Generate an llms.txt file for your website

Paste this into ChatGPT, Claude, Gemini, or any model with web access, replacing the SITE line with your own URL. It is written to constrain the model rather than just ask it nicely: the rules at the end are what separate a usable draft from a plausible-sounding fabrication.

Rule 5 is the one people delete, and it is the most valuable line in the prompt. Without it a model fills gaps silently, and you cannot tell a fact it read from a fact it produced. With it, the gaps arrive labelled, and your remaining work is a visible checklist rather than a hunt.

There is a second route that avoids the hallucination problem entirely. Instead of asking a model to recall your site, crawl it: our free generator walks your pages, extracts their real titles and descriptions, and assembles them into a structured crawl map. Because it reads pages that actually exist, it cannot invent a URL or a title, which removes two of the four validation passes below. It will not write your Key Facts for you, though — pricing, turnaround, and service areas still come from you, and those are the lines engines quote most.

Generation is optional. The file is short enough to write by hand in fifteen minutes, and hand-writing guarantees every fact is true, which is the only quality bar that matters here. Fill in the angle brackets and delete anything that does not apply.

The two lines that carry the most weight are the H1 and the blockquote directly under it. If an engine read nothing else, those two should still let it describe you correctly and without hedging. Write them last, once the Key Facts have forced you to be specific.

Key Facts is the most citable part of the file, because each Q→A pair is already the shape of an answer. What makes a good pair differs by what you sell. Use the row that matches you as the starting list, then cut anything you cannot state precisely.

Both are fictional, and both are written the way a finished file should read: no adjectives, every claim checkable, every URL real on the site it belongs to. The first is a local service business, the second a professional-services firm.

Notice what neither example contains: no "passionate", no "trusted", no "industry-leading". Every line is a number, a place, a name, or a boundary. The second example even states who it will not work with, which reads as a loss and behaves as a gain, because it stops an engine recommending the firm into enquiries that waste both sides' time.

This is the step that matters. A generated llms.txt is a draft written by something that cannot tell the difference between a fact it read and a fact it produced, and the whole purpose of the file is to be quoted verbatim. Four passes, in this order.

These five account for most of the corrections in a generated file. Knowing them turns validation from rereading everything into checking five specific things.

Is there an llms.txt generator?

There are tools that produce one, but the most reliable approach is to hand a language model a constrained prompt with your URL, then validate the output. The constraint that matters most is instructing the model to omit any fact it cannot find on the site and to label the gap, rather than filling it. Generation takes about two minutes; validation takes about ten and is the part that determines whether the file is usable.

Can ChatGPT write my llms.txt file?

Yes, and it produces a good draft when it can actually read your site. Give it browsing access or paste your homepage, about, and services pages into the conversation, because a model working from a URL it cannot open is recalling or inventing rather than reading. Then check every number and click every URL before publishing, since those are the two things models get wrong most often.

What should I put in an llms.txt for a small local business?

Lead with the facts buyers ask about: price range and what moves it, response or turnaround time, the specific suburbs or towns you serve by name, your hours, and your licensing or insurance. Name the places rather than writing "and surrounding areas", which matches nothing an engine can use. Six to ten atomic Q→A pairs in the Key Facts section carry more citation weight than several paragraphs of description.

How do I know if my generated llms.txt is any good?

Run four checks. Confirm every number against a page on your own site. Click every URL and delete any that 404. Resolve or cut the gaps the model flagged, then strip marketing adjectives. Finally, confirm the file returns a 200 as plain text at yourdomain.com/llms.txt with no redirect. If a line fails any of these, deleting it makes the file stronger.

Should I write llms.txt by hand or generate it?

Generate it if you want a draft in two minutes and are willing to verify it; write it by hand if the site is small, since fifteen minutes of hand-authoring guarantees every fact is true. Accuracy is the only quality bar that matters, because the file exists specifically to be quoted verbatim by AI engines. A hallucinated line does more damage than a missing one.

How often should I update my llms.txt?

Whenever a fact in it changes, particularly prices, service areas, and hours. Because the file is written to be quoted, a stale price keeps getting repeated back to customers long after you changed it. The durable fix is generating it from the same source of truth that feeds your structured data, so one edit updates both and the two never drift apart.