Dmytro Galko · open data

A map of your content, and a locked door

llms.txt conformance across the 1,500 most visited sites · 2026-09-19

How many sites publish an llms.txt has been counted several times this year, and the answers agree. This is the next question, which nobody seems to have asked: of the files that exist, how many actually work — the right shape, live links inside, and a robots.txt that lets the reader in.

An llms.txt is a map an assistant follows to your content. A map with dead streets on it is worse than no map: the assistant spends its one visit on a 404 and leaves. Nothing tells you when that happens.

The finding

13 of the 123 sites publishing a usable llms.txt — 10.6% — also name at least one AI crawler in their robots.txt and forbid it outright. They wrote a guide for a reader they are not letting through the door.

This is not the wildcard rule being read as a ban. Only a group naming the crawler explicitly counts here, and only Disallow: / counts as a block, because anything looser inflates the number.

Do the files have the documented shape?

Publish an llms.txt12316.1% of 765 answered
Have the documented shape10787%
No heading3
No sections7
No links at all12

A file with no links in it is a file that points an assistant nowhere. It exists, it would be counted as adoption by every published figure, and it does nothing.

Do the links inside still work?

Sites with at least one dead link1513.5% of 111
Links checked864
Resolved797
Dead364.2%
Could not verify31not counted as dead

At most eight links per site were followed, spaced out, so this is a floor and not a ceiling: a site with one dead link among its first eight may have more further down.

Who is banned by name

CrawlerBlocked outrightShare
GPTBot10615.7%
ClaudeBot10916.1%
Google-Extended9413.9%
CCBot12919.1%
PerplexityBot8913.2%
Bytespider11416.9%
Applebot-Extended7511.1%

Out of 675 sites that publish a robots.txt at all. A crawler that is never mentioned is counted as not mentioned, never as allowed and never as blocked.

How it was measured, and what it cannot tell you

Population: Tranco list V3YPN, top 1500 domains. Run on 2026-09-19 with well-known-audit/survey/llms-conformance.mjs.

Politeness: 4 domains at a time, 250ms between batches, 300ms between link checks, at most 8 links per site, no retries, honest User-Agent. A refusal is recorded as a refusal and never retried.

Of 1,500 domains, 735 could not be asked at all: 418 never answered — most of these are CDN and infrastructure names that carry no website — 133 refused the request, and 184 answer 200 to a path that cannot exist, so nothing they return is evidence of anything. None of the three is inside any percentage on this page.

each domain is first asked for a path that cannot exist; a 200 there means the site answers 200 to anything, and it is excluded from every rate.

only a group naming the crawler explicitly counts; the wildcard group is not evidence about a named bot, and only Disallow: / counts as a full block.

What this cannot tell you: whether any assistant actually read these files, and whether a site meant to publish one — a large share of llms.txt files on the web are written by a platform or a plugin rather than chosen by the publisher.

Want to know what your own domains publish? npx well-known-audit yourdomain.com, or write to hello@dkautomation.dev.