ChatGPT reads YouTube descriptions, not transcripts, when it retrieves a video

David Konitzny, writing on LinkedIn on 28 September, reports that ChatGPT does not watch or transcribe YouTube videos when it retrieves them. It receives a text snippet, and the transcript is not in it.

The snippet, per Konitzny's reading of the ChatGPT retrieval leak, contains the title, channel name, follower count, verified status, publish date, views, likes, and the full video description including links, timecodes and structured lists. He notes the description is often longer than everything else in the snippet combined.

Konitzny also flags that YouTube videos carry their own retrieval scoring in the leak, with views, followers and channel verification acting as small boosters. Retrieval and citation are separate steps, so a video can be pulled into consideration without appearing in the final answer.

What this changes about description writing

If the description is the largest block of text the model has, it is doing the work a transcript would otherwise do. Chapter titles and timecodes come through as readable text, which means the model can tie a claim to a specific point in the video.

Konitzny's advice is to treat the description like a landing page: say what the video covers, use the relevant terms, and keep the structure intact.

Checking this on your own channel

Before rewriting anything, verify the finding against your own videos. Open your top-performing videos and read each description on its own, without the video or any transcript in view. Check whether the text alone conveys what the video covers, who it is for, and what happens at each timecode.

Where the description falls short, the model is working from a thin snippet. Rewrite so the description answers the question a reader came with, and keep the chapter markers and lists that the snippet preserves.

This sits alongside the other signals Konitzny names from the leak. Description quality is one input to whether a video gets cited, not the whole story.