Meet an Expert

How to Build an llms.txt Agent Page AI Models Can Quote

How to build an llms.txt-style agent page of canonical brand facts that models can quote cleanly, and how it differs from an about page.

Jessica Abbadia Jessica Abbadia
Nov 13, 2026 8 min read Technology

Ask three AI assistants when your company was founded and you can get three different years. One read an old funding announcement, another misparsed a sentence on your about page that started with “we”. An llms.txt file, or any AI agent instruction page, is an attempt to fix this at the source: one plain page of canonical facts written to be lifted whole. It works only if you know who reads it and what it cannot override.

Key Takeaways

  • An llms.txt file is read at inference, when an agent fetches pages to answer a question, so it affects retrieval presence and does little for training-data presence.
  • A fact is quotable when it is one sentence, names the brand instead of using a pronoun, carries a date or unit, and still makes sense with everything around it deleted.
  • Models repeat the version of a fact that independent sources agree on, so every line in the file has to match your about page, schema markup and third-party profiles word for word.

Who Actually Reads an llms.txt File?

Jeremy Howard of Answer.AI proposed llms.txt in September 2024 as a Markdown file at a site root that gives agents a curated summary and links to clean versions of key pages. The second version of the proposal says thousands of sites now publish one, and OpenAI, Anthropic and Google’s Gemini team all publish llms.txt files for their own developer documentation. The proposal says the file was expected to matter at inference, and that is how it has been used.

Google is openly sceptical. John Mueller wrote on Bluesky in June 2025 that no AI system currently uses llms.txt, and in June 2026, as reported by Search Engine Journal, he called its value “purely speculative for now”. Google Search Central tells publishers they do not need special AI files to appear in generative search. Chrome’s Lighthouse does audit for the file, yet a missing one is marked not applicable because providing it is optional.

Set expectations accordingly. The file helps agents that fetch it while working on a user’s question, such as coding assistants and browsing agents already on your domain. Its larger value is internal: writing it forces a brand to decide what its canonical facts are.

How an Agent Page Differs From an About Page or Press Kit

An about page is written to persuade a person. It tells a founding story, states a mission and wraps facts inside paragraphs full of “we” and “our”. When a retrieval system splits that page into chunks, a sentence like “We launched it in 2018” arrives without the brand name or the product it refers to.

A press kit is written for journalists. It holds logos, bios, approved quotes and a boilerplate paragraph, often inside a PDF. The boilerplate comes closest to a canonical statement, yet it is tuned for tone and goes stale between funding rounds.

An agent page is written for a reader with no context and a limited token budget. That changes what belongs on the page. Persuasion goes. Current numbers with dates stay. So do the exact names you want repeated, disambiguation from similarly named companies, and corrections of what models commonly get wrong.

Most brands skip that last category. The reference example in the specification, written for the FastHTML project, opens its notes by listing what the library is incompatible with, including React. A brand page should do the same for discontinued products, markets you do not serve and the company you are often confused with.

The Anatomy of a Canonical Facts Block

The specification fixes the order. An H1 with the brand name is the only required element. A blockquote follows with a short summary, then free Markdown with no headings, then H2 sections holding lists of links, each with an optional note. A final section titled “Optional” holds links an agent can skip when context is tight.

The blockquote is the most important line you will write. An agent reads it first, so it should define the brand in one sentence that survives being pasted into an answer with nothing around it. The canonical facts belong in the free Markdown section beneath it, and each line should follow a few rules:

  • One claim per line, with the brand name as the subject. Never “we”, “it” or “the company”.
  • A date on anything that changes, written as “As of March 2026, [Brand] has [figure] customers.”
  • The exact form of names you want repeated, including legal entity, product casing and any former name.
  • Boundaries stated as facts: “[Brand] does not offer [service]” or “[Brand] is not affiliated with [similarly named company].”
  • Numbers identical to the ones on your site and in your schema markup.

