Video built for answer engines and AI search: what the engines actually publish

Google defines the two mechanisms behind a generated answer in its own words, on that same July 2026 page. Retrieval-augmented generation, "also known as grounding", relies "on our core Search ranking systems to retrieve relevant, up-to-date web pages from our Search index". Query fan-out is "A set of concurrent, related queries generated by the model to request more information and fetch additional relevant search results". Neither description names a media type of any kind.

The same mechanism was described in ordinary words by people under oath, and that account is not marketing copy. In the memorandum opinion filed in the District of Columbia on 2 September 2025, grounding is recorded as "anchoring the output of a model on factual information or [an] external database", with Google's Head of Search describing it as a model "ask[ing] for web results . . . to maybe fill in information it doesn't have, to correct information".

That opinion also records how the stores are arranged. The court limited its definition of Search Index to "databases that store and organize information about websites and their content that is crawled from the web", citing testimony "discussing Google's other various indexes that include data not crawled from the web, including video, images, and verticals". Scope that exactly. It establishes that Google maintains indexes separate from the web search index and that video is among them. It does not establish that a video watch page cannot be in the web index, and Google's own documentation says watch pages are indexed.

Among the SEO fundamentals Google publishes as still applying to its AI features are four that, read in order, are the practical spine of everything below. Ensure "that crawling is allowed in robots.txt, and by any CDN or hosting infrastructure". Make "sure that important content is available in textual form". Support "your textual content with high-quality images and videos, when applicable". Make "sure your structured data matches the visible text on the page". Video is named there as support for text, not as the thing being read.

Two independent counts were run for this article. One covered twenty-four publisher-facing documents from twelve companies. The other covered seventeen documents from six publishers across four source classes. Seven documents were shared between them, and all were read on 7 and 8 September 2026. In every document that describes how a generated answer is produced, video appears in the body zero times or once, and in each of the handful of exceptions the sentence is about video appearing in an answer or agreeing with the page. Every count here is a body count over extracted text, with navigation, header, footer and aside removed.

That is a statement about documents, not about machines. No operator publishes that its answer system reads a publisher's video file, and the absence of a published statement is not proof that it does not happen. Nothing here says an engine ignores video.

Microsoft publishes one publisher-facing instruction about video and AI answers, in a post dated 10 February 2026: "Reduce ambiguity across formats: Align text, images, and video so they consistently represent the same entities, products, or concepts." It is the exception that shows the shape of the rest, because it asks for agreement between the film and the page rather than for anything to be read out of the film.

Google's structured data guidelines, last updated 10 July 2026, say the same thing from the other side. "Don't mark up content that is not visible to readers of the page." And among the reasons Google lists for a feature not appearing: "The content referred to by the structured data is hidden from the user." A chapter list in markup has to correspond to a chapter list a person can see. That is a page and copy decision taken at delivery, not a template field.

The standard corrects the instinct most people bring to this file. RFC 9309, Standards Track, published September 2022, states that "These rules are not a form of access authorization" and that "The Robots Exclusion Protocol is not a substitute for valid content security measures", and frames itself as rules "that crawlers are requested to honor". Over its full text, video, transcript, caption and image each appear zero times. Rules match paths, not content. A site owner cannot express "you may not use my film" in this file, only "you may not fetch this path".

Two behaviours in that standard are commonly held backwards. A 4xx response on robots.txt means "the crawler MAY access any resources on the server". A 5xx means "the crawler MUST assume complete disallow". A broken server locks crawlers out and a missing file lets them all in.

Six of the eight companies below publish a training robot and an answering robot separately, each on its own documentation. The other two publish an answering robot only.

Common Crawl is the counter-example that makes the split visible. It publishes one bot, CCBot, with User-agent: CCBot and Disallow: / as the only published control, and no separation between training use and answering use.

The single most useful correction in this section is that the file is not a list of independent switches. Apple publishes that "If robots instructions don't mention Applebot but mention Googlebot, the Apple robot will follow Googlebot instructions." Amazon publishes a wider version of the same rule, scoped to other search bots rather than to Googlebot: where a file does not mention Amzn-SearchBot but allows other search bots, it "will crawl in accordance with the robots.txt directives given to other search bots". Brave publishes that its crawler "does not advertise a differentiated user agent" and that "if a domain or page is not crawlable by Googlebot, then Brave Search's bot will not crawl it either". Three companies inherit the rules written for somebody else.

Google itself now operates three separate switches over three different things, and they are frequently confused for one another.

The legal layer gets three sentences and no more. A United States court found in September 2025 that publishers "cannot ... opt out of Google's use of their content to fine-tune Google's Search models or for display in AI Overviews" except by opting out of being crawled altogether, which it called "not a tenable choice". On 3 June 2026 the United Kingdom's Competition and Markets Authority imposed a Publisher Conduct Requirement obliging Google to "provide publishers with effective controls to withhold their Search Content from being used in generative AI services and features", at directory level by 3 December 2026 and page level by 3 March 2027. Google's own Search Console control shipped worldwide on 31 August 2026, and this article does not adjudicate between the three.

