I live for music discovery, but Spotify’s recommendations had become a dead end—same hits, same artists, same playlists resurfacing week after week. I wasn’t quitting the app I use everywhere, from smart speakers to the car. Instead, I rebuilt my recommendation engine from the ground up.
For context, Spotify’s latest filings cite more than 600 million monthly users and over 235 million paying subscribers. With that scale, the recommender system optimizes for engagement at mass, not for the nuance of your personal taste. The result can feel like a perpetual popularity loop rather than genuine discovery.
Why Spotify’s picks go off track
Most mainstream streaming services lean on collaborative filtering—people who liked X also liked Y. The ACM RecSys community has documented how this approach amplifies popularity bias, rewarding what’s already big and flattening the long tail of niche tracks. That’s fine if you want crowd favorites; it’s lousy if you’re chasing depth or regional micro-scenes.
There’s also weak negative feedback. Spotify’s own research from the Echo Nest era showed roughly a quarter of songs get skipped within the first five seconds, yet the app doesn’t always treat a quick skip as a strong “no.” Unless you actively hide tracks or block artists, the same offenders tend to boomerang back in radios and mixes.
Finally, promotion and demographics steer results. Industry reporting has scrutinized programs like Discovery Mode, where artists trade lower royalties for added algorithmic exposure. Combine that with localized trending pushes, and your feed can skew toward what’s marketable nearby, not what you’d actually love.
What I changed to retrain Spotify
Start by cleaning the signal. I audited my Liked Songs and saved albums, ruthlessly unliking tracks that no longer reflect my taste. A smaller, truer core library sends much clearer cues to any recommender system.
Quarantine the noise. I toggled “Exclude from your taste profile” on kids’ music, sleep sounds, and gym playlists, and I use Private Session when I hand my phone to guests. This keeps situational listening from polluting my baseline recommendations.
Turn off drift. I disabled Autoplay so the queue doesn’t wander into generic filler after an album ends. If I want a radio, I start one deliberately from specific seeds.
Feed it better seeds. I built tight, theme-first playlists—think “instrumental post-rock under 110 BPM” or “melodic techno with low vocal presence”—and then ran song or artist radios from those lists. Narrow, coherent seeds yield far better adjacent finds than a grab bag of genres.
Use hard negatives. I lean on “Hide this song” in mixes and radios and “Don’t play this artist” for acts that consistently miss. Quick skips aren’t enough; explicit negatives reshape future queues.
Follow humans, not just algorithms. I subscribed to public playlists from tastemakers—NPR Music, KEXP, BBC Radio 6 Music, respected indie labels, and trusted friends. Human curation introduces off-grid records the algorithm often overlooks.
Third‑party tools that actually help
Chosic let me generate discovery playlists by energy, valence (“happiness”), acousticness, danceability, and even BPM. When I wanted moody, midtempo guitar work or hi-energy, low-vocal electronic tracks, I dialed in those attributes and exported the results straight to Spotify.
Playlost.fm matched my playlists to similar, user-made lists without forcing me to hand over full account access. The overlap-based approach cuts through local chart bias and surfaces deep catalog cuts from adjacent scenes.
For maintenance, I used tools like Sort Your Music to batch-view tempo and key, and Obscurify to identify overplayed comfort picks versus underrepresented corners of my library. Glenn McDonald’s Every Noise At Once map (a genre atlas created during his tenure as Spotify’s data alchemist) helped me jump between micro-genres with surgical precision.
What changed—and how I measure it
Before this reset, my Discover Weekly averaged one or two keepers. After six weeks of disciplined feedback and seed curation, I’m saving eight to ten tracks from each Discover Weekly and three to five from Release Radar. More importantly, the artists are genuinely new to me, not just the usual suspects in different packaging.
I also track skip behavior: fewer instant skips, more full plays. That lines up with broader listening patterns the IFPI reports—listeners reward relevance but punish repetition. In short, the recommender responds when you give it stronger, cleaner signals.
A simple weekly routine to keep it sharp
Once a week, I triage Discover Weekly and Release Radar: save the winners, hide the misses, block repeat offenders. I seed two fresh radios from tightly themed playlists, and I import one small Chosic or Playlost.fm set to nudge the graph in a specific direction. Fifteen focused minutes keeps the system honest.
Bottom line
Spotify’s default recommendations favor what’s popular and promotable, not necessarily what’s right for you. By cleaning your signals, isolating outlier listening, using explicit feedback, seeding with intent, and leaning on human curation and smart external tools, you can bend the algorithm back toward your taste. It won’t become a bespoke DJ overnight—but it can stop stinking.