How AI Search Reads a YouTube Video,
Tested on Our Own Brand Film

On the morning of 31 October I put a 2:48 film on YouTube. It's called Meet Pip, it's for our sister product, and before it went public we did everything you're told to do for video AI visibility: a checked subtitle file, chapters, a proper description. Then I looked at what YouTube actually serves to anything that comes asking. Alongside our captions, it had generated its own. They got 96.6% of the words right, and still turned "1,000 calls a day" into "a,000 calls a day".

Video AI visibility is whether an AI answer engine can find a video, read what it says, and repeat it correctly. In practice that comes down to text, because text is the part of a video the engines have documented ways of reaching. This article is about how AI search reads a YouTube video, using our own film as the worked example. It's a documented setup rather than a result. We haven't yet seen any AI assistant cite the film, and near the end I'll say how we'll measure it and when we'll report back.

Video AI visibility: the short answer

• The words are what travel. Google's own advice for its AI features is to make important content "available in textual form". A video's title, description, chapters, captions and host page are text. Its pictures aren't.
• Automatic captions are not a safe stand-in. On our film, YouTube's automatic track differed from our checked file in 12 of 353 tokens, and four of those changed a fact: the daily allowance, the domain, and two product claims.
• Some words can't survive speech at all. Pip says our domain as "mailbox, M C P, dot com". The hyphen is never spoken, so no speech recogniser could have written it.
• On-screen text is the weakest signal. The film shows three product facts it never says. To a system that reads text, they don't exist.
• VideoObject helps Google index the video, but Google doesn't document the transcript property. A transcript in the page's HTML is on firmer ground.
• Nobody outside the engines knows the rest. Google, OpenAI, Perplexity and Anthropic have published very little about how a video becomes a source. We'll report what we observe by 31 January 2027.

The worked example: Meet Pip, a 2:48 brand film on YouTube

Pip is the character for Mailbox MCP, which connects Claude, ChatGPT and other AI assistants to a mailbox someone already has. The film is his introduction: what he does, which mailboxes he works with, what it costs. It's a good test case for one reason in particular. It's full of the kind of facts people ask AI assistants about, product names, supported services, prices, a web address, and those are exactly the facts that are easy to get wrong. How the film itself was made, with Gemini, ElevenLabs and Lyria and six things breaking on the way, is 365i Web Design's AI brand video case study. This article starts where that one stops: the film is finished, so what can a machine make of it?

That transcript is on this page as ordinary HTML, collapsed but always in the document. It's the first thing on the checklist below, and it's here because this page is the only one on the web where the film's words exist as page text on a site we control. Here is what we did on YouTube itself, and nothing more.

• A checked subtitle file in English (United Kingdom). It was written from the script's own caption lines rather than by speech recognition, then checked by a person against the finished cut. YouTube serves it back word for word: we compared the two and found no differences.
• Nine chapters in the description. YouTube's rules are that the first must start at 0:00, there must be at least three, and each must be at least ten seconds long. Two pairs of scenes fell closer than that and were merged.
• A full written description covering what Pip does, which mailboxes he works with, and the price.
• Twelve tags. YouTube's own help says tags "play a minimal role in your video's discovery" unless a name is commonly misspelled, so we didn't spend long on them.
• A custom thumbnail, 1280 by 720.
• The AI-use disclosure, ticked. YouTube only requires it for realistic content, and lists AI animation in "a fully animated video" among the things creators don't need to disclose. A cartoon robot isn't going to fool anyone. But the pictures, voice and music all came from AI models, and saying so cost nothing. It shows as the small "AI" badge in the corner of the player above.
• A 1440p upload, so that high-definition versions exist at every quality setting. Our first upload, at 1080p, played to viewers at 240p by default, and 365i has written up why a YouTube video plays in 240p and the upload fix. That's a viewing problem rather than an AI one: none of the engines' documentation we read says anything about resolution.
• Two cut-downs from the same master: a 1:17 teaser for X and a captioned version for LinkedIn. Both have their captions burned into the picture, which is right for people scrolling with the sound off. It also means that on those platforms the words are pixels, and the only text a machine gets is the post above the video.

Which video signals AI search can read, and which it can't

Video AI visibility starts with a simple question: where does a video's meaning live? In four places. Three of them are text, or can be. The fourth is pixels, and it's where most of the effort in a well-made video goes.

Read the table as a statement about access, not about what any engine does with what it can reach. That second question is the harder one, and it has its own section.

