Google just debunked GEO and llms.txt. Here's why I keep using them anyway.

The Google post nobody saw coming

On May 15, 2026, John Mueller publishes a post on the Google Search Central blog with a deceptively quiet title: “A new resource for optimizing for AI features in Search.” Behind the title, a bomb for the entire industry that has spent its weeks talking about AEO, GEO, llms.txt, and brand mention signals.

Google ships an official guide. And in that guide, its position is sharp, almost dismissive: optimizing for generative AI in Google Search is just SEO. Not a new craft. Not a new file. Not a new discipline.

While part of the SEO ecosystem was selling GEO audits at $5,000 a pop, Google publishes a doc page that literally says: “you don’t need to create new machine-readable files, AI text files, markup, or Markdown to appear in generative AI search.”

I’ve been sitting in front of that guide since last night. I have a llms.txt on claudehub.fr. I have structured data. I’ve spent entire sessions optimizing for Perplexity and ChatGPT citations. And the first thing I think reading Mueller’s post is: he’s right on 80% of the points, and he completely misses the rest.

This article is my honest attempt to separate what Google rightly tore down from what they leave unmentioned because it’s not their problem.

What Google actually published

Let’s start cold. The official guide, Optimize Your Website for Generative AI Features in Google Search, is structured in five sections:

Is SEO still relevant for generative AI search?
Apply foundational SEO best practices to generative AI search
Mythbusting: what you don’t need to do
Explore agentic experiences
Next steps: what to focus on

The heart of Google’s position fits in one Mueller quote: “From Google Search’s perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO.”

Translated to operational terms, here’s the list of what is pointless if you want to rank in AI Overviews or get cited in Google’s generative answers:

No need for a llms.txt, llm.txt, or any specialized markdown file
No need for “AI-ready” markup
No need to chunk your content into pre-sliced passages for LLMs
No need to rewrite your articles with a tone or format tuned for models
No need to chase inauthentic “brand mentions” across the web
No structured data specifically for AI search visibility

And the list of what actually matters, per Google:

Unique content based on first-hand expertise, not generic
Pages indexed and eligible for standard snippets
Public crawlability
Clean semantic HTML
Properly served JS
Mobile and latency
Limited duplicate content
For commerce and local: clean Merchant Center and Google Business Profile

Read that list with no context and you read the same thing you’d have read in 2018, 2022, or 2024. That’s exactly the message Google wants to push.

Where Google is right

My starting position is that Mueller isn’t wrong. And I think plenty of people in the SEO/GEO space will have trouble swallowing that sentence, so let’s be direct about what holds in his argument.

First: AI Overviews eats from the standard SERP

When Google generates an AI answer in the search results, the system doesn’t look into a parallel index reserved for “AI-optimized” sites. It pulls from its main index. Pages cited in AI Overviews are pages that already rank in the standard SERP, or pages that could rank.

So if your foundational SEO is bad, your AI Overview visibility will be bad. If your SEO is good, AI visibility follows naturally, at least on classic commercial and informational queries. That continuity is consistent with everything observed since SGE then AI Overviews launched.

Second: llms.txt has zero official status

Let’s be clear on this one because that’s where GEO marketing went off the rails. llms.txt is a community proposal. Nobody adopted it officially. Not Google, not OpenAI, not Anthropic, not Perplexity. None of those players published documentation committing their engine to use the file as a ranking or inclusion signal.

For a year, the SEO ecosystem sold llms.txt as if it were a standard. It isn’t. It’s a file you publish, that nobody is obligated to read, and that changes nothing for your visibility in any AI search product today. Google says it explicitly. And they’re right to say it, because the confusion cost time and money to many teams.

Third: “chunking for LLMs” is a lazy myth

The idea that you have to break your articles into 200-word passages with headings every paragraph because “LLMs read in chunks” has no serious technical foundation for public SEO. Models doing retrieval-augmented generation on the web use their own chunking pipelines. You don’t control how Perplexity or ChatGPT slices your content. You control the quality of that content, its logical structure, its human readability. That’s it.

