Mueller’s markdown warning is the whole GEO argument

John Mueller spent one sentence this week saying something the GEO tool industry has spent eighteen months trying to avoid.

The context: someone on Bluesky pointed out that sites are now shipping parallel markdown versions of their HTML for AI agents to consume, while still failing basic accessibility — no proper heading structure, no landmarks, screen readers still locked out. Mueller's response, per Search Engine Journal's coverage, was that a well-designed website naturally functions for AI agents, search engines, LLMs, and — most importantly — people. Adding a separate agent-friendly version, he said, is technical debt you'll eventually have to rip out.

That's it. That's the whole argument against half the GEO playbook being sold to UK businesses right now.

I've been writing about the same tension for weeks — most recently on Lighthouse failing working llms.txt files over syntax and on Safari's MCP server as the quiet half of agentic browsing. Mueller's just made it official from inside Google. Worth pausing on why that matters.

The two-site fallacy

The GEO-tool pitch, stripped of its marketing, goes like this: LLMs and AI agents parse content differently from humans, so you need a machine-facing version of your site — a markdown mirror, an llms.txt manifest, a schema-heavy shadow — optimised for the way agents read. Sometimes it's positioned as a supplement. Sometimes it's positioned as strategic infrastructure. Either way, the assumption is that your normal site isn't enough.

The two-site solution is a symptom of not having done the one-site work.

Mueller's answer is that this assumption is wrong on its own terms.

If your HTML is well-structured — proper headings, semantic markup, clean information architecture, accessible landmarks — then agents can already parse it. They're not blocked. They don't need a separate lane. The reason people think they need a markdown mirror is that their primary site is structurally poor, and rather than fix it, they've built a second version that hides the problem.

The two-site solution is a symptom of not having done the one-site work.

Why this maps directly onto what the accessibility community has been saying for a decade

The Bluesky post that started this whole thread wasn't from an SEO. It was someone pointing out that sites now bother to ship markdown versions for AI agents while still ignoring the accessibility requirements they've been asked to meet for ten-plus years. Same underlying problem — content that only works for one kind of consumer — dressed up in new language.

The accessibility community's answer to "we need a separate accessible version" has always been: no, you need one properly structured site. The version-for-screen-readers approach became infamous because it always rotted. Content updates went to the main site, not the mirror. The accessible version drifted out of date. Screen reader users got a worse experience than they would have got from a properly built primary site.

Every argument against parallel accessible sites applies word-for-word to parallel machine-facing sites. The content will drift. The versions will diverge. Fixes will happen in one place and not the other. The engineering cost compounds.

The GEO tool vendors have a problem now

Not every GEO vendor is selling a parallel-site product. Some are doing genuinely useful work on citation monitoring, log analysis, brand tracking across LLM answers. Fine. But a meaningful chunk of the pitch deck economy is built on the premise that you need to *do something different for AI* at the content layer — different markup, different formats, different files, different pages.

Mueller just said, from Google's Search Relations seat, that this is the wrong direction of travel.

He wasn't speaking for OpenAI. He wasn't speaking for Anthropic. But it doesn't matter — the argument is structural, not brand-specific. Every serious LLM crawler in 2026 can parse well-structured HTML. Perplexity can. ChatGPT can. Claude can. Google's own AI Overviews obviously can. The bottleneck isn't format. The bottleneck is whether your content is worth citing.

If your HTML is bad, a markdown mirror is a workaround. If your HTML is good, a markdown mirror is redundant.

Either way, the mirror isn't the answer.

What the honest playbook actually looks like

I've been saying versions of this for months, but Mueller's given me cleaner language for it, so here it is again.

The machine-readable web isn't a separate destination. It's a property of your primary site being properly built. That means:

• Semantic HTML — headings that describe hierarchy, not headings picked for visual weight.
• Landmarks — main, nav, article, aside — used correctly, so screen readers *and* agents can identify sections.
• Structured data where it earns its place — product, article, organisation, FAQ where genuinely relevant. Not schema for the sake of schema.
• Speed and Core Web Vitals. Slow sites get partially indexed by everyone, humans and machines alike.
• Content that answers the question the URL implies it answers.

That list is boring. It has been the list for fifteen years. What's changed is that the cost of ignoring it has gone up, because now you've got LLM crawlers pulling from your pages to feed citations, and they hit the same structural problems your users have been complaining about since 2015.

The GEO industry keeps trying to add a new layer on top when what most sites need is to fix the layer they've already got.

The counterargument, taken seriously

There is a version of the parallel-content argument that isn't stupid, and it's worth stating fairly.

That version says: yes, well-structured HTML is table stakes, but LLMs benefit from *additional* signals — clean markdown extracts of key content, explicit llms.txt manifests pointing to canonical documents, schema that goes further than what humans need. Not a replacement site. A supplement. An extra layer of clarity for machine consumers who have less context than a human browsing your homepage.

I have some sympathy for this. llms.txt as a lightweight discovery aid isn't crazy — it's cheap to ship, it doesn't create maintenance drag if you generate it from your existing content, and if it helps a crawler find the right pages, fine. The Lighthouse piece I linked above was frustrated at the audit implementation, not the file itself.

Where the argument breaks is when the supplement becomes the strategy. When teams start writing content-for-humans and content-for-agents as separate outputs. When the markdown version has different language than the HTML. When the "AI-optimised" version starts including keyword stuffing dressed up as machine-readability. That's when Mueller's warning bites. That's when technical debt starts accumulating.

The line is: supplement is fine, substitute is not. And most of what's being sold as GEO infrastructure is drifting toward substitute.

What this means for buyers

If you're on the client side and someone is pitching you a machine-facing content layer, ask two questions.

First: is my current site well-structured? Get an honest audit. Not a GEO audit — a plain-boring accessibility and semantic-HTML audit. If the answer is no, fix that before you add anything on top. You will get more citation lift from fixing your heading structure than from generating a thousand markdown mirrors.

Second: if my current site *is* well-structured, what specifically does the proposed layer add that agents can't get from the primary site? Make the vendor answer that in concrete terms. "Better parseability" isn't an answer. "Faster retrieval" isn't an answer unless it's measured. A real answer sounds like "your product pages don't currently expose price and availability in a machine-readable way and we'll add ProductGroup schema to fix that." That's a supplement to good HTML, not a replacement for it.

Most of the time, once you ask those two questions honestly, the answer is that the money would be better spent on the primary site.

The measurement problem sits underneath all of this

The reason the parallel-content pitch works, commercially, is that nobody can measure whether it's working. I wrote about this at length in the Google reviews bug piece — the AI citation monitoring space is still years behind where it needs to be. When you can't measure whether markdown mirrors are actually lifting citation share, you can't disprove the pitch. You just keep buying the layer and hoping.

Mueller can't fix that. Nobody in Google can. But he can say, from a position of authority, that the parallel-site approach isn't the load-bearing answer. And this week, quietly, he did.

The industry should listen. Most of it won't, because too much revenue depends on not listening. But the buyers can. The buyers can ask better questions. The buyers can decline to fund the two-site fallacy.

The web that works for AI agents is the web that already works for people, done properly. That's been the answer for a decade. It's still the answer. The acronym on the invoice doesn't change it.