Google’s Mueller Questions Need For LLM-Only Markdown Pages via @sejournal, @MattGSouthern

Google Search Advocate John Mueller has pushed back on the idea of building separate Markdown or JSON pages just for large language models (LLMs), saying he doesn’t see why LLMs would need pages that no one else sees.

The discussion started when Lily Ray asked on Bluesky about “creating separate markdown / JSON pages for LLMs and serving those URLs to bots,” and whether Google could share its perspective.

Ray asked:

Not sure if you can answer, but starting to hear a lot about creating separate markdown / JSON pages for LLMs and serving those URLs to bots. Can you share Googleʼs perspective on this?

The question draws attention to a developing trend where publishers create “shadow” copies of important in formats that are easier for AI systems to understand.

There’s a more active discussion on this topic happening on X.

This has been the hot topic lately, I’ve been getting pitched by companies who do this https://t.co/rVnbPKUxZj

— Lily Ray 😏 (@lilyraynyc) November 23, 2025

What Mueller Said About LLM-Only Pages

Mueller replied that he isn’t aware of anything on Google’s side that would call for this kind of setup.

He notes that LLMs have worked with regular web pages from the beginning:

I’m not aware of anything in that regard. In my POV, LLMs have trained on – read & parsed – normal web pages since the beginning, it seems a given that they have no problems dealing with HTML. Why would they want to see a page that no user sees? And, if they check for equivalence, why not use HTML?

When Ray followed up about whether a separate format might help “expedite getting key points across to LLMs quickly,” Mueller argued that if file formats made a meaningful difference, you would likely hear that directly from the companies running those systems.

Mueller added:

If those creating and running these systems knew they could create better responses from sites with specific file formats, I expect they would be very vocal about that. AI companies aren’t really known for being shy.

He said some pages may still work better for AI systems than others, but he doesn’t think that comes down to HTML versus Markdown:

That said I can imagine some pages working better for users and some better for AI systems, but I doubt that’s due to the file format, and it’s definitely not generalizable to everything. (Excluding JS which still seems hard for many of these systems).”

Taken together, Mueller’s comments suggest that, from Google’s point of view, you don’t need to create bot-only Markdown or JSON clones of existing pages just to be understood by LLMs.

How Structured Data Fits In

Other individuals in the thread drew a line between speculative “shadow” formats and cases where AI platforms have clearly defined feed requirements.

A reply from Matt Wright pointed to OpenAI’s eCommerce product feeds as an example where JSON schemas matter.

In that context, a defined spec governs how ChatGPT ingests and displays product data. Wright explains:

Interestingly, the OpenAI eCommerce product feeds are live: JSON schemas appear to have a key role in AI search already.

That example supports the idea that structured feeds and schemas are most important when a platform publishes a spec and asks you to use it.

Additionally, Wright points to a thread on LinkedIn where Chris Long observed that “editorial sites using product schemas, tend to get included in ChatGPT citations.”

Why This Matters

If you’re questioning whether to build “LLM-optimized” Markdown or JSON versions of your content, this exchange can help steer you back to the basics.

Mueller’s comments reinforce that LLMs have long been able to read and parse standard HTML.

For most sites, it’s more productive to keep improving speed, readability, and content structure on the pages you already have, and to implement schema where there’s clear platform guidance.

At the same time, the Bluesky thread shows that AI-specific formats are starting to emerge in narrow areas such as product feeds. Those are worth tracking, but they’re tied to explicit integrations, not a blanket rule that markdown is better for LLMs.

Looking Ahead

The conversation highlights how fast AI-driven search changes are turning into technical requests for SEO and dev teams, often before there is documentation to support them.

Until LLM providers publish more concrete guidelines, this thread points you back to work you can justify today: keep your HTML clean, reduce unnecessary JavaScript where it makes content hard to parse, and use structured data where platforms have clearly documented schemas.

Featured Image: Roman Samborskyi/Shutterstock

LLMs.txt Shows No Clear Effect On AI Citations, Based On 300k Domains via @sejournal, @MattGSouthern

A new analysis from SE Ranking suggests the llms.txt file isn’t delivering measurable benefits yet.

After examining roughly 300,000 domains, the company found no relationship between having llms.txt and how often a domain is cited in major LLM answers.