Writing good paragraphs, clear sentences, descriptive headings: it worked before LLMs, it works with LLMs, it’ll work after. “AEO chunking” is an obsession that displaces honest editorial work.

Fourth: obsessive brand mention farming is counterproductive

Part of the GEO industry sold the idea that you must multiply brand mentions on Reddit, Quora, Hacker News, paid blogs, to train LLMs to associate your brand with a topic. Google says it politely: “avoid chasing inauthentic mentions.”

Translated: if you spam forums with generated reviews to train models, you risk more than wasting your time. You risk being detected, deprioritized, or even penalized on channels where reputation is built over time. Bad ROI plus bad ethical signal.

So yes, on these four points, Google usefully clarifies the debate. Mueller cleans up ground that needed it.

But Google is answering the wrong question

Here’s where I diverge from the official line. Google’s guide answers a specific question: how do I appear in the AI answers generated by Google Search? That’s a legitimate question, it’s even the main question for 99% of e-commerce sites, media outlets, and local businesses.

It is not my question.

My question, on claudehub.fr, is different. And I think my question is also the question of a growing number of content creators, builders, and experts publishing on the web in 2026.

My question is: how do I make my content usable by AIs, agents, and assistants, without them having to scrape my site on every interaction?

It’s not the same question. Google doesn’t answer it. Google has no reason to answer it, because Google sells ad slots on SERPs, not interaction protocols between agents and websites.

The real use of GEO in 2026

When a user opens Claude, ChatGPT, or Perplexity and types “explain Simon Willison’s concept of agentic engineering,” three scenarios unfold:

Scenario 1: the LLM answers from its internal knowledge. If Claude Hub has been ingested in the training set, my content can shape the answer. It’s slow (training cycle), opaque (I have no control), and doesn’t change based on what I publish today.

Scenario 2: the LLM does a live web search and scrapes the top SERP pages. Here, it’s classic SEO. Mueller is right. If my article is well-positioned on Google or Bing, it’ll be read and potentially cited. No GEO magic.

Scenario 3: an autonomous agent interacts with my site as a resource. This is where everything changes. An agent that wants to build a complete answer about Claude Code, compare multiple sources, follow internal links, understand my site’s structure, doesn’t operate like a Google Search user. It operates as an information consumer.

That third scenario is where files like llms.txt start to make sense. Not to rank. To reduce the cost of using your site by agents.

llms.txt: not a ranking factor, an accessibility protocol

Let’s reframe llms.txt correctly, because all the commercial GEO pitch sold it wrong. llms.txt isn’t an SEO visibility signal. It’s a navigation file for agents.

When an AI agent lands on claudehub.fr with no context, it has to figure out:

What is this site?
How is it organized?
What are the canonical pages on each topic?
How do I navigate efficiently without scraping everything?

If I have nothing, the agent has to do a full crawl, compare dozens of pages, guess which URL is authoritative on a given topic, handle duplicates, follow pagination. That’s expensive in tokens, time, and bandwidth. And the agent can get it wrong, pick the wrong page, cite a secondary URL instead of the main article.

With a well-built llms.txt, I solve part of those problems:

List of the site’s main sections
Pointers to canonical articles
Short description of each resource
Clear hierarchy

It doesn’t help ranking. It helps agentic accessibility. Very different.

The Anthropic, GitHub, Stripe case

If you look at the sites that have implemented llms.txt over the past six months, a pattern emerges. They’re mostly tech companies with a stake in their documentation being usable by AI assistants.

Anthropic publishes llms.txt on docs.anthropic.com. GitHub has an equivalent. Stripe and Cloudflare too. These teams don’t expect to rank better on Google because of that file. They’re building infrastructure for a world where developers use Claude and ChatGPT to search their docs, and where answer quality depends on the agent’s ability to find the right endpoint in the doc.

It’s product-driven use, not SEO-driven. And it has clear commercial sense: if Claude answers better on the Anthropic API, it makes life easier for developers using the API. That’s conversion by usage experience, not SEO traffic.

