On 15 May 2026 Google published a guide called Optimizing your website for generative AI features on Google Search. It carries a section headed Mythbusting generative AI search: what you don't need to do. Five tactics sit in it: llms.txt files and other special machine readable markup, chunking content into small pieces, rewriting content specifically for AI systems, seeking inauthentic mentions across the web, and overfocusing on structured data.
One of the five now has outside measurement attached to it. Ahrefs published a study on 15 June 2026 covering 137,210 domains and reported that 97 percent of the llms.txt files in that population received no requests at all during May 2026 (Ahrefs, 15 June 2026). The Google list governs Google Search, AI Overviews and AI Mode included. It says nothing about ChatGPT, Claude or Perplexity.
The Revenue Signal arrives on Thursdays. One decision, one verified source, one move you can make before the next issue. Signing up delivers both the newsletter and the Executive AI ROI Scorecard, which you can take in about five minutes. Take the scorecard and subscribe.
Somewhere in your accounts payable there is a line item for AI search work. It might be called AEO, short for answer engine optimisation, or GEO for generative engine optimisation, or simply AI visibility. Nobody in the room can say what it bought. The vendor sends a score each month, the score moves, and no one has ever traced the score back to a server log.
That invoice now has a public reference point. This article covers what Google published, what the outside data does and does not prove, and the three questions worth asking before you approve the next one.
Google published Optimizing your website for generative AI features on Google Search through Google Search Central on 15 May 2026, announced on its Search Central blog. It sits in the official documentation rather than in a blog post or a conference talk, which matters because documentation is the thing Google maintains and points people back to.
Most of it restates positions Google has held for a while. Generative AI features in Search run on the same core ranking and quality systems as the rest of Search. A page has to be indexed and eligible to appear with a snippet before it can appear in a generative AI feature at all. Crawlability, page experience and duplicate content still matter for the same reasons they always did.
The guide also takes a position on the vocabulary. It states that from Google Search's point of view, optimising for generative AI search is optimising for the search experience, and is therefore still SEO. It then points readers to Google's own guidance on how to evaluate third party SEO advice and services.
What is new is the mythbusting section. Google names tactics that are sold in the market and tells site owners they can ignore them for Google Search. Naming tactics is unusual. It is the reason this document is worth an executive's attention rather than only a marketing team's.
The five items, in the order Google lists them.
Read item five carefully before anyone deletes anything. Google's word is overfocusing, not abandoning. The instruction is to stop treating schema as an AI visibility lever, while keeping it for the job it actually does.
Because someone measured it from outside Google.
Ahrefs published a study on 15 June 2026, written by Louise Linehan and Xibeijia Guan. It covered all 137,210 domains in Ahrefs Web Analytics that received traffic during May 2026. The researchers checked each domain root for an llms.txt file returning a 200 response, confirmed the file was real Markdown rather than an error page dressed up as one, then used server log data to classify every request to those files by user agent.
Three findings carry the weight.
Twenty eight percent of the domains published an llms.txt file, about 38,000 sites. Of those, Ahrefs found 97 percent received zero requests during May 2026. No bots, no humans, nothing. All measured traffic went to the remaining three percent, roughly 1,100 domains and about 22,000 requests in total.
Inside that small pool, the composition is the interesting part. Twelve percent of requests came from the industry studying the file rather than consuming it: AEO and GEO scoring tools, llms.txt validators and directories, and research crawlers. AI retrieval bots, the ones that fetch pages to answer a live user query in an AI search product, accounted for 1.1 percent of requests in the same Ahrefs data. Ahrefs also reported that Slackbot, the link preview bot in a chat app, fetched these files more often than PerplexityBot did.
The last finding closes off a common argument. Ahrefs looked at requests for llms.txt files that do not exist, the 404s. The AI bot share of those was zero. AI systems are not probing for the file and finding it missing. They are not looking at all.
The study does not prove that llms.txt is dead everywhere, and it does not describe the open web. Four limits travel with the study.
Scope note. Ahrefs sells the bot analytics product this data comes from, and its customers skew technical and SEO aware, so the 28 percent adoption figure is an upper bound rather than a read on the open web. The behavioural findings cover only the three percent of files that received any request, about 1,100 domains. A request proves that something asked for the file, not that anything read it or acted on it, which makes every figure in the study a ceiling on real consumption. The study measured the index file at the site root and nothing else. Ahrefs did not test whether the files were well formed against the specification.
It also does not prove that the other four items on Google's list are worthless. Nobody has published a comparable log study on chunking or on AI specific rewriting. Those four rest on Google's own statement about its own systems, which is strong evidence about Google and no evidence at all about anyone else. What makes a site invisible to those other engines is a separate problem with separate causes.
And it does not prove that llms.txt is useless everywhere. The study found that AI coding agents were the most active identifiable readers. If your buyers use coding agents against your documentation, the file has a plausible job. That is a narrow, specific case, and it is not the case the tactic is usually sold on.
No, and treating it that way would be the wrong lesson.
