Hey Spotify, Play {Artist}

The Challenge

Artists were the second most frequently requested entity on Spotify’s voice platform. The intuitive solution was to play the artist’s “This Is” playlist, which collects a variety of tracks from their discography. I suspected we could meaningfully increase engagement and satisfaction by building an experience that better understood what a user was actually asking for.

My Role

I originated this project. While digging through utterance data, I noticed that people were asking for artists frequently but abandoning those sessions more often than you’d expect. My hypothesis was that we’d done what was easy instead of what was right. I rapidly designed a solution, pitched it to the PM and engineering team, and from there worked alongside ML and front-end engineers to build and test an MVP.

User Needs

When someone asks for an artist like “play Coldplay,” they’re trying to satisfy three core needs:

Hear their favorite songs by the artist (or, if they’re unfamiliar, hear iconic tracks)
Discover songs they might like from the artist
Stay aware of new releases from the artist

A “This Is” playlist can satisfy some of these needs, but it does so by handing the user a tracklist to browse for the specific song they want. “This Is” playlists aren’t personalized, and they’re biased toward an artist’s iconic songs rather than any individual user’s favorites.

This is Coldplay Spotify Playlist — "This Is" playlists provide a solid overview of an artist, but aren't personalized.

Voice is also fundamentally different from browsing. People don’t always have access to a screen — they might be on Google Home, Alexa, or using the voice feature in the Spotify app. When speaking, people don’t always know what to say, and artist names are easier to recall than specific album or track titles.

Sequencing Design

When designing algorithmic UX, it’s important to be clear about what the algorithm is optimizing for, and what heuristics or weights should tip the scale toward a particular outcome while still giving the model freedom to do the right thing for an individual user.

For this project, I wanted the algorithm to do a few specific things:

Optimize for the user’s favorite songs by the artist, but avoid staleness by mixing it up
Interleave unfamiliar tracks we believed the user might like alongside their favorites
When there’s a relevant new release, open the session with a popular song from it
Wrap the set by shuffling the artist’s catalog, biasing toward familiarity earlier in the set

Play Artist Sequence — An example of the sequence in action.

Fallback

Voice introduces ambiguity. Homonyms — words that share spelling or pronunciation but refer to different entities — combined with the looseness of spoken requests mean we won’t always get it right. Someone might say “play Coldplay” because that’s all they could remember, when what they actually wanted was the album Ghost Stories.

To handle this, I designed a companion experience that let people pivot to a different entity. Unlike Spotify search, the underlying algorithm was optimized for voice: it didn’t anticipate future keystrokes, it sought to resolve ambiguity, and it would switch playback in place rather than navigating users to an entity screen — so they could sample results by listening to alternatives instead of visually browsing them.

I considered routing to the entity page, but that collapsed the alternatives and defeated the point. The goal was to keep voice sessions in voice.

Voice fallback screen showing alternative results — The fallback screen reinforces what we thought someone wanted while showing other reasonable results, letting them easily pivot.

Dialog Framework

Beyond playing the right song, I wanted the experience to show that we knew the user, using text-to-speech. This was only possible inside the Spotify app, since third-party voice platforms control what their assistants say. Depending on how the user related to the artist and what we’d decided to play, the session was introduced differently:

Scenario	Utterance	TTS
Artist has a new release	Play Taylor Swift	“This is Taylor Swift, starting with Willow, a popular track off her latest album.”
Familiar artist, no new release	Play Taylor Swift	“This is Taylor Swift, starting with Shake It Off.”
Artist is unfamiliar	Play Taylor Swift	“This is Taylor Swift, starting with one of her most popular tracks, Blank Space.”

Results

This update was a meaningful improvement over control:

+10pp session success rate — people were measurably more likely to listen through rather than abandon the session.
+15% session length — sessions grew by roughly one additional song per session on average.
Qualitative satisfaction lift — not something you can measure quantitatively, but in research sessions people smiled when their favorite songs played, where previously their expressions had been neutral.

The artist-by-artist analysis also surfaced specific cases where “This Is” playlists were underperforming, providing actionable data to improve curation across the platform.