What The Data Says

Adoption Is Thin

SE Ranking’s crawl found llms.txt on 10.13% of domains. In other words, nearly nine out of ten sites they measured haven’t implemented it.

That low usage matters because the format is sometimes described as an emerging baseline for AI visibility. The data instead shows scattered experimentation. SE Ranking says adoption is fairly even across traffic tiers and not concentrated among the biggest brands.

High-traffic sites were slightly less likely to use the file than mid-tier websites in their dataset.

No Measurable Link To LLM Citations

To assess whether the llms.txt file affects AI visibility, SE Ranking analyzed domain-level citation frequency across responses from prominent LLMs. They employed statistical correlation tests and an XGBoost model to determine the extent to which each factor contributed to citations.

The main finding was that removing the llms.txt feature actually improved the model’s accuracy. SE Ranking concludes that llms.txt “doesn’t seem to directly impact AI citation frequency. At least not yet.”

Additionally, they found no significant correlation between citations and the file using simpler statistical methods.

How This Squares With Platform Guidance

SE Ranking notes that its results align with public platform guidance. But it’s important to be precise about what is confirmed.

Google hasn’t indicated that llms.txt is used as a signal in AI Overviews or AI Mode. In its AI search guidance, Google frames it as an evolution of Search that continues to rely on its existing Search systems and signals, without mentioning llms.txt as an input.

OpenAI’s crawler documentation similarly focuses on robots.txt controls. OpenAI recommends allowing OAI-SearchBot in robots.txt to support discovery for its search features, but does not say llms.txt affects ranking or citations.

SE Ranking also notes that some SEO logs show GPTBot occasionally fetching llms.txt files, though they say it doesn’t happen often and does not appear tied to citation outcomes.

Taken together, the dataset suggests that even if some models retrieve the file, it’s not influencing citation behavior at scale right now.

What This Means For You

If you want a clean, low-risk way to prepare for possible future adoption, adding llms.txt is easy and unlikely to cause technical harm.

But if the goal is a near-term visibility bump in AI answers, the data says you shouldn’t expect one.

That puts llms.txt in the same category as other early AI-visibility tactics. Reasonable to test if it fits your workflow, but not something to sell internally as a proven lever.

Featured Image: Mameraman/Shutterstock

New Data Finds Gap Between Google Rankings And LLM Citations via @sejournal, @MattGSouthern

Large language models cite sources differently than Google ranks them.

Search Atlas, an SEO software company, compared citations from OpenAI’s GPT, Google’s Gemini, and Perplexity against Google search results.

The analysis of 18,377 matched queries finds a gap between traditional search visibility and AI platform citations.

Here’s an overview of the key differences Search Atlas found.

Perplexity Is Closest To Search

Perplexity performs live web retrieval, so you would expect its citations to look more like search results. The study supports that.

Across the dataset, Perplexity showed a median domain overlap of around 25–30% with Google results. Median URL overlap was close to 20%. In total, Perplexity shared 18,549 domains with Google, representing about 43% of the domains it cited.

ChatGPT And Gemini Are More Selective

ChatGPT showed much lower overlap with Google. Its median domain overlap stayed around 10–15%. The model shared 1,503 domains with Google, accounting for about 21% of its cited domains. URL matches typically remained below 10%.

Gemini behaved less consistently. Some responses had almost no overlap with search results. Others lined up more closely. Overall, Gemini shared just 160 domains with Google, representing about 4% of the domains that appeared in Google’s results, even though those domains made up 28% of Gemini’s citations.

What The Numbers Mean For Visibility

Ranking in Google doesn’t guarantee LLM citations. This report suggests the systems draw from the web in different ways.

Perplexity’s architecture actively searches the web and its citation patterns more closely track traditional search rankings. If your site already ranks well in Google, you are more likely to see similar visibility in Perplexity answers.

ChatGPT and Gemini rely more on pre-trained knowledge and selective retrieval. They cite a narrower set of sources and are less tied to current rankings. URL-level matches with Google are low for both.

Study Limitations

The dataset heavily favored Perplexity. It accounted for 89% of matched queries, with OpenAI at 8% and Gemini at 3%.

