An embedded video is an iframe. The content lives on another server, in a format that most retrieval pipelines do not process, and what remains on your page is an empty container with some metadata around it.
Whatever expertise you demonstrated in those twenty minutes is invisible unless you put the words on the page. The fix is not complicated. It is just tedious enough that most teams skip it.
Auto-Generated Captions Are Not a Transcript
Platform captions get you most of the way and the gap matters more than it looks. Speech recognition handles ordinary language well and handles technical terminology, product names, and acronyms badly, which is exactly the vocabulary you most want indexed.
The failure is specific: your product name comes out as something phonetically similar, a schema type becomes two unrelated words, an acronym gets spelled out wrong. The transcript reads fine to a skimming human and contains none of the terms someone would actually search for.
Pull the auto caption file, correct the specialist vocabulary, and publish that. Correcting a twenty minute transcript takes maybe half an hour and it is the difference between text that matches real queries and text that does not.
Timestamps are worth keeping while you edit and worth stripping before publishing, unless you are using them for chapter navigation. A wall of text interrupted every eight seconds by a timecode is harder to extract cleanly.
Where the Transcript Goes Matters
Three common placements and they are not equivalent.
Directly in the page HTML, visible, below the video. This is the version that works. The text is in the server response, it is available to any crawler, and readers who prefer skimming to watching get something useful.
Behind a "show transcript" toggle. Fine if the text is present in the HTML and hidden with CSS. Not fine if the toggle fetches the transcript via JavaScript on click, which puts it in the situation covered in the SSR versus CSR guide. The visual difference is nil. The indexing difference is total.
On a separate transcript page. This works but splits the signal across two URLs, and you now have a page of raw transcript with no context competing with the page that has the video and the framing. I would keep them together.
Raw Transcripts Are Bad Reading and Mediocre Content
Speech is repetitive, full of false starts, and structurally loose. A verbatim transcript is accurate and hard to extract anything clean from.
What performs better is a lightly edited version: false starts removed, filler cut, headings inserted at topic changes, paragraphs broken where the subject shifts. You are not rewriting what was said. You are giving it the structure spoken language never has.
Adding headings does the most work here. A twenty minute walkthrough usually covers four or five distinct things. Marking those with headings turns an undifferentiated block into a document with navigable sections, and section level extraction is far more likely to produce a usable citation than pulling from the middle of an unbroken wall. The same structural logic applies to written support content, covered in the help center guide.
Schema Should Describe Both the Video and the Text
VideoObject schema, covered in the video AEO guide, has a transcript property that most implementations leave empty.
{
"@context": "https://schema.org",
"@type": "VideoObject",
"name": "Setting up schema markup for AI search",
"description": "A walkthrough of implementing Organization and Article schema.",
"uploadDate": "2026-07-16",
"duration": "PT18M42S",
"transcript": "Full corrected transcript text goes here.",
"embedUrl": "https://www.youtube.com/embed/example"
}Populating transcript is redundant with having the text visible on the page, and redundancy is cheap here. If one signal fails the other still carries.
Where a video sits inside a longer article rather than being the whole page, VideoObject should be one node in a graph alongside Article schema rather than the page's primary type, following the stacking approach in the schema markup guide.
The Argument Against Transcripts, and Why I Do Not Buy It
The objection I hear is that publishing the transcript removes the reason to watch, which hurts view counts and engagement metrics.
I think this is backwards for most business content. People who want to watch a video watch it. People who want to scan for one specific answer were never going to watch twenty minutes, and without a transcript they leave and find that answer elsewhere. The transcript captures an audience the video was never going to serve.
If your business model depends on watch time specifically, the calculation is different and I would take the objection seriously. For a company publishing educational content to build authority, view count is not the outcome that matters.
Podcasts Have the Same Problem, More Severely
Audio has no thumbnail, no visual context, nothing. An episode page without a transcript is a title, a description, and a player. The podcast guide covers this and the mechanics are identical: correct the auto transcript, structure it, publish it visibly, mark it up.
Podcast transcripts are also where the auto caption vocabulary problem is worst, because guest names and company names appear constantly and speech recognition mangles proper nouns more than anything else.
Checking Yours
Load a video page with JavaScript disabled. If the transcript is not there, it does not exist as far as most crawlers are concerned.
The NotionCue AI Crawler Audit checks the server response, which resolves the toggle question without you having to reason about how the component was built.
Start your free NotionCue trial and audit your video pages specifically. Teams that have done transcript work often find it is being injected client side and reaching nobody.
If you are starting from a back catalogue of untranscribed videos, do not try to do all of them. Take the five that cover topics you most want to be cited for and do those properly. The rest can wait or stay as they are.
Common Questions
Does duplicate transcript text across my site and YouTube cause a problem?
Not meaningfully. They are different platforms serving different contexts, and this is not the canonicalisation situation described in the canonicalization guide, which concerns duplicate URLs within a domain.
Should I transcribe every video?
No. Transcribe the ones covering topics you want citations for. A product announcement from three years ago is not worth the effort.
Do AI systems process video directly yet?
Some multimodal capability exists and it is not something I would build a content strategy around. Text remains the reliable path, and a transcript costs half an hour.