In the space of a week, Google made the same point twice. On its Search Off the Record podcast, John Mueller and Martin Splitt took apart the idea that stripped-down content files help AI find your site. And in a quiet update to its generative AI optimization guide, Google spelled it out in writing: you do not need llms.txt, markdown, or special AI text files to show up in Google Search, including its AI features, because Search does not use them.
For anyone who has been pitched an “AI SEO” package built on exactly those files, this is the part worth reading twice.
What Google actually said
Two separate statements, one message.
The help-doc update is the blunt one. Google added a line clarifying that maintaining an llms.txt file will not help or hurt your rankings, because Google Search ignores it. It added a second clarification that you do not need machine-readable files, AI text files, markup, or markdown to appear in Search or its generative AI results. Google has never been a fan of llms.txt anyway. It has compared the file to the old keyword meta tag, pointed out that no major system actually reads it, and at one point suggested you may as well noindex it.
The podcast covered the other half: markdown vs HTML. The pitch for markdown is that it strips a page down to clean content with fewer tokens, which feels friendlier to a language model. Splitt’s counter was that the “extra” stuff markdown removes, the links, the navigation, the headers, the structure, is exactly what tells a search engine how a page fits into the rest of your site and the rest of the web. Strip it out and you have damaged discovery, which is how crawlers find your other pages in the first place.
Mueller added the practical kill shot. Crawlers have decades of practice reading HTML, and converting HTML to plain text is trivial. So the core selling point of markdown for SEO, that it makes content easier for machines to parse, solves a problem that does not exist.
The parallel-site trap
Here is the part that should stop a few roadmap decisions.
The popular move is to generate a second, markdown or llms.txt version of your site to feed the machines. That means maintaining two versions of everything. Double the surface area for bugs, double the things that can drift out of sync.
And there is a quieter risk. If your hidden machine version breaks, no human ever sees it, so nobody reports it. Meanwhile an automated system can keep crawling the broken copy. You have built a failure mode that only robots can find.
This is the same instinct we keep flagging in our AI website audits: teams bolting on machinery to chase AI visibility while the fundamentals underneath go unchecked.
What this does not mean
Worth being precise, because the hype will overcorrect in both directions.
Google’s statement is about Google Search and its own AI surfaces. It is not a claim that markdown or llms.txt are useless everywhere. Some tools and agents do read those files, and Google itself says it is fine to keep them if a system you care about consumes them. Splitt allowed that markdown has real uses. The point is narrower and more useful: do not build a parallel file strategy expecting Google ranking or AI search gains, because that specific payoff is not coming.
If your goal is to be understood by machines, the lever is not a separate file. It is structured, clean HTML with schema markup where it helps, which is readable by browsers, screen readers, crawlers, and LLMs at the same time.
What actually moves AI visibility
Strip away the file debate and the fundamentals are unchanged, and a little boring.
Publish normal HTML pages. That is still the prerequisite for being crawled, indexed, and cited, by traditional search and AI systems alike. Then make the structure legible: clear headings, sensible internal links, and a site architecture that shows how pages relate. The link and navigation context that markdown throws away is the same context that helps you get discovered, which is why we treat internal linking and topic clusters as a ranking lever, not decoration.
Everything real about optimizing for AI answers, the work behind generative engine optimization and content built for AI, sits on top of that foundation. None of it requires a second copy of your website written in markdown.
So before you greenlight an AI-SEO project whose centerpiece is a pile of llms.txt and markdown files, ask the vendor one question: which system, specifically, is going to read this, and what does it change. If the answer is “Google,” the honest answer is nothing.
If you would rather spend that budget on the HTML and structure that actually gets you found, that is the work we do with B2B SaaS teams through our SEO services. Grab 30 minutes.