Google’s Mueller Says llms.txt Can’t Help LLMs Differentiate Sites

Google’s John Mueller argued that LLM techniques can’t use information like llms.txt to resolve which web sites to floor for a given question.

He made the feedback on a current episode of Search Off the Record, the podcast from Google’s Search Relations staff.

His remark factors to a broader sign drawback, not simply intentional gaming. Even a well-written llms.txt file remains to be self-reported info from the location that desires to be chosen.

For discovery, Mueller pointed again to regular HTML pages and inside hyperlinks.

What Mueller Stated

The dialog began with a query about whether or not publishers ought to convert web sites to Markdown for LLMs. Mueller and co-host Martin Splitt agreed that HTML remains to be the inspiration for crawling and discovery.

The dialogue obtained particular when Mueller turned to llms.txt. He described the invention use case as a lifeless finish:

“It’s principally you’re telling these techniques, like, I’ve one of the best web site ever. And listed below are the entire pages that everybody should go to. And it’s essential to purchase all of my merchandise or no matter you set in there. So in LLM system, it principally, by design, can’t belief what’s right here as a means of differentiating between totally different web sites.”

His argument comes right down to differentiating. If websites use llms.txt to advertise themselves, the information could make comparable claims. An LLM deciding which web site finest solutions a question nonetheless wants one other option to differentiate between them.

What ‘By Design’ May Imply

“By design” may imply two various things, and Mueller didn’t make clear which.

One studying is architectural. LLM techniques consider net content material and might’t use self-reported information when choosing sources.

The opposite studying treats it as a sign drawback. Self-reported alerts lose worth when everybody gives them. Meta key phrases stopped working for a similar purpose. Each web site stuffed them, and search engines like google couldn’t extract a helpful rating sign.

Each readings attain the identical conclusion on discovery. However they suggest various things about whether or not the limitation may change over time.

The place Mueller Sees A Position

Mueller didn’t reject all makes use of of llms.txt. He carved out one case the place it may assist:

“If somebody is already in your web site, perhaps some sort of automated system is useful.”

He used the instance of an agent attempting to purchase {a photograph} from a selected web site. The LLM would go to the location and search for directions on full the acquisition.

The argument splits discovery from navigation. llms.txt can’t assist an LLM select which web site to go to. Nevertheless it may assist as soon as the agent is already there, like a retailer listing for somebody who already walked in.

Past The Gaming Argument

Mueller has called building Markdown pages for bots “a stupid idea”. He’s additionally compared llms.txt to the keywords meta tag.

SEJ’s Roger Montti wrote that llms.txt is “inherently untrustworthy” as a result of nothing stops web site homeowners from including self-serving content material. SE Rating’s analysis of 300,000 domains discovered no hyperlink between llms.txt adoption and quotation frequency in LLM solutions.

These arguments centered on what occurs when individuals recreation the information. Mueller’s podcast remark provides the nuance that there’s no mechanism inside the information to assist an LLM decide one web site over one other.

Why This Issues

The gaming argument in opposition to llms.txt has all the time had a counterargument obtainable. Platforms may be taught to penalize manipulation, the best way search engines like google dealt with spammy structured knowledge.

The differentiation argument leaves a tougher drawback. Penalizing manipulation could handle abuse, however it doesn’t clarify how self-reported information assist an LLM select one web site over one other. Your most correct llms.txt file nonetheless can’t inform an LLM to select your web site over a competitor’s.

Wanting Forward

Requirements for a way brokers navigate websites haven’t settled but, Mueller acknowledged. He talked about WebMCP alongside different file sorts underneath dialogue.

None have turn into an ordinary. By his estimate, it may take six months to a 12 months, or longer, for agentic techniques to decide on a format. The invention layer, the place HTML and inside linking already work, isn’t a part of that dialogue.