AI Search and B2B Marketing: What Google Actually Documents

What Google documents about AI Overviews and AI Mode

Google documents two generative AI features in Search and describes both in the same terms as ordinary ranking. AI Overviews are meant to help users get to the gist of a complicated topic quickly and to act as a jumping off point to explore links. AI Mode is described as helpful where further exploration, reasoning, or complex comparisons are needed.

Two mechanisms are named. The first is retrieval-augmented generation (RAG — the model answers from passages retrieved out of an index), which Google says relies on "our core Search ranking systems to retrieve relevant, up-to-date web pages from our Search index." The second is query fan-out: Google issues multiple related searches across subtopics and data sources, which lets the features show a wider and more diverse set of links than a classic results page.

Google's own summary: "the best practices for SEO continue to be relevant because our generative AI features on Google Search are rooted in our core Search ranking and quality systems." If a page cannot be crawled, indexed and ranked, it cannot be retrieved. If it can, no additional step is documented.

Why the missing click lands differently on a B2B site

Consumer sites lose volume when an answer replaces a click. B2B sites lose something harder to see, because volume was never the point.

A company selling industrial automation has an addressable market of hundreds of organisations. Its load-bearing pages are definitional and comparative — what is X, how does X differ from Y, who are the vendors in X. Those are exactly the queries where AI features appear, and exactly the content most B2B teams spent five years building.

Fan-out cuts the other way. Because Google issues its own related searches, a page can be retrieved for a question the visitor never typed. Keyword-level thinking degrades here; topic-level depth does not.

Then there is the committee. Gartner's May 2025 research, from a survey of 632 B2B buyers fielded in August and September 2024, describes buying teams of five to 16 people across as many as four functions, looping rather than progressing in order. Several of them research with no intention of filling in a form, and an answer delivered without a session is invisible in your analytics.

There is no special file, markup, or format

From Google's guidance on optimising for generative AI features:

• "You don't need to create new machine readable files, AI text files, markup, or Markdown to appear in generative AI search."
• "Structured data isn't required for generative AI search, and there's no special schema.org markup you need to add."

Google also names what is not needed: chunking content into passages, and rewriting content just for AI systems. And from the AI features documentation: "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary."

Any service pitching a proprietary file, a required markup type, or a reformatting exercise is therefore selling something Google documents as unnecessary. The inverse claim needs care too: not required is not the same as does nothing. But no primary source establishes that any schema type causes inclusion in an AI feature.

Recent history should make anyone wary of markup-led strategy. FAQ rich results stopped appearing in Google Search on 7 May 2026, having been restricted since August 2023 to authoritative government or health sites. The markup is not harmful, but it now produces no documented Search feature for anybody.

The controls that exist, and the trade-off inside them

There is no AI-Overviews opt-out tag. The controls are the ones that already existed:

• nosnippet, data-nosnippet and max-snippet — snippet controls, at page or element level.
• noindex — removes the page from Search entirely.
• robots.txt directives for Googlebot — controls crawling.
• Google-Extended — documented as limiting AI training and grounding in some of Google's other systems, not as an AI Overviews control.

The trade-off is usually left out of vendor summaries. To keep a page out of AI Overviews specifically, the documented levers are the snippet controls or noindex, and nosnippet also removes your ordinary search snippet. Nothing documented lets you appear normally in classic Search while being excluded from AI Overviews.

For most B2B sites the right answer is to use none of this. The genuine cases are narrow: licensed data you have no right to have summarised, or a single passage that is dangerous out of context — where data-nosnippet on that element is proportionate and a sitewide block is not. Applying nosnippet broadly costs you snippets, clicks, and the retrieval you were trying to protect.

How to measure it without inventing numbers

Two first-party sources exist, and neither gives most people what they want.

In Search Console, traffic from AI features is included in the overall Performance report under the Web search type. In June 2026 Google announced Search generative AI performance reports in Search Console. Per the announcement and the trade reporting around it, the view is impressions-first — impressions, pages, countries, devices and dates; AI Overviews and AI Mode not segmented from each other; no click data; a subset of sites first. Treat that detail as reported rather than settled, and check the live help documentation before building a board report around a column.

The absence of clicks has one blunt consequence: no click-through rate for AI features can be calculated from Google's own data. Any AI CTR figure in a deck was built from assumptions.

In GA4, assistants are a first-class channel. The default channel group includes AI Assistants, which Google describes as covering arrivals from sources like ChatGPT, Gemini, Deepseek, Copilot or Grok, assigned automatically since May 2026 when the referrer matches Google's list. Google does not publish that list, so treat the channel as a floor rather than a count.

llms.txt, and the appetite for a new file to add

llms.txt is an independent proposal by Jeremy Howard, hosted on GitHub for community input. It is not a W3C standard, not an IETF standard, and not named in any search or assistant vendor's documentation that could be located.

What it specifies is modest: a markdown file at /llms.txt with an H1 naming the site as its only required element, an optional summary, and H2-delimited sections listing URLs to markdown resources.

Whether any major vendor consumes it in production is unverified; no documentation confirming production consumption could be found. Google's position covers the idea by description rather than by name — new machine readable files, AI text files, markup or Markdown are not needed for its generative AI features.

So the honest recommendation is unexciting. The file costs an afternoon, does no harm, and may turn out to matter. Presenting it as a requirement or a ranking factor is not defensible, and a supplier who does tells you something about the rest of their advice.

What is actually contested, said plainly

A reference page earns its keep by naming the claims that do not survive contact with a primary source. Here there are five.

• The size of AI's traffic impact. Nobody has a primary-sourced figure. Vendor estimates in 2026 run from AI traffic being roughly one percent of the web, to ChatGPT holding about 92 percent share of AI referrals, to growth above 500 percent. Those measure three different quantities, and none of the studies is independent.
• Formatting as a cause of inclusion. Answer-first paragraphs, question-shaped headings and passage chunking are sold as AI optimisation. Google says chunking is unnecessary and no rewriting for AI systems is needed. No replicable evidence shows any format causes retrieval.
• Structured data as an AI ranking factor. Contested and unproven; Google documents no AI-specific requirement.
• AI visibility scores and citation share. Vendor-defined denominators, unpublished sampling, no audit. Useful for tracking your own movement, not for measuring a market.
• llms.txt influencing Google. Not supportable.

Not contested: the features run on core ranking, fan-out surfaces you for queries nobody typed, and the technical requirements are the ordinary ones.

Be the source a system would want to quote

The defensible strategy is the one that was already defensible, which is why it is hard to sell. If retrieval runs through core ranking and no special preparation is documented, the work is to be the most quotable source on a narrow set of subjects your buyers ask about.

• Publish information that exists nowhere else — your own pricing structure explained, your own failure-rate data, your own reading of a standard. A restatement of what six other sites say has no reason to be retrieved.
• Answer the question in the first sentence, because a reader benefits and a retrieval system has less work to do. Do it because it is good writing, not because a vendor promised a mechanism.
• Name the author and date the page. Attribution is how humans and systems both decide whether to trust a claim.
• Keep the boring floor solid — crawlable, indexable, fast enough, not duplicated across four URLs. Most lost visibility originates here.

Watch one thing above all this year: whether Search Console's generative AI reporting gains click data. Until it does, treat every external estimate of your AI traffic as a hypothesis.

Frequently Asked Questions