One of the easiest ways Spotify lets its users discover new music is the "Fans Also Like" feature. Found on artist's profiles, it's stocked with a list of related artists surfaced by a metric called "artist similarity."

"Artist similarity is probably the second-most important piece of data we extract from listening patterns—after popularity," says Glenn McDonald, Spotify's data alchemist. "It's also the data behind radio, genres, and several of the features on Home and the Discover pages."

Unlike Spotify's editorial playlists, which are curated by humans, Fans Also Like is populated algorithmically. "Which is a fancy way of saying it's done with math,” explains McDonald. "But the goal of the math is to collect and make sense of aggregate human input. The computers don't have opinions—they're just organizing the collective opinions of people."

Using patterns to gather insights

When Spotify's back end is populating Fans Also Like, it relies primarily on two factors. The first one is shared fans—particularly among artists who might exist on the fringes. "The more fans two artists have in common, and the larger the share of each artist's total fans those shared fans represent, the more similar we consider them," says McDonald. "If an aspiring band has 10,000 fans, and 6,000 of them are also fans of another band that also has 10,000 fans, that's a pretty good sign that the two bands are probably similar. Whereas either of those bands might easily have 1,000 fans who also like Ariana Grande, but Ariana has tens of millions of fans, so those 1,000 shared fans aren't a significant share of hers."

The second one involves shared descriptions—using not just the content on Spotify but also on blogs, in magazines, and at other places where music is being discussed. "We analyze web pages about music every day, and combine [that information] with various other sources of artist descriptions, and then look for patterns in shared descriptive vocabulary," says McDonald.

The more specific a descriptor is, the higher the probability it’ll result in a match between artists who share it. "If two artists are both frequently described as 'pop,' that's not that interesting, because lots of artists are described as 'pop,’" McDonald notes. "If two artists are both frequently described as 'polka metal,' that's way more unusual and thus a much better sign that the two bands might be similar."

In many cases, shared fans are the factor that carries more weight. But new artists who might be enjoying a lot of press before a debut release also get a boost from shared descriptions.

How to troubleshoot

Pairing these two pattern-matching algorithms means that for the most part, Fans Also Like offers a solid way to identify artists that will appeal to similar audiences. But there are two ways things can go wrong, McDonald points out.

"If your band's Spotify page has stuff from other bands with the same name, then your fans are getting confused with theirs, and everything we try to deduce from their patterns is probably unreliable," explains McDonald, adding, "so be sure to let us know if this ever happens to you.” He notes that making sure other artists’ work isn’t being mistaken for yours (and vice versa) is important for lots of reasons—it also ensures that your fans get your new releases through Release Radar, that you don't show up on mismatched radio stations, and so on.

Descriptive pattern matching can also introduce errors—and newer artists who don't have a lot of press or who do have common words in their names should take time to look at their Fans Also Like list and make sure it checks out.

“If your band's name is a very common word or phrase, and you don't have very much material yet, the programs may sometimes get this wrong, and thus mistakenly think an unrelated article is talking about you,” McDonald says. “If you're a new or not-yet-famous band and your Fans Also Like list seems very weird to you, especially if it's full of bands with similar names but totally different music, then this might be why."

-Maura Johnston