Your llms.txt is a sign an AI can read outside a door it cannot open

The 60-second answer

An llms.txt file can give machines a concise guide to the content you consider important. It does not prove that ChatGPT, Gemini or another assistant will retrieve that file, and it does not enable an agent to do anything on your website.

There are two different jobs:

• identity: describing your business and useful resources;
• capability: exposing controlled functions an agent can use to search, filter, book or buy.

For an informational website, a correct llms.txt may be a low-cost experiment. For ecommerce, booking platforms and SaaS products, agent capability is the more important engineering decision. A brochure can explain the shop. It cannot operate the till.

In this article

1. What an llms.txt file actually does
2. Why publishing it does not prove value
3. What WebMCP changes
4. What to inspect now

1. What an llms.txt file actually does

The llms.txt proposal gives a website a root-level file containing a concise, structured guide to selected resources. The intention is sensible: reduce navigation noise and point language models towards useful information.

Publishing a file, however, is not the same as having it used. No mechanism requires an assistant to fetch it, trust it or prefer it to ordinary pages. Treat it as a descriptive layer, not a guaranteed shortcut into an AI answer.

Start with the obvious check: open yourdomain.com/llms.txt. If a plugin generated it, read it as if it were a page for a prospective customer. Are the products, services, markets and links still correct? Does it update when the main site changes? An automated file can be current at launch and quietly wrong six months later.

2. Why publishing it does not prove value

An Ahrefs analysis reported by Search Engine Journal found roughly 38,000 valid llms.txt files in a sample of 137,000 domains. In May 2026, 97% of those files received no requests. AI retrieval bots accounted for only 1.1% of observed requests (Search Engine Journal).

That finding does not prove the format will never be adopted. The sample skewed towards technically active websites and covered one month. It does show that, in 2026, the presence of a file is not evidence of meaningful use.

Separate three levels when you review it:

Most discussions stop at presence. An investment decision should depend on retrieval and outcome.

3. What WebMCP changes

WebMCP tackles a different problem: how can a website expose structured tools that an agent is permitted to use? Chrome is testing it as an experimental feature in a Chrome 149 origin trial. A site can describe an action, its parameters and its rules, allowing an agent to invoke it more reliably than by guessing how the visual interface works (Chrome for Developers).

The status matters. WebMCP is not a universal production standard, and Chrome describes its Gemini integration as a separate later step (Chrome AI developer preview). Rebuilding a product around one trial would be premature. Ignoring the operating model it reveals would also be short-sighted.

Transactional businesses can prepare sound foundations now:

• clearly defined functions such as search, stock checks, quotes and booking;
• stable inputs and outputs;
• explicit permissions for sensitive actions;
• human confirmation before payment, submission or an irreversible change;
• logs showing what an agent requested and what the system executed.

This is product and engineering work. The descriptive file can support it; the file cannot replace it.

4. What to inspect now

Begin with a limited review, not a broad transformation programme.

If you already have llms.txt

Check its source, update date, links and owner. Generate it from maintained content where possible rather than creating a second manual version of your offer.

Then inspect server logs. If no relevant system requests the file, that is useful evidence: keep its maintenance cost proportionate to its observed use.

If your website supports transactions

Choose one valuable customer task. It might be “find an item available in this size” or “check an appointment in this postcode”. Define the input, output, business rules, possible errors and point at which a person must confirm the action.

This work remains useful if WebMCP changes. Clear, controlled and observable functions strengthen APIs, automation and the product experience as well.

If AI visibility is the priority

Do not reduce the diagnosis to llms.txt. Review crawlability, page clarity, factual consistency and the external sources that describe your business. Test a stable set of buyer questions afterwards. Visibility and agent capability are related, but they are not the same outcome.

Conclusion: measure the door, not only the sign

An accurate, automatically maintained llms.txt can be reasonable experimental cover. The available evidence does not justify treating the file alone as an AI visibility strategy.

For a transactional website, the more durable question is what an agent can safely do, under which permissions and with which evidence. Prepare one useful capability before multiplying files. Capability, not compliance theatre.

Frequently asked questions