Google's guide states that creating content people find unique and useful will influence a site's presence in generative AI search more than anything else it suggests. It gives examples of what that means: a genuine point of view, first hand experience, material that is not a restatement of what is already on the internet. That is work, and it costs money, and it is the thing Google puts first.
The guide also keeps the technical floor in place. A page that is not indexed and not eligible for a snippet cannot appear in a generative AI feature. Sites also have to be included in Search generative AI features in Search Console to be eligible for display. Crawl behaviour, JavaScript handling and page experience still apply.
So the finding is narrower than the headline suggests. A named set of tactics does nothing on Google Search, the largest surface, while invoices for those tactics keep going out. The category is real. Part of the delivery inside it is not.
A score is a number a vendor computes. A measurement is a record of something that happened, which you can inspect.
Google's guide includes a direct caution here. It tells readers to be wary of third party tools that promise ranking success or claim to use internal Google metrics, and states plainly that no third party tool has access to its internal ranking or AI systems. It adds that such tools are fine to use if they help your workflow, as long as their advice is checked against the official guidance.
There are two records you can inspect without buying anything.
The first is your own server log. It shows which user agents requested which pages, and how often. That is what the Ahrefs study of 137,210 domains is built on, and there is no version of it a vendor can compute for you without touching it. Anyone reporting on your AI visibility should be able to produce the log lines behind the report.
The second is the Generative AI performance report in Google Search Console, which shows how content is performing in generative AI features on Google Search and Discover. It is first party, it is free, and it covers the surface Google's list is about.
Neither one covers ChatGPT, Claude or Perplexity. For those, the honest position today is that measurement is partial, and any vendor claiming otherwise is describing a model of those systems rather than the systems themselves.
We publish our own method for that partial measurement, including what it cannot do. It runs buyer intent prompts across three engines, repeats each one, and then opens every cited page to check whether it actually backs the claim it was cited for. That is a sample rather than a log, and it is described that way on purpose.
Three questions, in this order. None of them requires a technical background to ask, and each one has a right answer that a competent supplier can give in a sentence.
Ask the person who owns the website, not the supplier. You are looking for line items covering llms.txt or AI specific markup, content chunking, rewriting for AI systems, mention building programmes, or schema work sold as AI visibility. Work sold under those names is work Google says does nothing for its own surface.
Which AI user agents fetched which pages, on which dates, how often. A supplier who cannot produce those log lines is showing you a model, not a measurement. It may still be a useful model. It is not the same object and it should not be priced as though it were.
Unique, first hand content is the item Google names first. It is also the hardest to fake and the slowest to produce, which is why it rarely appears on an invoice from a supplier selling monthly scores. If the answer is a list of posts that any competitor could have published word for word, the budget is in the wrong place regardless of what the score says.
Google's guide ends with five takeaways: keep applying foundational SEO practice to generative AI search, build a clear technical structure, create non commodity content that carries genuine expertise, monitor performance in Search Console, and stay aware of agentic experiences as browser agents start visiting sites on behalf of users.
The content item is the one that takes real work. We catalogued the structural choices that separate pages which get cited from pages which do not, and almost none of them are tactics you can buy by the month.
None of it is exotic. That is the point of the document. The tactics Google named are the ones invented to fill the gap between how AI search felt new and how it actually works, and the gap turned out to be smaller than the market that grew inside it.
No. An llms.txt file is a markdown index file placed at a site's root, summarising what the site is and linking its main pages. Google states that maintaining one will neither harm nor help visibility or rankings in Google Search, because Google Search ignores it. Ahrefs notes a separate consideration: a research crawler in its dataset was studying these files as a prompt injection route, because agents are built to trust what they ingest. If you keep a file, treat it like code, control who can edit it, and review anything a platform generates on your behalf.
No. Structured data is machine readable markup that describes what a page contains. Google says it is not required for generative AI search and there is no special markup to add for it, while also saying it is worth continuing as part of overall SEO because it supports eligibility for rich results. The change is in what you expect from it, not in whether you keep it.
No. The guide covers Google Search and its generative AI features, including AI Overviews and AI Mode. It makes no claim about other AI systems. On the llms.txt item specifically, the Ahrefs log data does cover the named retrieval bots for those platforms, and found them requesting these files at 1.1 percent of measured requests (Ahrefs, 15 June 2026). The other four items have no equivalent outside evidence.
Use the Generative AI performance report in Google Search Console, which reports on how content performs in generative AI features on Google Search and Discover. Read it alongside your own server logs, which show which AI user agents fetched which pages. Both are records of events rather than computed scores.
The Executive AI ROI Scorecard is a self assessment for executives across four areas of AI spend, and section two of it is AI visibility. Twenty statements, about five minutes, and a written report that tells you which of these questions your company can already answer and which it cannot.
Take the Executive AI ROI ScorecardThe scorecard arrives with The Revenue Signal, published on Thursdays. One decision, one verified source, one move. You are told about both before you sign up, and you can leave either at any time.
Both sources checked live on 28 August 2026.
