Building a faceless YouTube channel and learned the algorithm reads your audio, not just your metadata
I run a faceless YouTube channel that makes long-form focus music (2–4hr sessions) for founders, builders, and coders. No voiceover, no lyrics, no transcript — which means title/thumbnail/description carry almost all the semantic weight of the video.
Three weeks ago I learned something I didn't expect: the actual acoustic content of an AI-generated track can route your video into the wrong algorithmic co-viewing graph, independent of your title or tags. One upload used "reverb pads" in the generation prompt for a dark-ambient track — technically correct genre description — and it got quietly classified alongside meditation/wellness content instead of focus/coding content. Impression budget for the whole content line dropped by roughly 90% for the next several uploads, and it took unlisting the offending video plus rebuilding the audio prompt language from scratch to start recovering.
The part that surprised me: this had nothing to do with description text or tags. Same tags, same title formula, same thumbnail style as videos that performed fine — the only variable was a couple of words in the audio generation prompt. Metadata and packaging get all the attention in "how to grow on YouTube" advice; the content's own acoustic fingerprint apparently gets read and classified too.
Today's upload is the first one I'd call a clean recovery test — checking whether Suggested CTR climbs back toward baseline now that the prompt language is quarantined.
Curious if anyone else building content products has hit a platform behavior this opaque — where the thing tanking your distribution wasn't your metadata or your quality, but something structural you'd never have thought to check?
