to on repeat or were saved to users’ libraries and playlists, all of these actions would give Spotify’s algorithms a chance to improve users’ Taste Profiles and come up with even better recommendations for future playlists.

Discover Weekly proved to be a hit among both users and the media. For Business Insider, Alex Heath praised: “Like magic, it just knows what I want to hear.” Some people even said they’d marry the playlist if it were to propose to them.

And as far as listening diversity went, Discover Weekly paid off for Spotify. In 2017, the company announced that since 2014, the number of artists whom users listen to on average had gone up by 37% year-over-year. More crucially for the company’s performance, personalized and editorial playlists had contributed to a 25% increase in listening hours. These were good signals for Spotify, as active use of a service usually indicates a lower likeliness of cancellations and a higher likeliness of conversion from free to premium.

2020: How Spotify’s algorithms get users to listen more today

Since the early success of Discover Weekly, Spotify has rolled out many more recommendation products. They can be roughly divided into four user behavior categories:

• High personalization, high intent: The user is actively engaged, and has a good idea of what they want to do by interacting with a feature with algorithmic recommendations.
• High personalization, low intent: Similar to the above, but in scenarios where users are much less likely to have any specific goal in mind, e.g. “just play me something I like.”
• No personalization, high intent: Higher engagement behavior in the service, such as browsing the catalog where the features are algorithmic recommendations, but not personalized.
• No personalization, low intent: Similar, but again more of a “just play me something I like” scenario.

Bear in mind that certain features can be used in many ways. For instance, Spotify’s Radio functionality typically caters well to “just play me something that sounds like X,” but can also be used to actively dig through similar music to add to one’s own playlists.

Normally, you would have a collection of problems you want to solve for users and then map your features alongside them. Since I don’t want to speculate too much, I’ve simplified it to the below table.

I am going to highlight a few of these to point out how some of these features specifically play into Spotify’s strategy:

Daily Mix

A clear example of Spotify incentivizing users to open the app more often — and not to leave without listening — are the service’s Daily Mixes. Launched in 2016, the feature was explicitly “designed to be the shortest path to a good musical experience.” The daily playlists contain a higher proportion of familiar music than Discover Weekly, which mitigates the risks involved with recommending new music. It provides a more passive way for users to tune into their taste, without having to find specific playlists or artists. In fact, Daily Mixes were referred to internally as personal radio prior to their launch.

The playlists apparently performed so well that Spotify integrated them into the core experience of the Free tier in 2018. While Daily Mixes address the value proposition of easily finding music for every moment well, how do you make sure that moment doesn’t pass?

Autoplay

Spotify’s automatic radio feature Autoplay is the most straightforward example of the company’s efforts to increase users’ session lengths. When you finish listening to a song, release or playlist, Spotify automatically continues your session by loading up a radio station based on what you were just playing. Launched quietly — likely some time in 2017, since that’s when the public complaints first emerged from users who wanted to disable it — the Autoplay feature reduces the need to actively engage with the service. If the algorithms do their job well, a user will not have to come back to the app to find more music to play until they want to listen to something else.

Spotify recently announced that artists will be able to use a feature called Discovery Mode to promote their music in Autoplay and Radio in exchange for a lower royalty rate for the promoted tracks. While plenty has been said and written about the artist’s side of the equation, it will be interesting to see how user engagement will be influenced by letting paid promotion play a part in the recommendation engine’s mechanisms for these features.

Your Daily Drive

Your Daily Drive, which launched in mid-2019, is a hybrid playlist that intersperses music with bits of news, specifically tailored to compete with radio during people’s morning commutes. Although Spotify announced that the music would include “tracks you’ve yet to discover,” it seems to be primarily composed of music users have already heard, which makes sense for a low-intent “hit play and put the phone away” type of scenario. Being able to compete with traditional radio is important for Spotify to get people to retain their subscriptions or even consider one.

Services like these don’t compete in the same way as traditional consumer products like detergents do: More than just claiming market share, online services compete to monopolize certain types of user behavior. With a detergent, you have the moment of a purchase decision, but with services it’s more about the behavior that you have to route through your product over time. Lose the engagement, and you lose the competition for that behavior. If you’re successful, people may start calling this type of behavior by your name, like Googling information or Ubering to a party. So Spotify and its competitors have to identify and cater to behaviors around music systematically, rather than just compete on isolated product features.

