Your best content is a video, and AI cannot watch it
Updated:
Quick answer: a video embedded on your page contributes nothing to what a language model can cite, because retrieval operates on text and an embed is a URL wrapped in markup. The forty minutes of expertise inside it are invisible. A transcript on the same page converts that forty minutes into the most quotable content you own, and almost nobody publishes one.
What does a retrieval system see when it reaches your video?
An <iframe> pointing at a video host. That is the whole payload.
The crawler does not download the video. It does not run speech recognition. It does not follow the embed and analyse what plays there. It reads your HTML, extracts text, and the text it extracts from a video embed is whatever happens to be in the title attribute and the surrounding caption.
So a page whose entire substance is a twenty-minute explanation from your most experienced person reads, to a machine, as a page with a heading, a paragraph of introduction and an empty box.
This is not a limitation anyone is working around on your behalf. It is how the pipeline is built.
- Transcript
- The full text of what was said in a recording, published on the page itself.
- Retrieval
- The stage where a system searches for suitable pieces of text before it generates an answer.
Why is the transcript better raw material than most articles?
Here is the part that makes this worth doing rather than merely correct.
A transcript of someone explaining their work is, structurally, close to ideal for extraction. Spoken explanation has properties that written marketing copy does not:
It is specific by necessity. People speaking out loud about work they do use numbers, name constraints, and describe what actually happens. "Usually it takes about three days, unless the site has more than fifty pages, then it is a week" is the kind of sentence that appears constantly in speech and almost never in a services page.
It answers questions in question-shaped units. Interviews and explanations are naturally organised as question and answer, which is the shape retrieval is trying to build anyway.
It comes from a named person. Attribution is the strongest measured content signal in the GEO-bench study, at a 28% to 41% increase in citations. A transcript is a document made entirely of attributed speech.
It is unedited enough to be uneven. Real explanation includes caveats, exceptions and hedges in the right places, which reads as expertise rather than as marketing.
The people who record video are already producing the highest-value text on their site. They just leave it in a format nothing can read.
What do you publish alongside the video?
The minimum that works, in order of effort.
A full transcript, cleaned lightly. Remove the filler, fix the obvious transcription errors, break it into paragraphs. Do not rewrite it into corporate prose, because the rewriting removes exactly the specificity that made it valuable.
Headings inside the transcript. A forty-minute transcript is a wall of text without them. Break it at topic changes and phrase the headings as questions, matching what is answered below.
A summary of 40 to 80 words above the video. Self-contained, stating what the video covers and what the main claim is. This is the fragment most likely to be lifted, and it sits in the first third of the page where extraction concentrates.
VideoObject markup with a transcript property. Declares that the text belongs to the video and gives the page a machine-readable relationship between the two.
<script type="application/ld+json">
{
"@context": "https://schema.org",
"@type": "VideoObject",
"name": "How a technical audit works, start to finish",
"description": "Walkthrough of the audit process with timings for each stage.",
"uploadDate": "2026-08-27",
"duration": "PT22M",
"transcript": "Full text of what is said in the video..."
}
</script>The markup is the smallest part of the value. The text on the page is the value.
What is the objection, and why is it wrong?
The usual reason people skip transcripts: nobody reads them.
Correct, and irrelevant. Transcripts are not published to be read start to finish. They are published because they convert a format nothing can index into a format everything can.
A secondary benefit that people underestimate: transcripts are searchable. A visitor looking for the one part about pricing can find it with a browser search instead of scrubbing a timeline. That is a real accessibility improvement even before considering machines.
And a third, which matters more than either: transcripts are where you discover what your own people actually say. Reading a transcript of a colleague explaining the work is the fastest way to find the sentences your website should have contained all along.
What is the automation trap?
Transcription is cheap now, which creates a specific failure mode.
An automatic transcript dumped onto the page unedited is a wall of unpunctuated text with misheard proper nouns, no paragraph breaks and no structure. It technically contains the words. It fails at almost everything that makes text extractable: paragraphs in the 40 to 120 word band, self-contained units, headings that mark topics.
The cleanup is not optional and it is not long. Twenty minutes of audio takes perhaps fifteen minutes to clean into something usable. Skipping that step produces a page that passes a checkbox and helps nothing.
Specific things to fix, in order of impact:
Proper nouns. Automatic transcription reliably mangles product names, company names and technical terms. Those are exactly the entities that carry retrieval weight.
Paragraph breaks. Speech has no paragraphs. Insert them at topic shifts, aiming for units of two to four sentences.
Headings. Every few minutes of speech covers a distinct point. Mark it.
Filler. Remove the verbal tics, keep the hedges. "Um" is noise. "Usually, though not always" is information.
What have we not done ourselves?
We do not currently publish transcripts, because we do not currently publish video. So this article is a recommendation we have no field data of our own to attach to, and it would be dishonest to imply otherwise.
What we do have is the mechanism, which is not in dispute: retrieval reads text, embeds are not text, and transcripts are text. And we have the measurement on attributed speech, which is the strongest content signal in the benchmark and which a transcript is made of.
That is enough to recommend it. It is not enough to attach a number to it, so there is no number in this article.
We do not publish video, so we are in no position to teach transcript practice. What I can say is this: everything we ever explained to a client by voice on a call, we had to write out as text afterwards. Otherwise a week later that explanation existed for nobody.
Where does this generalise?
The video case is the most visible instance of a broader rule: any content stored in a format that is not text on your page is content a model cannot use.
Podcasts. Same problem, same fix, and podcast episodes are often the single best-explained content a company produces.
Slides. A deck embedded as an object contributes nothing. The speaker notes, published as text, contribute a lot.
Infographics. The chart is a picture. The figures in it, written out below, are extractable facts.
PDFs. Better than video, because many crawlers extract PDF text, but worse than a page, because the extraction is unreliable and the structure is usually lost.
Images of text. A screenshot of a table is a screenshot. Alt text helps a little. An HTML table helps entirely.
The general instruction is unexciting: if it matters, it should exist as text on a URL you control.
Frequently asked questions
Does YouTube's own transcript count? It helps people find the video on YouTube. It does nothing for your page, because it lives on a domain you do not own and is attached to a video, not to your URL.
Should the transcript be visible or in an expandable block? Visible is safer. Content hidden behind interaction can be treated as less prominent, and some hiding patterns are treated as suspicious. An expandable block that is open by default is a reasonable compromise.
How much cleanup is enough? Enough that a stranger reading a random paragraph understands it. That is the same test as for any other content, because after chunking that is exactly what happens.
Does a summary alone work instead of a full transcript? Partly. A summary is extractable and better than nothing. A full transcript gives the model twenty specific claims instead of one general one.
What about VideoObject without a transcript? It tells a machine a video exists and what it is called. It does not make the contents available. The markup describes the box, not what is in it.
Is this worth doing for old videos? Start with the ones that answer questions people ask. An eighteen-month-old explanation of how something works is as extractable as one recorded yesterday.
The short version
The most useful thing many companies have produced is a recording of someone knowledgeable explaining their work, and it is sitting in a format that contributes nothing to how machines see the site.
Publishing the transcript converts it. The text is specific, attributed, organised as answers, and already written. The only work is cleanup, and the only reason it does not happen is that nobody thinks of a transcript as content.
Sources: GEO-bench measurements on quotation and attribution (Princeton University); schema.org VideoObject specification; research digest compiled from 165 sources on AI search, August 2026. Details in sources/research-notes.md.