I am always hungry for music discovery, but Spotify’s recommendations had led me to a dead end — the same old hits, the same old artists, the same old playlists surfacing once a week like a tired pit crew slapping the same four tires on the car.” I was just quitting the app I use across devices, from smart speakers to the car. Instead, I re-built my recommendation engine from scratch.
For context, Spotify’s most recent filings show it has over 600 million users a month and more than 235 million paying subscribers. At that scale, the recommender system is optimizing for engagement en masse, not for the nuance of your individual taste. It can result in a perpetual popularity loop, instead of discovery happening.
Why Spotify’s picks miss the mark
The vast majority of the mainstream streaming services rely on collaborative filtering — “people who liked X also liked Y” — and the ACM RecSys community has described how this method of personalisation compounds popularity bias, “giving rise to a rich-get-richer dynamic that benefits the already popular, at the expense of the vast majority of other tracks.” If what you’re after is crowd favorites, then that’s great, but if you’re seeking depth, or regional micro-scenes for that matter, it’s bad.
There’s also weak negative feedback. Through Spotify’s own research from the Echo Nest era, about a quarter of songs were skipped within the first five seconds, yet the app doesn’t consistently take a quick skip as a strong “no.” Unless you’re actively hiding tracks or blocking artists, the same offenders may boomerang back around radios and mixes.
And finally, promotion and demographics drive results. Industry reporting has probed programs such as Discovery Mode, where artists give up some royalties in exchange for algorithmic exposure. Throw in localized trending pushes and your feed can veer toward what’s marketable nearby, not what you’d truly love.
What I did to retrain Spotify
Start by cleaning the signal. I am trying something a little different: I have gone through my Liked Songs and saved albums, unliking the music I no longer believe to be an accurate representation of my taste. And a smaller, truer core library transmits much clearer signals to any recommender system.
Quarantine the noise. Kids’ music, sleep sounds and gym playlists are what I flip “Exclude from your taste profile” on, and I do Private Session when I give my phone to guests. This prevents situational listening from infecting my baseline endorsements.
Turn off drift. I killed Autoplay, which otherwise allows the queue to go truffling for generic filler once an album reaches completion. If I desire a radio, I start some radio by design from certain seeds.
Feed it better seeds. I constructed tight, theme-first playlists — like “instrumental post-rock under 110 BPM” or “melodic techno with low vocal presence” — and then cued up song or artist radios from those lists. And tight, focused seeds produce vastly superior neighboring finds to a mixed bag of genres.
Use hard negatives. I rely on “Hide this song” in mixes and radios, and “Don’t play this artist” for acts that perennially strike out; quick skips aren’t sufficient, not when explicit screening can distort future queues.
Follow humans, not just algorithms. I followed public playlists by tastemakers — NPR Music, KEXP, BBC Radio 6 Music, respected indie labels and trusted friends. There’s something about human curation that brings in off-grid records that algorithms tend to miss.
Third‑party tools that really work
Chosic allowed me to upload and generate discovery playlists by energy, valence (“happiness”), acousticness, danceability, and even BPM. When I needed moody, midtempo guitar work or hi-energy, low-vocal electronic tracks, I clicked in those parameters and sent the results right out to Spotify.
Playlost. fm matched me to similar playlists made by users, without my surrendering full account access. We find the overlap-based approach is able to remove local chart bias as well as deep catalog cuts from neighboring scenes.
Maintain, however, meant leaning on tools like Sort Your Music to batch-view the tempo and key, or Obscurify to figure out what is an overplayed comfort pick versus an underrepresented corner of some “obscure listening” profile. Glenn McDonald’s Every Noise At Once map (a genre atlas that the Spotify data alchemist made during his time at the streaming service) let me hop between micro-genres with surgical precision.
What has changed — and how I measure it
Prior to this reset my average with my Discover Weekly was about 1-2 keepers. After six weeks of disciplined nudging and seed curation, now I’m saving eight to 10 tracks from each Discover Weekly, and three to five from Release Radar. Most important, the artists are genuinely fresh to me, not merely the usual suspects in fresh wrapping.
And I track skip behavior: fewer instant skips, more full plays. That jibes with broader listening trends the IFPI’s research shows—relevance is rewarded, repetition punished. In other words, the recommender answers when you feed it stronger, cleaner signals.
An easy weekly regimen to keep it sharp
Weekly, I triage Discover Weekly and Release Radar: save the winners, hide the misses, block repeat offenders. I plant two new radios from closely-themed playlists, and I sow a small Chosic or Playlost. fm NOT to force the graph in any direction. Fifteen intense minutes keeps the system honest.
Bottom line
Spotify’s suggested defaults were biased toward what’s popular and easy to market to you, not what’s best for you. But by cleaning your signals, separating out the outlier listening, leveraging explicit feedback, seeding with intent, and relying on human curation and smart external tools, you can bend the algorithm back toward your taste. It will never become a custom DJ overnight — but it can stop reeking.