What a plain fetch of a YouTube watch page contains

Most crawlers that aren't Google fetch a page the simple way: one request, no JavaScript, no sign-in. So I fetched the film's watch page like that on 31 October and looked at what came back. Google owns YouTube and has no need to fetch its own pages this way, so this tells you nothing about Google. It tells you what everyone else gets.

You can repeat this on any video in two commands. The first fetches the page; the second counts how many times a phrase that is only spoken turns up in it.

# Fetch a watch page the way a simple crawler does: no JavaScript, no sign-in
curl -s -A "Mozilla/5.0" -H "Accept-Language: en-GB" \
  "https://www.youtube.com/watch?v=Ec1UEB5M7eg" -o watch.html

grep -c '"captionTracks"' watch.html   # the caption tracks are listed: 1
grep -c "tidy ghost" watch.html        # the words are not: 0

The point isn't that YouTube is hiding anything. It's that a video's spoken words live one step away from its page, and every crawler that takes only the first step gets the title and the description. That makes the description the most important piece of text you write for a video, and it's usually the most rushed.

YouTube automatic captions vs a checked subtitle file

Uploading your own captions doesn't switch YouTube's off. It made its own automatic English track for the film anyway, and it offers both. So I fetched both tracks with yt-dlp, which handles the signed addresses, and compared them word by word against the file we uploaded.

Our file came back exactly as uploaded. The automatic track came back with 12 differences in 353 tokens, a word error rate of 3.4%. (The script is 347 words; splitting "mailbox-mcp.com" and the like into separate tokens makes 353.) By the usual measure that's a good transcript. By the measure that matters for a product film, it isn't.

Each of the four is wrong in a different way, and that's the useful part.

• "a,000 calls a day". Pip says "a thousand". The recogniser heard it correctly, then turned it into a number and kept the "a". The allowance on the paid plan is now unreadable.
• "mailboxmcp.com". Pip says "mailbox, M C P, dot com". There is no way to pronounce a hyphen inside a domain, so no recogniser could ever have written it. We happen to own the unhyphenated domain and it redirects to the right place. A brand that didn't would have its own film pointing somewhere else.
• "filed in scent". Sent and scent are the same sound.
• "mark it as red". So are read and red. That sentence is the film's main promise, that reading an email doesn't mark it as read, and in the automatic track it no longer makes sense.

The fifth row, "favourite" as "favorite", is a spelling rather than an error: the automatic track is plain English, not English (United Kingdom). It's harmless here. In a film where a product name had a British spelling, it wouldn't be.

YouTube's own guidance is more careful about this than most of what's written about video and AI.

"However, automatic captions might misrepresent the spoken content due to mispronunciations, accents, dialects or background noise. You should always review automatic captions and edit any parts that haven't been transcribed properly."

YouTube Help, Use automatic captioning (verify quote at source)

I read that after running the comparison, and my first feeling was a slightly guilty relief that the number had come out as "a,000" and not "10,000". A garbled figure looks garbled. A plausible wrong one doesn't. What stays with me, though, is that none of the causes YouTube lists applies to our film. Pip's voice is synthetic and very clear, there's no accent to speak of, and the music sits well under him. The errors weren't noise. They were homophones and a hyphen, and better audio fixes neither of those. That changed how I think about the advice. "Review your automatic captions" isn't housekeeping for poor recordings; it's the only way a written fact gets into a spoken film intact. And note what the guidance doesn't say. It doesn't say that uploading your own file makes the automatic one go away. On our film it didn't. That track is still public today, still filing things in scent.

Which track does an answer engine use when a video has two? Nobody has said, and I'm not going to guess. The safe position is to assume either might be read, and to make sure the facts that matter also exist somewhere no speech recogniser is involved: the description, and a page of your own.

# Download both caption tracks for any public video, then compare them
yt-dlp --skip-download --write-subs --write-auto-subs \
  --sub-langs "en-GB,en" --sub-format vtt -o "film.%(ext)s" \
  "https://youtu.be/Ec1UEB5M7eg"

# film.en-GB.vtt is the uploaded file; film.en.vtt is YouTube's automatic track
python compare-captions.py film.en.vtt film.en-GB.vtt

The comparison script is short and published with this site's figure sources. It ignores punctuation and capitals, and prints every difference, so you can judge for yourself which ones matter.

