llms.txt is not a ranking signal: what actually gets you cited by AI

Every few months a new file promises to unlock AI search. The current one is llms.txt, a plain-text file you place at your site root to point AI systems at your most important content. It is a thoughtful idea from a respected technologist. It is also, as of mid-2026, inert: no major engine honors it, and there is no evidence that publishing one changes whether you are cited.

We are going to say that plainly, because the alternative is letting teams pour effort into a signal that does nothing. This post is the honest accounting: what llms.txt is, why it does not work yet, and what actually moves AI citation, each claim with a source.

What llms.txt is

llms.txt was proposed by Jeremy Howard in September 2024. The idea borrows from robots.txt: a Markdown file at the site root that lists, in a curated and machine-friendly form, the pages an LLM should read. It is well designed and well intentioned, and that is precisely what makes it dangerous. It sounds like it should work, so people assume it does.

Why it does not work (yet)

The decisive fact is that the engines have not adopted it. Google's Gary Illyes stated publicly, at Search Central Live in 2025, that Google does not use llms.txt. OpenAI, Anthropic, and Perplexity have made no public commitment to honor it. A standard is only a standard when the systems it targets read it, and none of the major ones do.

So the practical status of llms.txt is: a proposal awaiting adoption that may or may not come. Watching it costs nothing. Shipping one is harmless but does nothing. The error is paying for it, or worse, reorganizing your content strategy around it, on the belief that it earns citations. It does not.

An audit tool that scores you down for lacking an llms.txt is selling a signal that does nothing. We refuse to do that.

SEO for AI Agents, methodology

The pattern: myths that sound technical

llms.txt is the latest in a lineage of tactics that feel rigorous but rest on signals that do not exist. The same honesty applies to several others the legacy tools still score:

• Keyword density: there is no optimal percentage, and there never was. Covering the entities and questions a topic implies is what works, not hitting a keyword ratio.
• An "E-E-A-T score": E-E-A-T is a framework Google's systems approximate through many signals and that human raters use to evaluate quality. There is no single score to optimize, so a tool that hands you one is inventing a number.
• Meta keywords: ignored by Google for over a decade. Filling them in changes nothing.
• Submitting your site to hundreds of directories: low-quality citations and links that add no value and can look manipulative at scale.

The unifying lesson is that a signal is only real if the engine reads it. The way to know is to look for a primary source, the engine's own documentation, a credible statement from its team, or controlled measurement, not a confident claim in a marketing post.

What actually moves AI citation

Here is the part that is real, and it is more demanding than dropping a file at your root. The levers that measurably increase whether an engine cites you, in rough order of impact:

1. Be readable to non-JS crawlers. GPTBot, ClaudeBot, and PerplexityBot fetch raw HTML and do not run JavaScript. If your content is hydrated client-side, it is invisible to them. Server-render your primary content. This is the gate.
2. Make the answer extractable. Structure your most important answers as a question-shaped heading followed by a tight, self-contained block of roughly 40 to 170 words. Engines retrieve and compose at the passage level; lead with the answer.
3. Get corroborated. Independent research has found that AI answers lean heavily on a small set of trusted sources, with Wikipedia accounting for a large share of factual citations. A brand whose key facts agree across Wikipedia, Reddit, review sites, and its own pages is cited more than a brand twice its size that exists only on its own domain.
4. Establish your entity. Ship structured data that declares your organization and connect it to your authoritative profiles with sameAs links. A Wikidata entry, with a lower notability bar than Wikipedia, gives engines a machine-readable node to recognize.
5. Earn experience and authority signals. Named, credentialed authors and genuine first-hand experience are what the conversational engines, and Google's raters, weigh for consequential topics.

The academic backbone for the passage and corroboration levers is the 2023 paper on Generative Engine Optimization, which ran controlled experiments showing that adding cited statistics, direct quotations, and authoritative language measurably increased a source's visibility inside generative answers. That is the kind of evidence a real signal has. llms.txt has none of it.

The honest position is a competitive advantage

Telling clients the truth about llms.txt costs a little traffic from the "AI SEO hacks" search intent. It earns the trust of the operators and agencies whose careers depend on not getting fooled. We would rather have the second. Every claim we make in an audit traces to a primary source on a methodology page, and where a popular tactic does nothing, we say so and cite why.

So: skip the magic file. Spend the hour you would have spent on llms.txt verifying that ChatGPT can actually read your most important page with a JavaScript-free curl. That single check tells you more about your AI visibility than any file you could publish.