A worked example, as a fact about a file. On 8 September 2026 grolez.com's robots.txt disallowed nine agents, carried Content-Signal: search=yes,ai-train=no,use=reference, set no ai-input signal, which is the signal Cloudflare's vocabulary defines as covering generative AI search answers, and named no rule at all for OAI-SearchBot, PerplexityBot, Claude-SearchBot, Googlebot or Googlebot-Video, all of which therefore fall under Allow: /. Nothing is said here about why it is set that way.

Open your own file. It takes ten seconds and costs nothing. The reason to look is that six of the eight companies above document a training robot and an answering robot separately, and three companies, two of those eight and Brave besides, inherit whatever was written for somebody else's crawler, so a file can express a preference nobody sat down and gave. It is not an outcome, because no outcome is established.

Every body text read for this article was swept for a digit attached to a percentage, a multiple, a period or an outcome word. The total of quantified outcome claims found was zero.

Google's nearest thing to an outcome claim carries no figure either: "We've seen that when people click from search results pages with AI Overviews, these clicks are higher quality (meaning, users are more likely to spend more time on the site)." No number, no period, no method, no sample. The absence is the point, rather than the claim.

The figure the industry quotes most often is in the public court record, redacted: "At present, more than __% of Search queries trigger an AI Overviews response". The one party that knows the number gave it to a court under seal, so any figure a seller quotes for it did not come from Google. The single public quantified figure that party did put on the record is about its own query volume rather than about anybody's traffic: "Since its introduction, Google Search queries in the United States have increased 1.5 to 2%."

Google has published a position on the circulating third-party traffic figures, and it is a position rather than proof. It calls them "often based on flawed methodologies, isolated examples, or traffic changes that occurred prior to the roll out of AI features in Search", and Google is an interested party in saying so. The two sides of that argument are a platform with the data and no published number, and sellers with a number and no data.

There is also a reason not to trust a single screenshot. Peer-reviewed work published in July 2026, comparing Google organic search against five generative search systems from Google, OpenAI and Perplexity, reports that "the outputs of generative search can vary across time and executions, raising new challenges for robustness". That paper mentions video once across 22 pages, in a bibliography entry.

The largest open question is the one the pitch is usually built on. Whether any answer engine ever ingests, watches or cites a brand's video file itself is not documented in either direction by anything retrieved for this article. We have established no figure for what a transcript, a chapter list, structured data or a robots.txt change does to citations, impressions or traffic, because no engine publishes one.

Does ChatGPT or Google actually watch our video?

No operator publishes that its answer system reads a publisher's video file. Across the documents counted for this article, the ones that describe how a generated answer is produced mention video zero times or once, and the exceptions are about video appearing in an answer or agreeing with the page. That is a fact about documents on named dates. The absence of a published statement is not proof either way.

Should we buy an answer-engine optimisation retainer?

Read Google's own sentence first, dated 10 July 2026: "You don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search (including its generative AI capabilities), as Google Search itself doesn't use them." Google names AEO and GEO itself and calls the work SEO. The honest limit is that this is Google's position about Google Search, and it is not evidence about any other engine.

What is the one thing that changes what we film?

Chapters. A segment that can carry a label is a segment that answers exactly one question, and that is decided in the brief and on the day rather than in the edit. Media Fragments URI 1.0 describes a segment as an in-point and an out-point. Google's Clip wants a descriptive title for each one. YouTube requires at least three timestamps, ascending, starting at 00:00, with a ten-second minimum.

Where should the transcript go, on the page or in a file?

On the page, as readable text. That is page text of the sort Google's own fundamentals ask for, which is the only mechanism read here that a transcript actually reaches. Transcript, caption and subtitle appear zero times across five named Google video and structured-data documents. Accessibility is a separate and genuine reason to produce caption and transcript files, and the Journal covers it elsewhere as a delivery specification.

Can we stop AI answers using our site?

Google publishes three switches: robots.txt for Googlebot for crawling, Google-Extended for training and grounding, and a Search Console control for appearance in AI Overviews and AI Mode, rolled out worldwide on 31 August 2026 and taking effect "within 1-2 days after the control goes live". Each is scoped to Google. Other operators publish their own separate controls, and three companies publish that their crawler inherits rules written for somebody else's.

How much extra will all this cost us?

Nothing in our own record prices it. No line prices a transcript, a chapter set, an Open Graph block or a schema implementation, and the production bands live in the Dubai video production cost article. This article makes no offer on the subject and quotes no figure for it.

In the documents read for this article, no answer engine publishes that its answer-generating system reads your video file, and the company that owns the largest surface publishes that optimising for its generative AI features is the ordinary work of being findable. What is reachable is the page, in words, and the film supports it. The queue runs page first, video second, answer third, so a film with no page is not in it. Chapters are the one part of this that changes what gets shot, because a film can only be chaptered if it was built out of parts that survive being named. The transcript belongs in the page HTML rather than in a file only the player opens. The robots file expresses a preference about fetching rather than about use, and three companies inherit the rules written for somebody else's crawler. And nobody publishes an effect size for any of it, which is worth knowing before somebody quotes you one.