It's a ten-minute file that costs nothing to publish and, as far as any AI engine has confirmed, changes nothing when you do. Here's what it actually is, and where the evidence stands.
Jeremy Howard, co-founder of Answer.AI and fast.ai, published the proposal on September 3, 2024, at answer.ai and llmstxt.org. His argument: a language model's context window is too small to hold a whole website, and turning a page's HTML into clean model-readable text is laborious - navigation, ads, and scripts get in the way of the actual content.
llms.txt is his answer - a markdown file at yoursite.com/llms.txt, short, with a one-line summary of the site and a curated list of links to the pages a model should read, grouped and described in plain sentences. Some sites pair it with an llms-full.txt, the same content pages exported in full. It's deliberately low-ceremony: no new schema, no registration body, nothing to install.
The three files solve three different problems, and having one says nothing about whether you have the other two.
Adoption skews hard toward developer-tool and API companies. Anthropic's own documentation, Stripe, Vercel, and a long tail of dev-tool vendors ship an llms.txt or llms-full.txt for their docs. That's a specific and explainable pattern: these are companies whose product is consumed by other developers building AI-integrated tools, where a clean, curated docs export saves real engineering time regardless of whether any consumer-facing AI engine ever reads it for citations.
Outside that niche, adoption is thin. The proposal has never been submitted to the IETF or W3C for ratification, and in its own framing it remains a convention, not a standard - something a site can choose to publish, not something any crawler is obligated to respect.
Google is the one company that has answered this directly, and answered no. Gary Illyes said at Search Central Live APAC in July 2025 that Google "doesn't support llms.txt and isn't planning to." John Mueller repeated the point on Reddit in 2026, comparing it to the old meta keywords tag, and called any claimed benefit "purely speculative" - the file has existed for years and, in his words, no AI system uses it. Google's own AI optimization guide, updated mid-2026, tells site owners plainly that they don't need to create llms.txt or any similar machine-readable file to appear in Google Search, including its generative features.
Nobody else has said differently, in either direction. OpenAI, Anthropic, and Perplexity have not published a statement confirming their retrieval or answer systems read llms.txt files on third-party sites. Anthropic and others publishing their own llms.txt for their own documentation is a separate fact from any engine consuming other sites' llms.txt files during retrieval - conflating the two is the most common mistake in guides written about this file.
One nuance worth naming honestly, because it cuts against the flat "nobody at Google cares" read: Chrome's Lighthouse tool added an llms.txt presence check in version 13.3, released May 2026, inside a new "Agentic Browsing" audit category alongside checks for emerging AI-agent tooling like WebMCP. That's the Chrome team checking whether the file exists - it isn't Google Search using it, and Lighthouse marks the check "not applicable" rather than "fail" when the file is missing, because even the team that added it treats it as optional. Two parts of the same company disagree with each other, which is a reasonable summary of how unsettled this still is.
Because the cost is close to zero and the downside is close to zero. Writing an llms.txt takes minutes, touches nothing about how your site is indexed, needs no developer time to speak of, and can't get you penalized - a short markdown file at your root is invisible to everything except whatever bothers to fetch it. If a future version of some engine starts reading it, you're already there. If none ever do, you spent ten minutes.
There's a second reason we mention it to clients that has nothing to do with any engine reading it. Naming the handful of pages that actually represent your business, in one clear sentence each, is the same discipline behind good page structure and clean entity signals - it just happens to also produce a file with no confirmed reader. Do it for that reason. Don't do it because someone told you it moves a citation count.
Keep the format plain: an H1 with your company name, a one-line summary of what you do, then H2 groupings of links with a one-sentence description each - Docs, Products, Pricing, About, whatever your actual structure is. Point it at pages that already state a clear answer near the top; a curated link to a page that buries its own point doesn't fix the underlying problem, it just curates the problem. Skip marketing copy, skip anything your sitemap already covers in more exhaustive form, and don't bother with an llms-full.txt unless you genuinely have a documentation surface worth exporting whole.
A check earns a place on a client's readiness sheet only if it traces to something an engine has documented or something we've watched move a real answer. llms.txt currently traces to neither - the one company that has spoken clearly about it has said, twice, that it doesn't use it. We cover the checks that do belong on that sheet - access, rendering, structure, entity evidence - in the readiness pass, and where a file like this sits relative to the actual work of getting recommended in how to navigate AEO. Publish llms.txt if you want. Just don't pay anyone to optimize it.