What AI can't read from a video: words that are only on screen

When we went through the film frame by frame, three of its product facts turned out to be shown and never said. They aren't in the voice, so they aren't in either caption track, and they aren't in the description either.

None of this harms the product. All three facts are on the Mailbox MCP website, in words. But inside the film, they're pixels. Could a model read them? If it were asked to watch the film, probably. Google's Gemini API says that when it analyses a video "the model samples the video at a rate of 1 frame per second", and these cards stay up long enough to be sampled. Whether any answer engine watches a video's frames when it builds an answer to someone's question is a different matter, and none of them has said that it does.

The film also makes the opposite point, which is easy to miss. Some of what's on screen only means something because of what's said.

A picture carries a fact only when the words beside it say what the fact is. A model describing that frame would say "a robot dressed as a ghost". The transcript says why.

What Google, OpenAI, Perplexity and Anthropic say about video sources

This is the section most guides to video SEO for AI search skip, and fill with confident claims instead. Here is what each company has actually published, read on 31 October 2026, and where each one goes quiet. If you want the wider picture of how these systems find and choose sources, how AI search works covers the retrieval pipeline.

Google

Google has said the most, across several products that work differently and shouldn't be lumped together.

• Search and its AI features. There are no extra requirements to appear in AI Overviews or AI Mode. The guidance is ordinary SEO, including "Making sure that important content is available in textual form" and "Supporting your textual content with high-quality images and videos, when applicable". The video is support for the text, not a replacement for it.
• AI Mode shows video results. On 15 October 2026 Google said it had started testing "clickable creator and publisher names on video results" in AI Mode. It didn't say how those videos are chosen.
• Gemini. The help centre says "Gemini uses public information from YouTube" and can help you understand the content of a video. It doesn't say whether that means captions, audio or frames.
• Gemini Notebook. Here Google is explicit: "Only the text transcript of the video is imported as a source." It accepts public videos with uploaded or automatic captions, and videos uploaded less than 72 hours earlier may not import yet. This is the clearest statement anywhere that a Google AI product reads a YouTube video as its captions.
• The Gemini API. When a developer passes it a YouTube address it processes frames and audio. That's a capability on request, not a description of how Search builds answers.

The most direct thing anyone at Google has said about video in search came from the head of Search herself, in a long podcast interview in March.

"Now you can understand audio much better. Now you can understand video much better. Now you can understand not just the [...] video transcription but [...] what is the video more about or what's the style or other things like that."

Liz Reid, VP, Head of Search at Google, on the Sources podcast, What happens to Google when AI answers everything?, 6 March 2026, at 1:01:00 (verify quote at source)

I should be honest about how that quote reached me, because it's this article in miniature. I didn't hear Reid say it. I read it in YouTube's automatic captions of the episode, an hour in, and only trusted it once it matched an independent write-up of the same interview. The two "like"s I've cut are marked. What strikes me about her wording is the order she puts things in: the transcription first, then what the video is "more about", then its style. She's describing the transcript as the floor, with understanding built on top. I find that oddly reassuring. The floor is the only part of that sentence I can check, and the only part I control. Everything above it is Google's to work out, and I'd rather it worked out from the right words.

OpenAI

OpenAI documents that "OAI-SearchBot is used to surface websites in search results in ChatGPT's search features". We found no OpenAI statement that ChatGPT search plays a YouTube video, reads its captions, or treats a video differently from any other page. We checked the ChatGPT search help article, the publisher FAQs, the crawler documentation and the file-upload help pages. Uploaded video files are a separate feature, and not how search works.

Perplexity

Perplexity's own use-case page says it "acts as your AI remote, finding the exact YouTube video (and moment) you mean with a single sentence". It doesn't say what it reads to find the moment. We found nothing in its help centre or changelog about captions, transcripts or structured data.

Anthropic

Claude's list of supported uploads covers documents and images, with no video or audio. Its web search fetches pages. We found no statement about YouTube videos or their captions.

What the citation data shows

Observational studies fill some of the gap, with the usual caveat that they show what was cited, not why. The most careful one we found ran 5,000 US B2B software prompts every week from June to September 2026 and found YouTube cited in 36.5% of AI Overview answers and 35.6% of AI Mode answers, against 2.1% of ChatGPT answers. Almost half of Google's YouTube citations carried a timestamp, which points at a particular moment. So videos matter a great deal on Google's surfaces and very little, so far, on ChatGPT's. It's one category of prompt, in one country, and it can't tell you what in those videos was read.

