Blog · Third-party sources

YouTube, transcripts and video citations

Published 8 October 2026 · Updated 12 October 2026 · 8 min read

YouTube, transcripts and video citations
Photograph: Jakob Owens / unsplash

The short answer: A video is usually cited for the text around it, because the spoken words are not in what a crawler can fetch. On the three watch pages we checked on 8 October 2026, YouTube handed a crawler a title, a thumbnail, an upload date and a stack of viewer comments, and nothing else. The caption track sits behind paths YouTube's own robots file disallows. If you want a video quoted, publish its words as readable text on a page you control.

That is the mechanism, and it is the part most video advice leaves out.

What a crawler actually receives from a watch page

We fetched three public YouTube watch pages on 8 October 2026 with a Googlebot user agent and parsed the JSON-LD block each one carries. All three returned a VideoObject with the same four keys: name, thumbnailUrl, uploadDate, and comment. No description. No duration. No transcript. Three pages on one day is a spot check, not a census of YouTube, and the markup can differ by video type, region or rendering path. Run the same fetch on your own video rather than taking our three as the rule.

Compare that against what Google itself asks for. The video structured data documentation, updated 24 September 2026, lists exactly three required properties: name, thumbnailUrl and uploadDate. YouTube ships the minimum, then adds a list of top comments underneath it.

So when a retrieval system treats a watch page as a document and chunks it, the richest prose on offer is what strangers typed in the comment box. That is a strange thing to let represent your product.

The transcript is blocked, not merely absent

Captions are served from a timed-text endpoint under YouTube's /api/ path. Open https://www.youtube.com/robots.txt and read it yourself. Disallow: /api/ is there. So is Disallow: /timedtext_video, and Disallow: /youtubei/, which is the internal player API that the on-page transcript panel calls.

A crawler that obeys that robots file cannot take the caption track. It can render the watch page, and it can read the description, which is often three lines and a link tree. The spoken explanation in the audio is not in what such a crawler collects. Whether a given engine has it by some other route, a licensed feed or a product integration, is not something we can see from outside.

That constraint fits a pattern we see in the citation checks we run during audits: when a video is cited, there is usually text elsewhere carrying its substance, a detailed description with chapter timestamps, a transcript republished on a blog, a news write-up, a forum thread where somebody paraphrased the argument. We have not run a controlled test of this, so treat it as a working explanation rather than a measured finding.

Chapters are the one transcript-shaped thing you control

Google's video structured data documentation, updated 24 September 2026, says Google derives key moments for YouTube-hosted videos from timestamps written into the description, formatted [hour]:[minute]:[second] with a label on the same line, each on its own line, in chronological order. For video you host yourself, Clip markup does the same job explicitly, with name, startOffset and a url that deep-links to the second. SeekToAction tells Google your URL pattern so it can pick moments itself. The same documentation sets a thirty second minimum total video duration for Clip and SeekToAction eligibility and requires that no two clips for one video on one page share a start time.

Here is why that matters beyond a search feature. A chapter list is a block of human-written, topic-labelled sentences sitting inside the only crawlable text field on the page. Fifteen chapters is fifteen short answers, each bound to a timestamp. Write them as claims rather than labels. "Pricing" is a label. "Why per-seat pricing breaks for agencies" is a claim, and a claim is the thing a passage retriever can match a question against.

Auto-captions are not a transcript you can rely on

YouTube's help page on automatic captioning, read 12 October 2026, lists sixty-seven languages for automatic captions on long-form videos and Shorts, and says quality varies with mispronunciation, accents, dialects and background noise. It names the cases where captions may not generate at all: an unsupported language, audio that is too long, poor sound quality, a long silence at the start, overlapping speakers, or multiple languages at once.

Even where the captions are good, they arrive as unpunctuated drift with no speaker labels and no paragraph breaks. Paste one into a page and there is no clean sentence for a retriever to lift, which is the unit most citation surfaces quote. The fix is editorial, not technical. Someone has to turn the caption dump into paragraphs with headings, fix the product names the recogniser mangled, and cut the throat-clearing at the top.

