My SEO plugin made an llms.txt. I read what nine AI companies document, and thirty days of its traffic.

On my WordPress site there is a file I asked for on purpose. It lives at kovalweb.com/llms.txt; if Yoast SEO has made you one, yours is at your domain followed by /llms.txt, and a line near the top reads Generated by Yoast SEO. LLMs are large language models, the programs behind ChatGPT, Claude and Gemini, and the file is a note addressed to them. I switched it on on purpose. The plugin’s makers say that is how a site gets cited in AI answers. If yours is there and you do not remember asking, Yoast’s launch post says “we’ve introduced this feature as opt-in”, so somebody with the login did — you, or whoever looks after the site.

The answer first. If it is on, leave it; if it is off, do not bother. Will it hurt your rankings? Not on anything I can find: Google says in writing that it will “neither harm nor help” a site in its search, and no other company has written anything about reading yours. Will it get you cited? Nobody the file is addressed to has put that in writing, and on my own site a month of Cloudflare’s count of self-declared AI bots shows none was ever handed it — three limits go with that, below. What is written down about being found by ChatGPT is not a file you add but three doors that let its program in, and an earlier article walks through them. Whether a local business should chase AI answers at all is a larger question; this file is not the way to chase them. The rest is how I know.

The file itself

On 19 September 2026 mine was 1,499 bytes. It opens with the site’s name and a one-line description, cut here all but its last words; then the line that says who wrote it, then the first heading; then seventeen links, also cut:

> […] and WordPress\.

Generated by Yoast SEO v28.5, this is an llms.txt file, meant for consumption by LLMs.

## Pages

I did not write a byte of it. The stray backslash is the plugin’s: Yoast’s functional specification lists it as a known limit, “Currently these characters will be escaped.” The same page says “This file will be updated weekly by a scheduled action”, and it does — between two readings of mine nine days apart, the file’s date moved exactly seven days and its version line moved from v28.4 to v28.5.

What it was sold as

The file is a proposal, by its own account. Its page at llmstxt.org says who is meant to read it — an agent being a program acting on somebody’s behalf: “Agents are expected to view or search llms.txt to find the information they need, then follow the relevant links.” The strongest thing the proposal says about anything reading the file is on its changes page: “coding agents use them reliably” — with no name and no number behind it. That is not search, and not an assistant answering a question about your shop; it is a programming tool reading software documentation, where the proposal says the files “are used most heavily”.

Yoast’s press release for the feature is specific, and specific is checkable. The file is “in a format optimized for how large language models (LLMs) — including ChatGPT, Gemini, Claude, and Perplexity — ingest, extract, and summarize information.” And, from Yoast’s Principal SEO: “If you want your site to be cited and included in AI answers, this is how you show up.”

What the companies write down

Nine companies send programs to read websites: OpenAI, Anthropic, Google, Perplexity, Microsoft, Apple, Meta, Amazon and Mistral. On 19 September I read the page each publishes about those programs — for Microsoft, whose Bing help pages would not open for any tool I have, three of its posts on being cited in AI answers; for Mistral, where I found no such page, the documentation of its web-search feature — and searched each for the string llms.

They know the file. Four of the nine publish one for their own developer documentation — OpenAI, Anthropic, Perplexity and Mistral each answered with an /llms.txt on 19 September. Two name it on the very page that describes their crawlers, both as the way into their own documentation. OpenAI’s page about its crawlers, in one line: “For the complete documentation index, see llms.txt.” Perplexity’s says it twice over, in a paragraph and a banner, and addresses the programs themselves: “Fetch the complete documentation index at: /llms.txt. Use this file to discover all available pages before exploring further.”

And not one of the nine documents reading one on somebody else’s site. Amazon’s page on its programs lists, by name, the instructions they honour and the one they do not; this file is in neither list. For the rest, no page I could read says its programs read yours. A page that is merely silent proves little. A company that tells agents to fetch its own copy first, and documents nothing about yours, is not silent.

Google is the one company that has written a position down, and it has written two. Its guide to generative AI features in Search, “Last updated 2026-07-10”, puts the file first under “what you don’t need to do”:

Google’s guide to generative AI features in Search, under “what you don’t need to do”

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 will neither harm nor help your site’s visibility or rankings in Google Search, as Google Search ignores them.