The order of Spotify’s search results

If you’re up for an interactive element: Here it is. Otherwise you’ll just have to take my word for it, or see the Twitter replies from when I researched search personalization a short while back.

If you’re a Spotify user, open the app, go to search, type yung as I did, and check out the artist results. Unless you’re in Germany, it’s highly unlikely you’ll see the same order of results as in the included screenshot above.

Why? Let’s bring back the service’s value proposition: “With Spotify, it’s easy to find the right music for the right moment.” What might be the right music for someone in one place, may not be right for another person in another context. With over 60 million tracks in its catalog and 40,000 tracks being added daily, Spotify’s search applies some type of ranking mechanism to serve you the right content for your distinct needs and tastes. This means users can get to play music they want to hear faster, there are less sessions without any playback and sessions in general are longer. All of that implies higher user retention.

That’s what Spotify wants, that’s what their investors want, and they do that by addressing what users want. But what do artists and other rights holders want?

Longer sessions means lower royalties per stream

Currently, Spotify’s payment model works by pooling together the revenue generated by users’ paid and ad-supported streams, and distributing that pie on a pro rata basis, i.e. according to an artist’s proportion of stream counts. It gets more complex when you get into the details, but the simple version is enough to know for the sake of the next paragraphs.

When users in general have longer sessions and listen to a wider variety of music, that royalty pool will have to be divided among more tracks. This means the total amount of money paid out remains the same (or may be larger in the case of ad-supported streams), but the average payout rate for each individual stream would be lower.

A user-centric model doesn’t change this, because the service is still designed to boost the average number of streams per user by trying to boost users’ session lengths. As a matter of fact, the design of these features may be at odds with user-centricity as a better payment model, since they encourage listening diversity. Diversity means playback gets spread out over a lot of artists, rather than being concentrated on a few.

So if a user contributes $7 of revenue per month, and they listen to an artist radio station rather than playing solely through the artist’s catalog, that artist will see a much smaller portion of that money. (To clarify semantics, in this case the term “diversity” refers to statistical diversity rather than demographic diversity — so it could be someone listening to two white, middle class EDM artists, instead of just one white, middle class EDM artist.)

Does that mean that services like Spotify should make listening diversity less of a priority, for the sake of paying more to artists who drive playback? Well, it depends.

We all want artists to make more money, so we should consider how a service that depends on end-users to generate that money would accomplish higher payouts at the same time. Whether the payout model is pooled or user-centric, higher per-stream royalty rates for artists would require either less consumption, or more revenue per user.

The former option seems unrealistic and undesirable on all sides. If users were to spend less time listening to music, would they still be willing to pay for Spotify, Apple Music, Amazon Music or any other service? A service with fewer paying users will have less money to pay out, no matter how it’s distributed. On the flip side, one may find that higher listening diversity goes hand-in-hand with retention, which means less churn and more money in the pool to distribute. Besides, the degree of listening diversity may not be as important of a factor with the streaming service’s current pooling method of distributing user revenue.

Merely increasing royalty rates alone might not help, either. Expecting a penny per stream is highly prescriptive for a service’s user experience. The current levers for increasing a stream’s average royalty rate are the number of total streams and the amount of money individual users bring in. If Spotify can’t charge its users significantly more, making a penny-per-stream model work depends on people spending less time listening to music rather than more. This would run counter to how Spotify has been executing on their value proposition and undo much of the impact of the features described in this article, possibly increasing subscriber churn.

This is why the second option — more revenue per user — might be more feasible in the short term. Spotify is already experimenting with price increases in select markets. Perhaps there could also be spin-off products powered by their recommendations, for which users can purchase add-on subscriptions to further grow the pie (Spotify is already researching this approach for podcasts, per some reports).

Spotify’s current mission is to have a million artists make a living off of the platform, and it may have to rethink its economics to accomplish that. While the answer to that challenge is complex, what’s certain is this: If we’re going to discuss growing pies, we need to include the oven they’re baked in, and that’s the user experience.