Publish the words on a page you own

The durable move is to stop treating the video as the asset and treat the words as the asset. A video page on your own domain, with the embed at the top and an edited transcript below it, gives an engine something to chunk. Mark the page up with VideoObject, and use the transcript property, which schema.org defines for exactly this purpose on audio and video objects. Google's video indexing documentation, updated 18 December 2025, adds the conditions: the watch page has to be indexed and performing well in Search before its video is considered, the video has to be embedded on that watch page and not hidden behind other elements, and the thumbnail needs a stable URL of at least sixty by thirty pixels.

That page is yours. It is also the asset that earns coverage, because a journalist or a forum poster can quote a paragraph from it without scrubbing through a timeline.

Video sits inside the wider question of the third-party sources AI engines keep citing, and it behaves like the rest of them: you influence it, you do not own it. The same logic runs through review sites, G2 and what AI repeats about your product, where the text on a profile page does the work that your marketing site cannot. On the social side, Grok and social sourcing: when posts become citations covers an engine that leans on a feed rather than a video platform, with a different freshness profile.

A check you can run in ten minutes

Take one video you care about.

  1. Fetch the watch page with curl and a plain user agent. Search the HTML for the first

    2. sentence your presenter speaks. If it is not there, a crawler reading that same HTML does not have it either.

  2. Parse the application/ld+json block. Note which properties exist. If description and

    3. transcript are absent, nothing in that markup carries your factual claims, so they need a home elsewhere.

  3. Read the description as a document with the video removed. Does it answer a question on

    4. its own, or is it a link list?

  4. Search your own domain for the video's title. If the only page that carries the content

    5. is on a platform you do not control, that is the gap.

If citations to a video page you built have fallen away rather than never existed, diagnosing a drop in AI citations is the right order of operations. The ten checks in our self-check cover the access side of the same problem for ordinary pages.

Honest limits

This post establishes what a robots-obeying crawler may fetch from a YouTube watch page. It does not establish what an engine actually holds. An engine may have caption text through a licensed feed or a product integration, and we have no way to confirm or rule that out from outside, so read every statement here as being about the crawl path rather than about retrieval.

Our own check is three public watch pages, fetched once, on 8 October 2026, with a Googlebot user agent. That is a spot check. It does not show that every watch page ships the same four JSON-LD keys, and we have not tested whether the markup differs for Shorts, live streams, members-only videos or other regions.

The claim a careful reader should push hardest on is the link between republished text and citations. We have not run a controlled test, so we cannot say how much a transcript page helps on its own; publishing one usually happens alongside other work, and we report no number for its effect because we have not measured one. The chapter and markup advice is sourced to Google's documentation for Search features, and we are not claiming Google's documented behaviour describes other engines.

FAQ

Do AI engines read YouTube transcripts? Not from the watch page by crawling, as far as we can tell from outside. The caption endpoint sits under paths YouTube's robots file disallows, and the JSON-LD on the three pages we checked carried no transcript property. What is left to crawl is the title, the description, the comments and whatever else on the open web describes the video in text. Whether an engine has captions by a licensed route is not visible to us.

Should I upload a transcript to my own site? Yes, if the video says something worth quoting. Edit it into paragraphs with headings, correct the names the recogniser got wrong, and put it under the embed on a page where the video is the main content.

Do views and subscribers decide whether a video gets cited? We have no measurement either way, and we have not found engine documentation that names view counts as a citation input. What we can say is that a retrieval step matches a question against passages, so a channel with a small audience can still own the clearest written answer to a narrow question. Treat audience metrics as a growth signal first.

Are Shorts worth making for this? We have not measured Shorts against long form. On the mechanism above, a short clip carries little spoken content, thin description text and no room for chapter timestamps, so there is less text for a retriever to match against. That is a reason to expect less, not evidence of it.

What markup should a video page carry? VideoObject with name, thumbnailUrl and uploadDate at minimum, plus transcript where you have one, and Clip entries with startOffset if you want key moments defined rather than guessed.

Sources

Related reading

The third-party sources AI engines keep citingReview sites, G2 and what AI repeats about your productGrok and social sourcing: when posts become citationsDiagnosing a drop in AI citationsWhen the answer names a competitor instead of youDigital PR when the audience is a model