VideoObject schema for AI search, and the transcript property

Structured data is the one part of video AI visibility that Google has documented in detail, and it's the part most often described wrongly. Google's video structured data guide requires three properties, name, thumbnailUrl and uploadDate, and says plainly: "If you don't include the required properties, Google may not be able to extract any information about the video." It recommends description, duration, embedUrl, contentUrl, hasPart for key moments, and a few more.

Three details trip people up.

• transcript isn't on Google's list. It's valid Schema.org, and this page uses it. But the word doesn't appear anywhere in Google's guide, so nothing Google has published says it reads it. Don't let it stand in for a transcript people can see.
• contentUrl is the video file itself, an MP4 or a stream. A YouTube watch address isn't one. For an embedded YouTube video the player address goes in embedUrl, and Google says to use that when contentUrl isn't available.
• An article with a video in it isn't a watch page. Google defines a watch page as one whose "main purpose is to show users a single video", and says pages where the video is "complementary to the rest of the content" aren't. That includes this one. For Google's video results the watch page is YouTube's. We mark the video up here anyway, because the markup is accurate and Google isn't the only thing that reads Schema.org.

On YouTube itself, chapters do the job of hasPart. Google says you can set key moments with timestamps in the description, and that "you can tell us more explicitly about the important points in your video and we will prefer that information". This is the VideoObject this page publishes for the film, shortened to fit:

{
  "@type": "VideoObject",
  "@id": "https://www.ai-visibility.org.uk/blog/video-ai-visibility/#video",
  "name": "Meet Pip: Claude and ChatGPT for your Gmail, Outlook and Microsoft 365 inbox",
  "thumbnailUrl": "https://i.ytimg.com/vi/Ec1UEB5M7eg/maxresdefault.jpg",
  "uploadDate": "2026-10-31T06:50:39+00:00",
  "duration": "PT2M49S",
  "embedUrl": "https://www.youtube.com/embed/Ec1UEB5M7eg",
  "inLanguage": "en-GB",
  "transcript": "Oh! Are we on? Is it recording? Right. Ahem. Hello! I'm Pip...",
  "caption": {
    "@type": "MediaObject",
    "contentUrl": "https://www.ai-visibility.org.uk/videos/meet-pip.en-GB.vtt",
    "encodingFormat": "text/vtt"
  },
  "hasPart": [
    { "@type": "Clip", "name": "Meet Pip", "startOffset": 0, "endOffset": 18,
      "url": "https://www.youtube.com/watch?v=Ec1UEB5M7eg&t=0s" }
  ]
}

The caption property points at a copy of our checked subtitle file, published on this site as WebVTT. It's the same words as the transcript on this page and the file on YouTube, and that agreement matters more than any single property. Google's AI guidance asks for structured data that matches the visible text on the page, and here it does, line for line.

Listing a video in llms.txt, ai.json and identity.json

This site publishes the specification for ten AI Discovery Files, so it would be easy to invent a video field and announce it. We haven't, and we won't: a field that nothing reads is worse than no field. None of the ten files has a place for video. What they can do is point at the page where a video's words are written down, and say who it belongs to.

In llms.txt, list the host page like any other important page, with a description that says what the video says rather than that a video exists:

## Guides
- [How AI search reads a YouTube video](https://www.ai-visibility.org.uk/blog/video-ai-visibility/): the Meet Pip film for Mailbox MCP, with its full checked transcript, and which parts of a video AI answer engines can read as text

In identity.json, add the channel to sameAs, so the channel and the organisation are tied together in one place:

{
  "name": "365i",
  "url": "https://www.365i.co.uk/",
  "sameAs": [
    "https://www.youtube.com/@365i"
  ]
}

That exposes a real wrinkle in our own setup, and I'd rather point it out than paper over it. The film is about Mailbox MCP, but it's on the 365i channel, because that's where our videos live. The channel belongs in 365i's identity file, not Mailbox MCP's, because sameAs says "this is the same entity". The connection between the film and the product is made by the film itself, in its title and its first line, and by the about property in the VideoObject. It isn't something an identity file should claim.

An ai.json file needs nothing new either. Its permissions and attribution rules cover the text on your pages, and once the transcript is on your page, that includes the video's words. Google has said its Search doesn't need any of these files, which we've looked at in detail. They're for the other systems that do read them, and they work only as well as the page they point to.

The film isn't embedded on the Mailbox MCP website yet. When it is, that page is the one that product's files should list.