Chrome, the same company, ships a check for the file in Lighthouse, its tool for checking web pages. The page for that check, “Last updated 2026-05-05”, says under “How to fix”: “Create an llms.txt file and place it in the root directory of your website”. Chrome’s scoring page calls the category “experimental” and says the check “Checks for the presence of a machine-readable summary at the domain root” — presence, not reading. So: Search, in July, “ignores them”. Chrome, in May, “Create an llms.txt file.” Both are Google’s; I print both and choose neither. Gemini, one of Yoast’s four, is Google’s; its search says it looks the other way.

Thirty days of the file’s own traffic

Documentation is what companies say. The file also has a log. Cloudflare sits in front of kovalweb.com and counts what asks for its pages; its AI Crawl Control screen files some of those requests as AI bots. On 19 September I opened its Metrics tab, set the range to the last thirty days, and filtered for any path containing llms. Cloudflare’s page for the feature says that on free plans this tab shows the past 24 hours; my screen, on a free plan, offered thirty days, and thirty is what I read.

The status-code panel showed one request in the whole month, answered in the 400s, on 16 September. The paths panel, same filter, same range, showed “No request data found.” Two panels on one screen, disagreeing; what that one request asked for, the screen did not say.

What the number does say. When I ask for the file it answers 200 — here it is. A request answered in the 400s did not receive the file, whatever it asked for. So: in thirty days, no request that Cloudflare counted as an AI bot was answered with this file. Three limits belong in the same breath. Only requests Cloudflare filed as AI bots are in that chart; a browser, a feed reader or a program it files elsewhere never enters it. On this plan Cloudflare files a request as an AI bot by the name it gives itself — the earlier article is about what that count contained on this site — so these are requests that called themselves AI. And thirty days is shorter than the file’s life. I cannot tell you nobody has ever fetched it. I can tell you that for a month, the programs that announce themselves as AI did not.

Nothing points at the file

That reading is less surprising once you ask how a program would find the file. There is one file a visiting program fetches without being told it exists: robots.txt. There is no line for this file in it. The standard for robots.txt leaves room for extra lines and names the one that took the slot — “Sitemaps”, the list of a site’s pages written for search engines — and no line for llms.txt exists in the standard, in the proposal, or on any vendor page I read.

The proposal’s own answer, added in its second version, is a link relation: a tag in a page’s code, or a line the server sends along with the page, saying the file that describes this page is over there.

On kovalweb.com, on 19 September: robots.txt, 356 bytes, contains the string llms zero times. The 96 pages in the site’s four sitemaps contain it zero times. No page carries the link relation, and neither the blog nor the file itself sends it along with the page. Sixteen of the seventeen links inside the file lead to ordinary web pages and the seventeenth to the sitemap; the Markdown versions the proposal expects them to point at — the same pages as plain text, for programs — do not exist, and asking this site for /index.md gets a not-found error. So a visit to this file would have to be a guess. Ahrefs, in its study of 137,210 sites with traffic that May, found that AI bots do not guess: “Zero requests came from AI bots for llms.txt files that don’t exist. They never go looking.” That is also the reason behind if it is off, do not bother.

None of this is the plugin’s fault: Yoast’s specification documents the file and its weekly rewrite, and nothing about making it reachable. That half is nobody’s job.

Leave it, or switch it off

Leaving it on proves nothing, and the one caution in that study is that a stale file misleads any agent that does read it — nothing above measures what an AI answer does with a site, only what asked for the file. If you want it gone, Yoast’s help page is written for switching it on, and the same screen reads the other way: in your WordPress dashboard, go to Yoast SEO → Settings → Site Features → under AI tools, LLMS.txt → set the toggle → Click Save changes. That last step is the help page’s too, and the one that makes the rest count.

On your own site, the check is the same two steps as mine. Open the file. Then, if Cloudflare stands in front of your site, open AI Crawl Control → Metrics, set the widest range it offers, and filter paths for llms. If it shows nothing, you hold what I hold, and there is nothing to do about it — the file is not broken; it is unvisited. If it shows requests answered 200, something that called itself an AI bot fetched the file; a fetch, which Ahrefs is careful to say is not a read, and is not a citation. Without Cloudflare, the same number lives in your host’s access log, the record of every request the server answered; there is no screen I can point you to, so ask whoever runs your site to search the log for llms.txt. Either way, with the same three limits, that is the file’s whole audience among the programs that say what they are.

Written by Oleh Koval at Specia1ne, who reads these reports for a living.