What started as a casual experiment quickly turned into an unexpectedly successful journey in music discovery. Treating Gemini as my personal DJ and pushing it to find exact song-level matches—not just broad artist suggestions—led me to a track that instantly soared to the top of my playlist: “Holy Mother” by Starbenders. The path wasn’t flawless, but it revealed how conversational AI can cut through the usual noise in music discovery, surpassing the limitations of traditional recommendation engines.
Turning Gemini Into a Personal DJ for Song Matches
I’m a listener who fixates on the best track by an artist, often ignoring the rest of their catalog. This behavior tends to confuse many recommendation systems, including Spotify’s intelligent AI DJ, which typically pushes me toward related artists or large genre categories with traits I don’t necessarily vibe with. As a result, my “Liked Songs” list grows agonizingly slowly.
So, I gave Gemini a straightforward directive: no more playlists I’d ignore—only individual songs, with no unnecessary recommendations for extra tracks or artists. I provided details about what I appreciate in reference tracks and asked for one sonic and one structural match, along with brief explanations for each suggestion. I also instructed it to avoid overexposed hits and dive deeper into lesser-known gems.
Hallucinations Disrupted the Music Search
The biggest issue I encountered was hallucination—Gemini confidently recommended nine songs that didn’t actually exist across several sessions. Initially, it blamed punctuation errors or streaming availability but eventually admitted to mistakes. This wasn’t just frustrating—it cost time and trust. Researchers at places like Stanford HAI have documented that large language models can produce believable but fictional outputs, and music recommendations are no exception.
The fix was to enforce prompt discipline. I told Gemini to confirm each song’s existence on major streaming platforms before recommending it, provide release details, and clearly mark any uncertainties. This single condition substantially reduced fictional suggestions and saved me from fruitless chasing.
The Breakthrough Prompt That Unlocked Better Matches
What finally clicked was grounding Gemini in my real listening data. I uploaded a playlist screenshot and asked it to extract the tracks into a table, then infer common features among the songs I frequently replay. Instead of generic genre labels, it generated nuanced descriptions like “theatrical rock” and “showtunes-meets-rock,” inspired by Foxy Shazam’s “Oh Lord.” That perspective shift was the key.
From there, Gemini quickly found two hits for my Liked Songs: “Holy Mother” by Starbenders and “Could Have Been Me” by The Struts. “Holy Mother” especially captured an elusive edge—anthemic without being saccharine, dramatic but not parody—which I found myself replaying daily. Alongside commentary from others, I uncovered six more favorites: “In Between” and “Disease” by Beartooth; “Hometown” by Cleopatrick; “45” by Shinedown; “Medicate” by Hollywood Undead; and “Black Holes (Solid Ground)” by The Blue Stones. Some made it to my main list, while others joined themed playlists for future exploration.
Why This Matters for Modern Music Discovery
Streaming dominates music listening today—two-thirds of recorded music revenues come from streaming, per IFPI’s latest Global Music Report, and that share is growing. Yet, most leading algorithms are engineered at scale to serve broad tastes, not the unique, idiosyncratic preferences of many listeners. Traditional collaborative filtering excels at “You liked this artist, so here’s more like it” but struggles with “I love this one oddly specific track.”
This is where chat models shine. They can translate fuzzy preferences—“this scratches a brain itch”—into semantic features, often contrasting: dynamic builds, theatrical vocals, major-key anthemic choruses, crunchy mid-tempo guitars, or retro glam minus the hair-metal camp. When these become explicit constraints, discovery can bias toward hidden gems rather than recycled genre clichés.
How to Try This Approach Yourself
- Ground the AI: Provide a screenshot or copy/paste of a 20–40 song playlist you frequently play. Ask the model to
tabulate the tracks and describe common attributes without defaulting to genres. - Demand song-level recommendations: Specify no generic artists and ask for matches based on mood, tempo, vocal
style, dynamics, and production details, with explanations for each pick. - Combat hallucinations: Insist the AI verify each track’s availability on major streaming services and allow it
to express uncertainty. Correct mistakes and refine descriptors accordingly. - Iterate and enrich your profile: As your Liked Songs list grows, standard recommendation engines improve as
well. Combining semantic cues from Gemini with collaborative signals creates a powerful discovery funnel that
expands without watering down your taste.
Judgment After Testing Gemini as a DJ
Gemini didn’t replace algorithmic playlists or radio-style features, but it became a valuable co-pilot. Despite the hallucination hiccups, it surfaced eight truly enjoyable songs this week, adding three new favorites to my daily listen. If you’re a picky listener often overlooked by typical recommendations, the conversational AI layer on top of your music library might be the catalyst that uncovers your next obsession.
