Google’s John Mueller has directly questioned the growing trend of creating special LLM-only Markdown or JSON pages designed purely for AI models. He says LLMs already understand HTML perfectly well, and publishers shouldn’t build pages “that no user sees.”
The idea sounds logical at first. After all, large language models thrive on clean, structured inputs.
So wouldn’t giving them simplified “shadow pages” improve how AI reads your content? Would it help content surface better in AI answers or citations?
Should developers start optimizing specifically for AI bots instead of humans?
This week, Mueller made Google’s position clear and his comments position a very different picture from what many SEOs feared or hoped for.
What Sparked the Debate About Creating LLM-Only Pages?
The conversation began when SEO expert Lily Ray posted a question on Bluesky.
She asked whether Google had any perspective on websites that were starting to create “separate markdown / JSON pages for LLMs” and serve those URLs directly to AI bots.
This reflects a larger trend happening inside SEO circles:
Publishers, worried that AI search experiences (like ChatGPT, Gemini, or Perplexity) may not read or parse their websites.
And this is the same way Googlebot does, experimenting with alternate, ultra-simple versions of their content.
Some are creating:
- Markdown-only copies
- JSON versions
- AI-optimized summaries
- Simplified “shadow” pages
These are not intended for users, only for LLMs. The assumption?
Cleaner formats = better AI interpretation = better visibility in AI search.
But Mueller’s reply pushed back strongly against this entire concept.
What Did John Mueller Say About LLM-Optimized Markdown Pages?
Mueller responded candidly, saying he was not aware of anything on Google’s side that would require such pages or benefit from them.
His reasoning was straightforward:
“LLMs have trained on, read & parsed normal web pages since the beginning. It seems that they have no problems dealing with HTML.”
In other words, LLMs do not struggle with HTML. They have been trained on trillions of tokens of standard HTML across the public web.
So why would they suddenly need alternate file formats?
Mueller then added the key question:
“Why would they want to see a page that no user sees?”
This is the core of Google’s argument. If an LLM is meant to model human knowledge and language, why should it be fed content that humans themselves never encounter?
Mueller’s point suggests that creating bot-only experiences goes against how search engines and modern AI models are designed.
Would Markdown Make It Easier for AI to Understand Content?
Lily Ray followed up with a logical question: even if HTML is fine, wouldn’t a cleaner, simpler file format help AI “get to the key points faster”?
Mueller wasn’t convinced. He said that if file formats significantly impacted AI quality, companies like Google, OpenAI, Anthropic, or Meta would publicly tell publishers to use those formats.
His argument was blunt:
“AI companies are not really known for being shy.”
Mueller believes that if Markdown, only versions improved AI answers, the LLM providers themselves would already be loudly recommending it.
Instead, we have heard nothing from them about switching formats only about improving clarity, structure and reducing heavy JavaScript.
Do LLMs Really Struggle With HTML, or Is This an SEO Myth?
According to Mueller, LLMs don’t struggle with HTML at all. He acknowledged that some pages perform better for AI systems than others but he made an important distinction:
It is not about file format but about how readable the content is, whether by humans or bots.
Then he added one caveat: “Excluding JS which still seems hard for many of these systems.”
This is where the nuance lies. LLMs struggle with:
- Heavy client-side rendering
- Dynamic content hidden behind JS
- React or Vue apps without server-side rendering
- Content that requires interaction to load
But plain HTML? LLMs have zero trouble with it. So the problem is not that AI needs Markdown, it is that AI struggles with JavaScript-heavy setups.
What About Structured Data or JSON Feeds & Do They Matter Here?
Yes, but only when the platform explicitly requests them.
In the Bluesky discussion, Matt Wright pointed out that OpenAI’s eCommerce product feeds already use JSON schemas.
These allow ChatGPT to ingest, organize and display product data reliably.
This is different from creating Markdown pages “just in case.”
In the OpenAI example, the platform provides an official specification, and sites follow it.
The conversation also referenced Chris Long’s observation from LinkedIn: editorial sites using product schemas are more likely to appear in ChatGPT citations.
This reinforces a clearer rule: Structured data matters when a platform publishes requirements.
Shadow pages matter when… someone in an SEO forum suggests it.
Only one of these is a valid optimization approach.
So Should Websites Create LLM-Optimized Pages Today?
Mueller’s comments suggest a strong no.
From Google’s perspective: LLMs understand HTML
- AI companies have not requested alternate formats
- File formats don’t control quality of AI answers
- Bot-only pages violate long-standing SEO best practices
Instead, Google wants publishers to focus on:
Clean HTML
- Reduced JavaScript dependencies
- Clear content structure
- Proper metadata
- Using schema where appropriate
The idea of LLM-only content is still speculative, experimental and unsupported by any major AI platform.
Why Does This Debate Matter for SEO and AI Search?
Because it signals where SEO energy should and should not be spent. SEOs are increasingly concerned about:
AI Overviews
- AI Mode ranking factors
- ChatGPT citations
- Perplexity answers
- Data ingestion rules
- AI-only crawling behavior
But without clear documentation from LLM providers, the risk is that teams invest heavily in tactics that do not move the needle.
Mueller’s response serves as a grounding point: HTML is enough. Structure is enough. Schema is enough. Clear content is enough.
Until AI companies publish formal ingestion rules, chasing alternate formats is premature.
What Should Publishers Actually Do Today to Prepare for LLM Search?
The conversation points toward practical action, not speculation. For now, the best approach is to:
- Keep HTML clean and accessible
- Limit dependency on JavaScript for key content
- Use structured data where platforms provide schemas
- Focus on semantic clarity
- Improve information architecture
- LLMs don’t need Markdown.
They need clarity and that starts with improving the pages users already see.
Key Takeaways
- Google’s John Mueller says LLMs don’t need Markdown-only or JSON-only bot pages.
- LLMs have always been trained on normal HTML, so alternate formats offer no clear benefit.
- Mueller believes AI companies would openly request specific formats if they mattered.
- Heavy JavaScript remains an issue for LLM parsing, not HTML itself.
- Structured data matters when platforms publish official specs (e.g., OpenAI eCommerce feeds).
Dipti Arora
AuthorDipti Arora is a Senior Content Writer with over seven years of experience creating impactful content across Digital Marketing, SEO, technology, and business domains. She has a strong background in managing news verticals and delivering editorial excellence. Dipti has contributed to leading publications such as The Times of India and CEO News, where her research-driven storytelling and ability to simplify complex subjects have consistently stood out. She is passionate about crafting content that informs, engages, and drives meaningful results.