Video AI visibility checklist: nine steps

This is video AI visibility as a working list. None of these steps guarantees a citation, because nothing does. Each one makes it more likely that if an engine uses your video, it gets the facts right.

1. Say every fact you want repeated. Product names, supported services, prices and limits belong in the script, not only on a card. A web address with a hyphen needs to be written somewhere as well, because it can't be spoken.
2. Upload your own caption file. Write it from the script and check it against the finished cut. It took us a few minutes for a 2:48 film.
3. Read the automatic track anyway. YouTube keeps publishing it. Look hardest at numbers, domains, names and words with a twin that sounds the same.
4. Write the description as a summary of what the video says, with the important facts in words, and add chapters: the first at 0:00, at least three, each at least ten seconds long.
5. Give the video a page of your own with the transcript on it, as HTML people can read. Collapsing it behind a "read the transcript" toggle is fine; leaving it out isn't.
6. Mark it up with VideoObject: name, thumbnailUrl and uploadDate, then description, duration and embedUrl. Add transcript if you like, but as well as the visible text, not instead of it.
7. Let the search crawlers reach the page. Googlebot, OAI-SearchBot, Claude-SearchBot and PerplexityBot all need to fetch it, and Google needs to be allowed to show it with a snippet. If you're not sure, check whether your site is blocking AI crawlers.
8. List the page in llms.txt with a one-line description of what the video says, and add the channel to sameAs in identity.json.
9. Put on-screen facts into words. Go through the video once with the sound off and list everything it shows but doesn't say. Each item needs a home in the description, the transcript or the page.

Step nine is the one I'd have skipped before doing this. It found three facts in a film I'd watched dozens of times.

What nobody has said yet about video in AI answers

The honest list of what's unknown about video AI visibility is longer than the documented one.

• Whether answer engines read captions when they build an answer. Gemini Notebook does, on request. Nothing says that AI Overviews, AI Mode, ChatGPT search, Perplexity or Claude do.
• Which track wins when there are two. Our film has a checked track and a wrong one, both public.
• Whether any of them watch the frames outside a request to analyse a particular video.
• Whether anyone reads VideoObject.transcript. It's valid, it's cheap, and no engine has said it's used.
• Who gets the credit. When a video is on YouTube and embedded on your page, nobody has said whether a citation goes to the watch page, your page, or neither.
• Whether chapters affect citation. The timestamps in Google's citations suggest moments matter. That's a correlation in one study, not a rule.

And our own limits, which are bigger than any of those. This is one film, about one product, a day old when this was written. We have not seen any AI assistant cite it, describe it or repeat anything from it. Nothing above shows that the setup works; it shows what an engine would have to work with. It's also a case where the facts in the film are written down elsewhere, on the product's own site, so even a correct answer may not have come from the video at all.

How we will measure the Meet Pip film, and when we'll report back

Measuring AI answers is hard for the reasons we set out in the Mailbox MCP discovery case study: a mention in an answer often leaves no trace at all. So the plan leans on something the film gives us that most videos don't, which is its own errors.

• Canary phrases. "a,000 calls a day", "mailboxmcp.com" and "filed in scent" exist only in YouTube's automatic track. If an AI answer ever repeats one of them, it read that track. The on-screen facts aren't clean tests, because they're on the product site too.
• The same questions, repeated. A fixed set of prompts about Mailbox MCP and its film, asked in Google AI Mode, AI Overviews, Gemini, ChatGPT search, Perplexity and Claude, several times each, because answers vary from one run to the next. We've written before about why one-off prompt tests mislead.
• Gemini Notebook import, once the 72-hour window has passed, to see which caption track it takes.
• YouTube Studio traffic sources and Search Console for this page and the product site, read for anything arriving from AI surfaces.

It's a small test of video AI visibility, on one film, and we'll treat it that way. We'll do a first check in late November and publish the full results here, as a dated update to this article, by 31 January 2027. If the canaries never appear and nothing cites the film, that's a result too, and it'll be reported the same way.

Your own site's inputs are easier to check than a video's. The free AI Visibility Checker tests whether a site publishes consistent, machine-readable facts that AI systems can find, which is what this site's definition of AI Visibility Checking covers. If your site already publishes AI Discovery Files, you can submit it to the directory for free verification, and the validation guide sets out what can be tested and what can't. The 365i Web Design team builds sites with these files in from the start.

Frequently asked questions