You're watching a show and a song comes in over a scene — something you've heard before, something you almost recognize — and you Shazam it. Thirty seconds later your phone tells you the title and the artist. That part is solved. It's been solved for years. But the question you actually wanted answered is the one that comes next: where else have you heard this? Which other shows, which films, which games does this song live in? That answer is nowhere in one place. We've spent the last several months building toward it.
SceneSong started from a simple premise: song recognition is a commodity now, but song context is not. The fingerprint match is easy. The cultural map behind the match — all the scenes and credits and moments where this particular recording has appeared — that's the hard part. It's hard because the data has never really existed as a single coherent thing. It lives in fragments across a dozen different places, most of them built for other purposes.
Where the data actually lives
There is no official "song to scene" database that covers TV, film, and games comprehensively and is available for apps to build on. What exists instead is a collection of partial answers. Soundtrack albums on streaming services tell you which songs a studio selected for the official release — but most of what gets licensed for a TV show never makes the album. Official production credits are accurate but scattered across individual show pages and subscription data services. And then there's the part nobody predicted: Wikipedia.
Search for "Music of Breaking Bad" and you'll find a Wikipedia article that runs thousands of words, organized by season and episode, logging every significant song placement the show ever made. "Music of The Sopranos" goes deeper. "Music of Grey's Anatomy" is a document of almost archaeological completeness, covering sixteen seasons' worth of needle-drops, episode by episode, with notes on where in the episode each song appears and what emotional register it was playing in. These articles exist because music supervision in prestige television has become its own art form, and the audience noticed. The fans who maintain them are doing something quite serious under the modest cover of a wiki.
We mine these articles. We also work through MusicBrainz — an open music database that catalogs every soundtrack album ever released, fully credited. If a song appears on a released OST, MusicBrainz knows the relationship between the recording and the film. Wikidata, the structured-data layer behind Wikipedia, links musical works to the films and shows they were present in through formal semantic relationships. All of this is free, openly licensed, and maintained by communities that care about accuracy for its own sake.
The AI extraction layer
Structured sources only go so far. For everything not in those databases — mid-tier TV shows without a Wikipedia music article, smaller films, games outside the major franchise releases — we use web search combined with AI extraction. The approach is straightforward in concept: search for a song, pull the most relevant results from reputable sources, fetch the actual pages, and ask an AI model to extract structured appearance data from what it reads on those pages.
The constraint we built in from the start: the AI is not allowed to report an appearance unless it can point to a specific URL in the search results that supports the claim. It cannot reach back into training memory and produce something plausible-sounding. Every data point needs a citation that we can verify — we check each one. About ten percent of what the model extracts gets dropped at this stage because the URL turns out to be a redirect loop, a deleted page, or a result that didn't actually discuss the song in context. That's the cost of doing it this way. The benefit is that the data we keep is grounded.
The AI is not allowed to report an appearance unless it can point to a specific URL that supports the claim. Plausible isn't good enough. Every data point needs a citation we can verify.
We use Claude, Anthropic's AI model, for the extraction step. Haiku — their smaller, faster model — turns out to be exactly the right tool for this. It doesn't need deep reasoning. It needs to read a page, identify the relevant facts, and structure them correctly. It does this reliably and quickly enough that we can process thousands of songs at a time.
The cover problem
"Mad World" is a Tears for Fears song from 1982. It is also, for many people, the Gary Jules version from the Donnie Darko soundtrack, released twenty years later. The two recordings are the same song in the way that two performances of the same piece of music are the same piece of music — the underlying composition is shared, but the recordings are distinct, and they have different histories.
This creates a specific problem for a recognition app. Someone hears the Jules version playing in a scene. Shazam identifies it — correctly — as "Mad World" by Gary Jules. But our database might have the Tears for Fears version mapped to its appearances, and the Jules version mapped to its own. If the user got back the wrong artist from the fingerprint match, or if we only have one version in the database, the result is nothing. And nothing, when you're trying to answer a question you've had since a song came in over a scene, is the worst possible outcome.
We now show a separate section on results pages: "Other versions also heard in." If you search "Mad World" by either artist, you see that artist's appearances first, and then, clearly labeled, the appearances associated with the other recording. Same song, different versions, different histories. The section exists to catch the cases where the recognition was slightly off, or where the database knows one version better than another. It's not a complete solution, but it's the honest one.
How you can help
No automated system is going to catch everything. Music supervisors license songs for scenes that get no write-up anywhere. Niche shows develop die-hard followings who know every song placement by heart and have never had a useful place to share that knowledge. A fan of Skins or The O.C. or Chuck carries around a catalog of song placements that doesn't exist in any database, and probably never will unless someone puts it there deliberately.
Every result screen in SceneSong has a "Submit a sighting" button. If you heard a song in something we don't have listed — episode number, season, title, any detail you remember — tell us. Submissions get reviewed, and when enough users confirm the same placement, it gets marked verified. The community layer is not a fallback. It's the part of the database that no algorithm can build on its own: first-person knowledge from people who were paying attention when the song came in.
The database is never finished. Music is still being licensed for new shows every week. Old shows get rediscovered. People remember things. We add to what we have every day, and we are nowhere close to done — which is, if you have any tolerance for an ongoing problem, a reasonably good place to be.