The Claude Hub case

I can tell mine because it’s more modest but representative.

Claude Hub publishes long-form articles on Claude Code, the Anthropic ecosystem, and LLM usage patterns in dev. My target audience massively includes people who ask their questions directly to Claude, ChatGPT, Perplexity. Before coming to my site. Sometimes without ever coming to my site.

My objective isn’t only to rank in Google AI Overviews. It’s that when someone asks Claude “what are the good French blogs on Claude Code,” my name comes up. It’s that when a dev asks ChatGPT “how does Claude Code’s memory system work,” my article on memory gets cited, ideally with a link.

For that, the work splits into two layers:

Classic SEO layer. Long articles, real expertise, Article structured data, clean internal linking, fast mobile, clean Search Console. Google is right: it’s the baseline. Miss it and nothing happens.

GEO/agentic layer. A llms.txt pointing to my canonical articles by theme. Stable URLs. Explicit headings that describe content, not just editorial hooks. Exportable MDX format. Rigorous sitemap. Explicit reformulation of the questions my readers ask.

This second layer won’t boost my SEO traffic. It increases the probability that an agent passing through my site understands my structure, finds the relevant article, and cites the right URL rather than a tag page or a duplicate.

Over the last month, my GA4 stats confirm something interesting: sessions from chatgpt.com and claude.ai appeared for the first time (5 and 2 sessions respectively). Ridiculous in volume. Massive in signal. Six months earlier, those sources didn’t exist in my analytics. The origin page for those sessions, in 100% of observed cases, is an article I structured well, with a descriptive H1 and an intro that directly answers a question.

That doesn’t prove llms.txt serves any purpose. It proves that content structure matters for agents. And llms.txt is just one more layer of that structure.

Verdict: SEO first, GEO if you want AIs to work with you

I’ll close where Google should have started, in my view. Mueller isn’t wrong to say that for AI Overviews, SEO is enough. Where his guide falls short is that it mentions agentic experiences only in an “explore” section, like a future curiosity.

Yet agentic experiences are already here. They’re in Claude, ChatGPT Pro, Perplexity, in the custom agents companies deploy internally to summarize their feed, in the MCP servers that expose content to LLMs.

If you’re a mattress retail chain, Google’s guide is enough. Do local SEO, put your products in Merchant Center, ignore GEO.

If you’re a blog, a documentation platform, a technical SaaS, a newsletter, a B2B media outlet, or if you publish expertise your readers will consume via an AI before coming to you, then:

SEO first, no compromise. Bad SEO breaks everything.
GEO as a second layer. Not to rank. To reduce the friction of using your content for agents.
llms.txt if you have the time. Honestly, it’s half a day of work. Low but positive ROI, trivial to maintain.
Article and FAQ structured data. Not because Google requires it for AI, but because it serves classic SEO AND agent accessibility.
Stable URLs and descriptive headings. Most important and least sexy thing.
No artificial chunking. Write well. Period.
No brand mention farming. Build real reputation.

Google’s position is useful to frame expectations. It prevents companies from wasting 20% of their marketing budget on “GEO audits” with no technical basis. But it doesn’t exhaust the question. It answers how to appear in AI Overviews. It doesn’t answer how to be a site useful to agents.

And these two questions will keep diverging. AI Overviews is a Google product. The agentic ecosystem is being built in parallel, with its own dynamics. Anthropic, OpenAI, Mistral, MCP players, internal agent builders, don’t read the world the same way Google Search does.

The right reflex in 2026 isn’t to choose between SEO and GEO. It’s to do both in the right order, with the right expectations, and without paying a consultant $5,000 for a GEO audit that boils down to “add a llms.txt.”

You can do it yourself in half a day. And if you do, do it for the reason that actually matters: not for Google, but for the agents that will consume your content more and more often without ever going through Google.

Going further

Sources: Google Search Central - A new resource for optimizing for AI features · Official guide - Optimize Your Website for Generative AI Features in Google Search · llms.txt specification - llmstxt.org