Why Your Sleep Playlist Shouldn’t Define You
Playlists serve many masters. There’s a reason beyond pure enjoyment you hit play: lullabies for a toddler at 7pm, focus beats during a work sprint, or white noise for the nightly wind-down. Yet for recommendation algorithms, all those plays look identical. A stream is a signal of taste, and your sleepy ambient marathon ends up competing with—and often beating—the rock and funk you actually want to discover.
Spotify recognizes this functional listening problem. For years, recommendations were built on streams as "representative" signals, and three user behaviors broke that assumption: playing audio for tasks (sleep, study, workouts), playing music for others (kids, parties), and sharing accounts with different listeners.
In short, these are plays that reflect a function, not an identity — yet they were confusing every recommendation surface.
Redefining "Taste"
Your "taste profile" is Spotify's behind-the-curtain interpretation of your listening behavior. It's not a file or a report; it's the collective signal powering recommended sets and features like Blend. Every stream nudges that interpretation in one direction. For users with heavy functional listening, streams skewed the profile so much that other music nearly vanished from the rotation.
This wasn't a niche complaint. An internal audit revealed millions of users were consuming functional music, and that group engaged with recommendations less than listeners with purely standard listening habits. That gap translated to a poorer experience for a large cohort — a problem worth solving.
Engineering a "Simple" Button
The product answer is a menu toggle: Exclude from your taste profile. The technical reality is more intricate. Spotify's recommendation stack isn't a single AI but a web of models handling recent listens, top tracks, user embeddings (a vector representing your position in the musical universe), and other interactions.
To make the toggle tasteful globally, Spotify applied the exclusion simultaneously across several core profile systems: recent-listening history, top-stat determination, and interaction pipelines. Applying filters inside those systems means downstream surfaces such as Home or Discover Weekly need no proprietary changes—they simply query a refined profile.
However, not every view wants a filtered peek. For Home Shortcuts and Recently Played (where accurate past playback data informs quick resume behavior), Spotify preserved an unfiltered lookup for those teams. The answer for each surface came down to a single question vis-à-vis discovery versus retrieval: Are you recommending new music or showing the listener their actual history?
Why Only Playlists?
Playlists were the scapegoat and the savior. Initial research tagged playlists as the primary vehicle for functional listening. Given limited resources, Spotify narrowed the field to a single entity type rather than building metadata-assumptions for tracks, artists, or podcasts. Sound engineering grounded in harsh reality: master playlists—and everyone else later.
Three key decisions emerged during development:
- Play-level exclusion: Any stream originating from an excluded playlist is quarantined. But identical tracks played elsewhere (e.g., your "Sing Along" vs. a specific artist page) still count toward a taste profile.
- Retroactive reach: Exclusion skims an echo of past listens, not just future ones. A year of nightly rain-sounds would linger for months without this. Acting retroactively promised faster recommendation corrections.
- Explicit, intuitive semantics: Spotify errs toward the literal. Excluding a playlist recommends similar music; it doesn’t auto-assume you dislike the entire genre or artist from just one action. This signals clean intent to the machines that produce your profile.
Measured Improvements, Millions of Users
The launch followed a structured A/B test targeting roughly a hundred editorial playlists (e.g., Peaceful Piano, Deep Sleep) where functional listening is common. Since the feature wasn’t fully integrated during the test phase, the grading was rolled out only to user embeddings, the profile embeddings influencing "On Repeat". The experiment was a standout: users who adopted Excluding a playlist started consuming more recommended music and increased overall streaming activity.
Post-A/B, the feature rolled globally. Spotify supported UI translation across 88 languages for the feature. Most intriguingly, the product dashboard shows that even a prompt or controlled trial induces learning: about 12% of test users clicked "Open Menu" to explore the feature but declined to use it—nonetheless expressing curiosity about personalized outputs.
The shipping did not come without measured outcomes.
Roughly 20 million active users routinely exclude a playlist from their profile. Analyses found that these users listen, on average, 4% more music from recommendations and consume nearly 10% more music overall compared to a baseline before using the feature—by no means a negligible product uplift.
Launched initially to remedy accidental genre bubbles, the feature largely hit its goal. So, if YouTube suggests one more "Baby Shark" encore tomorrow afternoon on your child’s watch, you own one more way to seal that sound off from your own musical soul.
Shipping a Feature That Started as a Side Project
Getting a feature from idea to launch when the resources aren’t already assigned to you is a coordination problem as much as an engineering one. The playbook we used for Exclude from Your Taste Profile works in any organization where you have to build support before you can build software.
- Prepare to wear multiple hats. Expect to cover roles outside your job description and put in extra hours.
- Do an offline analysis first. Use existing data to estimate impact and identify which metrics the feature should move. If those metrics already matter to the company, your case is stronger.
- Build the simplest possible prototype. Don’t worry about metric movement yet. If the backend isn’t feasible, a “painted door”—a UI without the underlying implementation—is enough to validate interest.
- Develop a minimal MVP that could move one or two target metrics if tested. Keep scope small but aim for measurable impact, then run an A/B test.
- Switch into product management mode. Draft a one-page brief that covers the roadmap, required stakeholders, and dependencies for the full feature.
- Use the test data to get buy-in. Present the A/B results and the roadmap together, and escalate through leadership to get teams coordinated.
- Set a target launch date. Even a straw-man date helps with prioritization. Align everyone on the plan and ship.
This approach got the feature live and gave users more direct control over their recommendations. The launch isn’t the end of the work, though. Feedback since release has been substantial, and we’re continuing to iterate on the feature to improve recommendation quality across Spotify.
Many teams across Spotify contributed to making this possible—product, design, data science, engineering, user research, data infrastructure, marketing, PR, and legal all had a hand in the launch. It’s a reminder that even a self-directed project ends up depending on a wide network of collaborators.
