Why Music Metadata Matching Is Harder Than It Looks
A missing Kesha remix exposes three real-world pitfalls in music metadata matching — and the fixes any multi-source pipeline needs.
A 2025 Kesha remix returned zero results from a music API even though it existed on Apple Music, Deezer, Spotify, and MusicBrainz. The data wasn't missing — the matching layer that decides whether two provider records describe the same recording simply couldn't recognize it.
Three mundane failure modes were responsible: raw string comparison broke on punctuation and Unicode dash variants that render identically but differ at the byte level; the primary source's search API turned out to be a partial, popularity-weighted index rather than a full catalog view; and naive 'take the first result' matching against MusicBrainz silently picked wrong recordings roughly one in seven times.
The fixes form a pattern: normalize strings into a canonical form before comparing anything, treat a single source's search miss as inconclusive rather than authoritative, and replace first-result matching with multi-signal scoring across title, artist, and duration — accepting a match only when at least two independent signals corroborate. A wrong match is worse than no match, since it poisons every downstream field without any visible error.
For anyone building systems that reconcile metadata across multiple providers, the takeaway is concrete: never trust raw string equality or a single ranked result as ground truth.