llms.txt will not put your pages into AI answers. Google has said in writing that it ignores the file. I would still publish one, because a narrow group of agents does read it and the cost is twenty minutes. Here is what the request logs show, and what a file worth reading looks like.
What llms.txt is
A Markdown file at the root of your site that lists the pages worth reading, with one line of description each. Jeremy Howard of Answer.AI proposed it in September 2024. The spec at llmstxt.org was revised to v2 in August 2026. It is a map for a machine that has already arrived at your domain, and it does nothing to bring that machine to you.
What the numbers say
Two studies in 2026 describe the same picture. Adoption is real, and almost nobody is reading the files. Ahrefs checked 137,210 domains that had traffic in May 2026 and found that 97% of the valid llms.txt files received zero requests that month. On domains without a file, no AI bot asked for one.
| Source | What it measured | Result |
|---|---|---|
| Ahrefs, 15 June 2026 | 137,210 domains with traffic in May 2026 | 38,360 served a valid file, a 28% figure the authors call an upper bound. 97% of those files got zero requests. AI retrieval bots were 1.1% of requests. |
| SEOmator, 13 September 2026 | Top 100,000 domains | 8,598 publish one, or 8.6%. 20,171 answered 200 at /llms.txt, so a status code proves nothing. 16.4% of the real files contain no links at all. |
| Google Search Central, June 2026 | Its own crawler and ranking | Ignores the file, and says so in the changelog quoted below. |
What Google actually said
Google’s June 2026 changelog added a note to its AI optimisation guide, and the wording is worth reading twice.
“It’s completely fine if you decide to create and maintain LLMS.txt files (or other similar files) for other services or systems that use these files. Doing so won’t harm (nor help) your visibility or rankings in Google Search, as Google Search ignores them.”
Its AI features documentation, last updated 10 December 2025, says the same thing in fewer words: no new machine-readable file is needed for its generative features. So a sales page promising AI Overviews through llms.txt is selling you a file that a Google crawler will not open. What decides whether a passage gets quoted is how well it answers one question, which is why query fan-out is the thing to study instead.
Who does read the file
Coding agents, mostly. In the Ahrefs data, Claude-Code out-fetched every AI bot, while GPTBot led the crawlers at 4.51% of total requests. A developer agent opens your file to find the three pages it needs without crawling four thousand. That is a real use case, and it is the one I would write for: API references, changelogs, support pages, one line of plain description each.
What a useful file looks like
- An H1 with the site name. 95.1% of published files get this right, so it is not the hard part.
- A one-line summary underneath, inside a blockquote, saying what the site is for.
- Grouped sections with absolute URLs, each carrying a description of five to ten words.
- Under about 200 lines. A file that lists every post is a sitemap with extra steps.
- A date or a version line, so a reader knows how stale it is.
What I would not do
- Let a plugin generate it and never open it. Shopify accounts for 1,561 of the files in the SEOmator sample, Yoast 463 and Rank Math 146. Generation is cheap. Review is not.
- Treat it as an access control. Crawler permissions live in robots.txt and headers, and the ChatGPT crawler rules are where that fight is happening.
- Put it in a client report as a visibility win. There is no measurement that supports the claim.
Why a 200 status fools people
Check any WordPress site at /llms.txt and you may get a 200 back with an HTML page in the body. That is a soft 404, not a file. SEOmator found 10,775 domains doing exactly that in September, which is more domains than the 8,598 that publish a real file. Any audit that counts status codes without reading the body will report adoption at more than twice the true figure.
Read the response, not the code. If the body opens with an H1 and ends with a list of links, you have a file. If it opens with a doctype, you have your homepage wearing a different hat.
When I would skip it
If your site is forty blog posts and a contact page, your sitemap already does this job and an agent will find what it needs. Write the file when you have documentation somebody would want to quote, or when your own tools read it during a task.
The test worth running
Publish a short file and check the server log after thirty days. If the only hits are your own monitor, you have reproduced the study on your own domain and lost nothing. Passages get cited for answering a question with a date, a number and a source, which is the point I made in why blog posts never show up in AI answers. A text file does not fix a page that says nothing.
If you already have one, open it today and count the links. A map with no destinations is a title card.