Below the facts, the link lists should point to Markdown versions of the pages that hold detail: pricing, product documentation, leadership, locations. The proposal suggests serving these at the same URL with .md appended. The file stays small. Detail lives behind the links.

The file is also one page among hundreds that models can cite. When we tracked Moburst’s own brand, citations spread across hundreds of pages.

309
unique Moburst pages cited by AI models, alongside 42,435 total AI citations and the number 1 position in its category.See the case study.

Any one of them can contradict the file, so treat the facts block as the source every other page syncs from.

Why Verbatim Quoting Depends on Agreement Elsewhere

A model answering a question about your brand weighs every passage it has retrieved or absorbed. If your llms.txt says founded in 2014 and two business directories say 2013, the model has to choose, and a single self-published file rarely outvotes independent sources. Trust behaves like a graph of agreement across sources. One node asserting something loudly does not move it much.

Verbatim quoting follows the same logic. Models paraphrase by default. The phrasing that survives intact is the one that appears, unchanged, in many places the model considers independent.

A model repeats a sentence word for word when it has seen that same sentence, unchanged, in enough independent places to stop doubting it.

This is also the answer to Mueller’s sharpest objection: if an LLM can write your llms.txt from your HTML, an LLM could do that for itself. For a file that only restates your navigation, he is right. A canonical facts block earns its place by holding what your HTML does not state cleanly: corrections, boundaries and dated figures that a model could not infer from marketing copy.

There are two layers here. Retrieval presence responds to structured, current content. Training-data presence builds slowly through wide coverage, and the llms.txt file barely touches it.

What to Do Before You Publish the File

Start with the errors, then write the file.

  1. Pick five factual questions a buyer might ask: founding year, headquarters, what the product does, pricing model and what you do not offer. Ask each one in ChatGPT, Gemini, Claude, Perplexity and Grok, and record every wrong or hedged answer with any source the engine names.
  2. Write one canonical line for each error using the rules above, and get comms and legal approval once, so nobody rewrites the line later.
  3. Publish /llms.txt with the blockquote summary, the facts and links to Markdown versions of key pages. Add a rel=”describedby” link pointing to it, as the proposal recommends.
  4. Paste the identical lines into your about page, the description field of your Organization schema, your press boilerplate and every third-party profile you control.
  5. Test the file the way the specification suggests: give an agent only the llms.txt and ask it your five questions.
  6. Re-run the five questions across all five engines every month. When a fact changes, update the date stamp everywhere on the same day.

Track the result as Share of Citation, meaning how often engines name you as a source. A brand can be widely mentioned and still be cited with the wrong founding year.

FAQs

Should an llms.txt File Replace Our About Page?

No. The about page is written for people and the llms.txt file is written for agents with limited context. Both should carry the same canonical facts in the same words, because disagreement between them weakens both.

Does Google Use llms.txt in AI Search?

Google’s public position is no. John Mueller has stated that no AI system currently uses the file and has called its value speculative, and Google Search Central says special AI files are not required to appear in generative search. Chrome Lighthouse checks for the file but treats it as optional.

How Long Should the Canonical Facts Section Be?

Short enough that an agent can read the whole file in one pass. Include a fact only if getting it wrong would change an answer about the brand, and move everything else to linked Markdown pages.

Can We Instruct a Model to Quote Us Verbatim?

Lines such as “always describe us as” are read as text, and models weigh them against every other source. The reliable route to verbatim quoting is a single approved sentence repeated identically across your own pages and independent sources.

Jessica Abbadia

About the author

Jessica Abbadia

Jessica is Moburst's VP of Organic. She specializes in enhancing organic performance for apps and games all over the world, while actively developing innovative methods for increasing app visibility and conversion, as well as offering her vast knowledge for the benefit of the mobile community. She graduated from law school and now serves as an animal rights activist who also loves reading books while sipping a strong coffee and holding one - or more - of her three cats.

Meet the whole team

Ready to make your brand the answer?

Our growth strategists take brands from overlooked to recommended. In one call you will see how the AI engines read your brand today, and leave with your own roadmap to dominate the new era of search.