Few things in technical marketing have travelled further on less evidence than llms.txt. It is worth setting out precisely what it is, because the file itself is reasonable and the claims made for it are not.
What the proposal actually says
The llms.txt proposal suggests a markdown file at
/llms.txt containing a concise, curated description of a site: what it is, followed by
lists of its most useful pages with one-line descriptions. The stated motivation is that language
models operate with limited context, and that a curated summary is more useful to them than an
unstructured crawl of a large site.
The format is deliberately simple: a heading with the site name, a blockquote summary, then sections of links with short descriptions. That is the whole specification in substance.
# Example Clinic > A CQC-registered dermatology and medical aesthetics clinic at 12 Example Street, > London W1A 1AA. Consultant-led, open Monday to Friday 09:00 to 18:00. > Telephone +44 20 7946 0000. ## Treatments - /treatments/acne-scarring: what acne scarring treatment involves, who it suits and recovery - /treatments/skin-cancer-checks: how a mole check appointment works and what happens next ## Practitioners - /team/dr-example: consultant dermatologist, GMC registered, special interest in acne ## Policies - /complaints: how to raise a concern and the timescales we work to - /aftercare: what to expect after treatment and when to contact us Last reviewed 2026-08-01.
What it is not
- It is not a standard. It is one person's proposal that gained attention. No standards body has adopted it and no major model provider has documented support for it.
- It is not robots.txt for language models. It grants and denies nothing. Access control is robots.txt, which is standardised in RFC 9309, and at the network layer.
- It is not a ranking or citation factor. No published documentation from any provider states that the file is retrieved, parsed or weighted.
- It is not a way to give models different content. If the summary contradicts the site, you have created an inconsistency, which is the opposite of what you want.
If someone proposes llms.txt as an AI visibility measure, the accurate framing is: it costs an hour, it cannot hurt if it is accurate, and there is no published evidence that any major system uses it. Publish it as tidy housekeeping, not as a lever. Anyone selling it as a lever is selling a claim they cannot substantiate.
Why it is still worth publishing
Three modest reasons.
- Writing it forces you to decide which pages actually matter, which is a useful exercise regardless of who reads the output.
- It is a stable, plain-text statement of your key facts and best pages, which is a reasonable thing to have at a predictable URL.
- If adoption does arrive, the file is already in place and already accurate.
None of those justify a project. All of them justify an hour.
What a clinic's file should contain
- The clinic name and one paragraph stating what it is, where it is and what it provides.
- The canonical facts: address, phone, hours, registration where applicable.
- The treatment and information pages that best answer real patient questions, each with a one-line description of the question it answers.
- The practitioner pages, with the registers they appear on.
- The policy pages: complaints, aftercare, consent, privacy.
- A last reviewed date.
Everything in it must match the site exactly. A file that disagrees with your own pages is an active liability rather than a neutral one.
AI answers and assistants
LOW CONTROL- Whether your content is crawlable by the relevant user agents
- How clearly a page states the fact you want quoted, and where
- Whether the same fact is consistent across your site, profile and listings
- Structured data that removes ambiguity about who and what you are
- Whether a model chooses to cite you
- Which passage it lifts, or how it paraphrases it
- Whether an answer appears at all for a given prompt
- The training data of any model already released
- Run the same prompt repeatedly and record what is cited, with dates
- Check server logs for AI user agents rather than guessing at access
- Verify the fact you want quoted appears verbatim on a crawlable page
Keeping it accurate
A stale llms.txt is worse than none, because it publishes wrong facts at a predictable location. Add it to the same review cycle as the location pages: twice a year, and immediately after any change to hours, address, staff or services.
What actually affects retrieval
If the objective is being usable by an answer engine, the levers are the ones documented elsewhere on this site: crawlability and server-rendered content, self-contained passages, consistent facts across every surface, accurate structured data, verifiable provenance, and speed. Google's own documentation on AI features describes controls that do exist, and none of them is a markdown file.
Related proposals
You may encounter llms-full.txt, markdown mirrors of individual pages, and various
schemes for exposing content in model-friendly formats. Treat them the same way: harmless if accurate,
unproven, and not a substitute for the page itself being crawlable and clear. Be particularly careful
with any scheme that serves different content to model crawlers than to people, which raises a
cloaking question you do not want to answer.
A defensible position
Publish an accurate llms.txt because it is cheap and tidy. Do not report it as an AI visibility initiative. Do not pay a meaningful amount for one. Do not let its presence substitute for the work that has an evidenced mechanism. And when a supplier tells you that publishing one improved a client's citations, ask what the baseline measurement was, because the answer to that question is usually the end of the conversation.
Why it spread so quickly
The proposal arrived at a moment when clinics and their suppliers were being asked what they were doing about AI search and had no defensible answer. A file that can be published in an hour, that sounds technical, and that nobody can prove does nothing is a comfortable thing to sell into that gap. That is not a criticism of the proposal, which is modest and clearly written by its author. It is an observation about why a modest proposal became a product.
The same pattern is worth recognising because it will repeat. Whenever a new discovery surface appears and its mechanism is undisclosed, an artefact will be proposed, adopted enthusiastically, and sold as a lever long before anyone establishes whether it is one. The defence is the same each time: ask what the mechanism is, ask who documented it, and ask what the baseline measurement was.
Testing it yourself
If you want to know whether the file does anything for your clinic, the method is the one described in our measurement article. Establish a baseline across a fixed question set over at least two months, publish the file, change nothing else, and continue measuring for three more. Any conclusion drawn from a shorter window, or from a period in which other work was also done, is not a conclusion. Most suppliers claiming an effect have run neither the baseline nor the isolation.