We scored songs the model had never seen during training, then checked Spotify plays, playlist counts, and stream thresholds. The result is a held-out map from audio score to real market outcomes.
We grouped the validation catalog into nine score bands and measured Spotify performance in each band. The chart shows median plays, P75, and P90. The y-axis is log-scaled, so every gridline is ten times the previous one.
The climb is monotonic. A song in the 60s typically clears a hundred and twenty thousand plays. A song in the 90s typically clears roughly six hundred thousand. The breakout tail of the 90s band crosses ten million.
The second chart shows threshold-crossing rate by band. As scores rise, more songs cross 100k and 1M Spotify plays. These are traction markers, not Gold or Platinum hit markers.
| Score band | Songs | Median age | Median plays | Strong (P75) | Breakout (P90) | Median playlists | ≥ 100k | ≥ 1 M |
|---|
A 60 can look mediocre because people read 0-100 like a school grade. That is wrong here. In this validation catalog, the 60s are where songs cross 100k Spotify plays more often than not.
Some market signal. Above the floor but not yet a strong traction tier. A starting point worth iterating from.
The practical crossing point. Songs here clear a hundred thousand plays more than half the time. Strong for independents.
Aligned with meaningful traction. The bar most serious release campaigns clear. A quarter cross a million plays.
Three quarters cross a hundred thousand. A third cross a million. Algorithmically loved by the catalog at large.
Top of distribution. Forty-one percent cross a million Spotify plays. High-streaming tracks are densest here.
Useful framing: "This score estimates market pull from audio alone, mapped to realistic stream tiers under comparable catalog conditions."
In the catalog's top 1% of streaming tracks, most scored in the 80s, many scored in the 70s, and thirteen percent scored below fifty. The score pushes high-streaming tracks to the right of the distribution, with leakage on both ends.
Audio is one input. Exposure, brand, sync placements, and viral timing are not in the file. A score below 50 means the model does not hear typical market signal. A score above 80 means the audio has it. Reach still decides who hears it.
Seventeen and a half percent of the catalog's top 10 % of streamers scored below fifty. Twelve and a half percent of the top 1 % did. Most were driven by reach, placement, or timing no audio model could hear.
If you scored low and have momentum, trust the momentum.
Just over three percent of songs scoring 80+ landed under a hundred thousand plays. These usually have real audio pull and limited distribution: bedroom producers, slow-burning indie cuts, or regional catalogues.
If you scored high but stalled, the gap is almost always reach.
The overall score blends four sub-scores. We tested each one against Spotify plays and playlist counts. Three carry market signal. Distinctiveness does not.
Catchiness is strongest for playlist placement (rank correlation 0.51). Genre Competitiveness is strongest for raw plays (0.43). Production Quality is steady across both. Distinctiveness correlates with neither.
How to read Distinctiveness. Distinctiveness can help positioning, press, and sync conversations. In this catalog, it did not move plays on its own. Treat it as identifiability, not scale.
We computed rank correlation per genre. Classical, Latin, and Hip-Hop sit at the top: coherent genres with more standardized market behavior. Jazz is the floor. Its long-tail consumption is curated, mood-based, and archival, which weakens the audio-to-stream link.
If your song is in Jazz, Blues, or R&B, treat the score with a wider error bar. It is less precise than the catalog average.
The score comes from audio. It does not know how big the artist already is. We split the validation catalog by Spotify artist popularity, a 0-100 measure of artist listening attention, and re-ran the analysis inside each tier.
Across each row, the ladder persists. Higher-scoring songs outperform lower-scoring songs inside every popularity tier. Down each column, exposure dominates. A 60 from a major-market artist can beat an 80 from an emerging artist by orders of magnitude.
This report uses the 20% validation split of our catalog, a sample our model never saw during training. Every song is a real Spotify release with a real play count. The numbers reflect listener behavior over the 15 to 21 months following release.
The four principles below are how we keep the data honest.
Plays and playlist counts come directly from Spotify, not proxies or estimates. No synthetic labels.
Scores in this report were produced on songs the model never saw during training: a genuine 80/20 split.
Every score band carries its median track age (15 – 21 months). You can read how long Spotify had to find the songs.
We do not see exposure, brand, sync placements, or release strategy. We rate the audio. We report probability, not destiny.
Horizon matters. The label is total Spotify plays at catalog read time, so older releases had more time to accumulate plays. We anchor each analysis to median track age inside the band. Future cuts should add explicit 30-day and 6-month windows.
This is a validation study, not a causal claim or guarantee. It asks whether today's score sorts songs in the same direction as later market outcomes. The answer is yes, with the caveats above.