While building a YouTube transcript exporter (a browser extension I submitted to Edge Add-ons, currently in review — not live yet), I found that the selectors in older guides return zero hits on the live transcript panel. Here is what the panel actually renders now, measured on 2026-11-15 and 2026-11-16.
The old selector is dead
ytd-transcript-segment-renderer — the element older guides and code samples target — returned 0 hits on the live panel.
What the panel renders now
Each transcript row is a <transcript-segment-view-model> element. The parts I use:
- Timestamp:
.ytwTranscriptSegmentViewModelTimestamp - Text:
span[role="text"]
My extraction tries the new DOM first and falls back to the old selector, so one code path covers both renderings.
Measured on a real video
Video: Rick Astley — Never Gonna Give You Up (dQw4w9WgXcQ). With the transcript panel open, the extension read 24 cues / 2,066 characters from the panel, and built all four formats from those real cues: TXT 2,335 bytes, MD 2,668 bytes, SRT 3,142 bytes, VTT 3,162 bytes. Spot-checked content was correct.
Two caveats before you rely on the panel
- The panel only renders the rows currently on screen. On this video that was the first 24 cues — not the full transcript. The full transcript still needs the timedtext endpoint.
- timedtext does not work from a datacenter IP. From my environment, curl got HTTP 200 with 0-byte bodies across formats, then 429s; a real browser on the same network got empty json3 through srv3 responses too. Residential IPs presumably work, but I could not verify that from here. So the order that works for me is: try timedtext first, fall back to the panel DOM.
If you maintain a transcript tool, check your selectors against the live panel — the old ones silently return nothing now.
Top comments (0)