What matching 445 albums to streaming data actually taught us
The rating maths is the easy part. Almost every real problem in the pipeline turned out to be the question of whether a row of data is about the song you think it is about.
The premise sounds trivial. Take a song, find its play count, look the number up on the ladder. In practice, the step where you find its play count is where nearly all the errors live, and a wrong match does not announce itself. It produces a plausible number attached to the wrong song, which then propagates into a rating that looks fine until someone who knows the album sees it.
Reprise hijacks
The worst category, and the one that gave the failure mode its name. Albums are full of tracks that share a title with a much bigger song: reprises, interludes, intros, skits, alternate versions. Match by title alone and the reprise inherits the hit's streaming numbers.
The result is a two-minute interlude carrying the evidence of a generational single, and a rating to match. Put plainly: an artist's biggest video was getting stapled onto their deep cuts. Because every individual number in the chain was real, nothing in the pipeline looked broken. The reprise simply appeared to be one of the biggest songs on the record.
Short-form tracks now get a deliberately buried popularity key and a hard rating ceiling, and the matcher rejects pairings where the durations diverge by more than about a fifth. A ninety-second interlude cannot claim the numbers of a four-minute single regardless of what its title says.
Duplicate identifiers
A related failure with a cleaner fix. If two tracks anywhere in the catalog end up carrying the same streaming identifier, at least one of them is wrong, and both will be rated off the same evidence. This is treated as a hard build failure rather than a warning, because there is no scenario where it is correct and no way to spot it by reading ratings.
These arise more often than expected, usually through compilations, re-releases, and songs that appear on both a studio album and a later greatest-hits package.
Deluxe editions and the identity problem
Track identifiers are index-based, derived from the album and the track's position. That is compact and readable, and it means inserting or reordering a single track shifts the identifiers of everything after it. Which is fine, until you consider what "the album" means.
A record might exist as a standard edition, a deluxe with four extra tracks, a platinum edition with different extras again, and a television edit. Search does not reliably resolve these, and picking the wrong one shifts every position and quietly mismatches most of the record. The pipeline therefore supports pinning a specific known-good album identifier per release, and after any tracklist change the affected album has to be rematched by title rather than by position.
The 80% ceiling on YouTube views
Views carry real weight in the blend, second only to streams, and they feed the within-album ordering as well. They are fetched without an API key by reading the public search page directly, which removed a hard quota that previously capped coverage at around a hundred tracks a day.
Coverage tops out at roughly eighty percent, and the missing fifth is not random. It is skits and interludes, which nobody uploads. It is remixes and alternate versions whose only video is the original, which is exactly the case where a loose match would be most damaging. And it is album tracks from artists who never released an official video for anything beyond the singles.
That last category is a genuine finding about how albums are structured rather than a limitation of the scraper. The videoless portion of a record is a reasonable proxy for the part of it the label never promoted, which is not the same as the part listeners ignored, and the distinction matters when you are drafting.
Because coverage is incomplete, missing views renormalise out of the blend rather than counting as zero. A song is never penalised for lacking a signal. It is rated on the evidence that exists for it.
The rule underneath all of it
Every matching decision follows one principle: a dropped match costs coverage, a wrong match corrupts a rating. Those are not comparable harms. Coverage gaps are visible, bounded, and degrade gracefully, since the blend simply renormalises. A wrong match is invisible, unbounded, and actively misinforms.
So the matcher is strict to the point of being annoying. A title must match either as a full substring or on at least eighty-five percent of its significant words at word boundaries, and the artist must be the primary credit rather than a feature. Validation runs offline against the cache and drops anything that fails, which is why coverage occasionally goes down after a maintenance pass. That is the system working.
The audit suite exists mostly to enforce this. Six checks have to pass before ratings ship, and the one that matters most for matching verifies that a track rated above a less popular labelmate has actual evidence for it: a chart peak, a certification, or a view count. A deep cut outranking a hit with nothing behind it still fails the build. The instruction written into the pipeline documentation is not to weaken that check to get a green run, but to fix the data instead, and it is there because the temptation is real at eleven at night.
Play #1 Debut