Researchers matched queries using semantic similarity scoring. Paired queries expressed similar information needs but were not identical user searches. The threshold was 82% similarity using OpenAI’s embedding model.

The two-month window provides a recent snapshot only. Longer timeframes would be needed to see whether the same overlap patterns hold over time.

Looking Ahead

For retrieval-based systems like Perplexity, traditional SEO signals and overall domain strength are likely to matter more for visibility.

For reasoning-focused models like ChatGPT and Gemini, those signals may have less direct influence on which sources appear in answers.

Featured Image: Ascannio/Shutterstock

Google Brings Gemini 3 To Search’s AI Mode via @sejournal, @MattGSouthern

Google has integrated Gemini 3 in Search’s AI Mode. This marks the first time Google has shipped a Gemini model to Search on its release date.

Google AI Pro and Ultra subscribers in the U.S. can access Gemini 3 Pro by selecting “Thinking” from the model dropdown in AI Mode.

Robby Stein, VP and GM of Google Search, wrote on X:

“Gemini 3, our most intelligent model, is landing in Google Search today – starting with AI Mode. Excited that this is the first time we’re shipping a new Gemini model in Search on day one.”

(1/5) Gemini 3, our most intelligent model, is landing in Google Search today – starting with AI Mode. Excited that this is the first time we’re shipping a new Gemini model in Search on day one! 🚀

In Search, Gemini 3 with generative layouts will make it easy to get a rich… pic.twitter.com/iFoXJSMw37

— Robby Stein (@rmstein) November 18, 2025

Google plans to expand Gemini 3 in AI Mode to all U.S. users soon, with higher usage limits for Pro and Ultra subscribers.

What’s New

Search Updates

Google describes Gemini 3 as a model with state-of-the-art reasoning and deep multimodal understanding.

In the context of Search, it’s designed to explain advanced concepts, work through complex questions, and support interactive visuals that run directly inside AI Mode responses.

With Gemini 3 in place, Google says AI Mode has effectively re-architected what a “helpful response” looks like.

Stein explains:

“Gemini 3 is also making Search smarter by re-architecting what a helpful response looks like. With new generative UI capabilities, Gemini 3 in AI Mode can now dynamically create the overall response layout when it responds to your query – completely on the fly.”

Instead of only returning a block of text, AI Mode can design a response layout tailored to your query. That includes deciding when to surface images, tables, or other structured elements so the answer is clearer and easier to work with.

In the coming weeks, Google will add automatic model selection, Stein continues:

“Search will intelligently route tough questions in AI Mode and AI Overviews to our frontier model, while continuing to use faster models for simpler tasks.”

Enhanced Query Fan-Out

Gemini 3 upgrades Google’s query fan-out technique.

According to Stein, Search can now issue more related searches in parallel and better interpret what you’re trying to do.

A potential benefit, Stein adds, is that Google may find content it previously missed:

“It now performs more and much smarter searches because Gemini 3 better understands you. That means Search can now surface even more relevant web content for your specific question.”

Generative UI

Gemini 3 in AI Mode introduces generative UI features that build dynamic visual layouts around your query.

The model analyzes your question and constructs a custom response using visual elements such as images, tables, and grids. When an interactive tool would help, Gemini 3 can generate a small app in real time and embed it directly in the answer.

Examples from Google’s announcement include:

• An interactive physics simulation for exploring the three-body problem
• A custom mortgage loan calculator that lets you compare different options and estimate long-term savings

All of these responses include prominent links to high-quality content across the web so you can click through to source material.

See a demonstration in Google’s launch video below:

Why This Matters

Gemini 3 changes how your content is discovered and used in AI Mode. With deeper query fan-out, Google can access more pages per question, which might influence which sites are cited or linked during long, complex searches.

The updated layouts and interactive features change how links appear on your screen. On-page tools, explainers, and visualizations could now compete directly with Google’s own interface.

As Gemini 3 becomes available to more people, it will be important to watch how your content is shown or referenced in AI responses, in addition to traditional search rankings.

Looking Ahead

Google says it will continue refining these updates based on feedback as more people try the new tools. Automatic model selection is set to arrive in the coming weeks for Google AI Pro and Ultra subscribers in the U.S., with broader U.S. access to Gemini 3 in AI Mode planned but not yet scheduled.