llms.txt v2 is here. 62 of 100 major brand websites have no root llms.txt.

In August 2026 the llms.txt proposal at llmstxt.org was revised, and the second version is less about the file than about the web around it: pages can now say which llms.txt covers them and where their own Markdown version lives, a site can keep separate files for separate sections, and the new links can be declared in a page's HTML or in an HTTP header. It is a thoughtful extension, but it extends something, and before asking how ready the web is for v2 it seemed worth asking whether the first version ever arrived. So on 4 October 2026 we checked one hundred well-known brands, five sectors of twenty each, and put the same question to every one of them: what does your own brand domain serve at /llms.txt?

What v2 adds to v1

The first version, published by Jeremy Howard in September 2024, asked for a single Markdown file at a known address: an H1 with the site's name, an optional one-paragraph summary and sections of links to the pages that matter. The revision keeps the same file, with the same shape, and adds ways for the rest of a site to refer to it.

Two of those additions are about llms.txt files themselves, the link that points a page to the file covering it and the files for sections, and only the Markdown link stands apart from the file. A page cannot usefully point to an llms.txt its site does not serve, which is why the question of v1 comes first.

The foundation is missing on most of these brand domains

Of the hundred domains, 34 served a root llms.txt, 62 had no llms.txt at that address, and 4 could not be determined. In software the file has found a real foothold: fifteen of the twenty SaaS brands publish one and ten of the twenty developer-first brands do. Outside software it has barely arrived, with nine of the sixty brands in finance, professional services and large corporates, retail and publishing serving a file, and none of the publishers.

None of the twenty publishers we checked, from the BBC and Reuters to The Atlantic and Nature, serves a file we could read on its brand domain: nineteen answered with a page saying there was nothing there, The Independent's most pointedly, with a "Gone" message for this one path that differs from its normal not-found page, and The Economist's returned an error page at that path, which we left undetermined. Among the five AI-native companies in the sample only Mistral AI serves one; OpenAI, Anthropic, xAI and Perplexity each answered /llms.txt on their brand domain with a not-found page. Their developer documentation was out of scope by design, and five companies are too few to generalise from, but on their brand domains these are not the companies where the file is most often found.

How far the 34 have to go

For the 34 brands that do serve a file, v2 is a shorter step, and some are already part of the way there. Thirteen of the 34 files link to at least some .md addresses, the kind of Markdown resource that v2 lets each page announce for itself, and three go further and state that every page on the site is available as Markdown: Cloudflare, Netlify and Siemens, the last two of which name the Accept: text/markdown request header as one way to ask for it, with Siemens serving the llms.txt itself as Markdown. Vercel's root file already links to a second llms.txt for one section of its site, and ServiceNow's has a section of related llms.txt files, which is the shape v2's files for sections formalise.

What we have not yet measured is whether any of these sites declare those relationships on their pages, in a describedby link or a Markdown alternate in the HTML or the HTTP headers. That needs a page-by-page pass of its own, and the brands above are the natural place to start it.

What the files that exist look like

The proposal asks for very little, an H1 with the name of the site being the only required element, and thirty-one of the 34 files open with one. The three that do not are worth a look, because two of them are very large files that somebody went to the trouble of producing, and all three miss the one element the format insists on. Salesforce serves a single flat list of 3,777 links (3,682 distinct addresses, every one on its own domain) with no heading of any kind, no summary and no sections; it is a little over a megabyte, and it reads as an export rather than an introduction. Twilio's file, at 2.3 MB the largest whose size we recorded, opens with a list item rather than a title, carries two H1 lines part-way through and names many of its 153 sections after URL path segments such as "Lp" and "Report", which together suggest a marketing-site index joined to a documentation file. Notion's is the opposite case, a short and tidy file whose only fault is that it begins with its summary and never gives itself a title.

Beyond that, the files fall into two kinds. Some are written for a reader: Etsy's is titled as an official reference for AI assistants and is organised around what an assistant may state with confidence, what it must not assert and where each kind of request should be routed; Postman's opens with a hierarchy of authoritative sources and a rule for resolving conflicts between them, and says it was last modified two days before we checked; McKinsey, Sephora and Intercom each include a section of notes addressed to language models; and several carry a visible date, among them American Express (18 September 2026), Etsy (25 September 2026) and PayPal, which also numbers its versions. Others are exported from the site: at least six of the 34 are larger than 200,000 characters and all six read more like exports than selections, and seven of the files we counted in full list more than 200 links, the point at which our own guidance on the format suggests a curation review, although two of those seven, McKinsey and ServiceNow, reach that size by curating a large site rather than by exporting it. In four files, including the title line of Intercom's otherwise careful one, Chrome displayed garbled characters of the kind that appear when text is stored in one encoding and declared in another, and Chrome reported an unexpected character set for a fifth; we did not confirm the cause, and it is the kind of fault that only shows when someone opens the file itself.

The near misses

A check that counted anything served at the path as a file would have miscounted several of the domains that do not have one. ASOS sends the request to its homepage, Zalando answers with a page listing the countries it serves, H&M's regional redirect rewrites the path to an HTML page on another of its hosts before reporting it missing, IKEA returns the words "Not Found" as plain text and GitLab redirects to its sign-in page. None of these is llms.txt, and under v2 the distinction matters more, not less, because a page that points to an llms.txt is only as good as what is actually served at that address. Fidelity is a different case: its domain returned a storage service's access-denied message, which does not say whether a file exists, so it is one of the four we left undetermined.

What good looks like

• Get v1 right before v2. One H1, then a summary you would be content to see quoted and sections of links you chose, because v2's describedby links and section files point back to it.
• Choose, do not export. A file that lists every URL repeats the sitemap and buries the pages that matter; the Optional section exists so that priority and completeness can be kept apart.
• If you offer Markdown, make it real. A promise that every page is available as Markdown is only worth the pages that actually answer that way, and it is the promise v2's page-level links will make visible.
• Date it and give it an owner. A visible date and a named owner make it possible to tell when the file has fallen behind the site it describes, and a file that pages will point to needs one more than a file nobody links to.
• Serve it as text, in UTF-8, and open it in a browser now and then, since several files here displayed garbled characters that nobody would notice without looking.
• Let a missing file answer as missing. If you do not publish /llms.txt, the path should say so plainly, not return a homepage, a country picker or a sign-in page that a careless check will mistake for a file.

More on the format and how Flowpane assesses it in llms.txt, and on keeping it consistent with your crawler rules in cross-file coherence.

Method

Checked on 4 October 2026 in Chrome, one site at a time, with no logins, no forms, no crawling and no attempt to pass bot checks. The sample is 100 prominent brand websites chosen in advance in five cohorts of twenty (SaaS and software; technology and developer-first, including five AI-native companies; ecommerce and retail; publishers and media; financial, professional and large corporate), so the figures describe these brands and not the web as a whole. One URL was read per domain, https://<domain>/llms.txt, on the primary brand domain (so stripe.com rather than docs.stripe.com), following any redirects and recording where they led. A file counted as present when the content was a text resource intended as llms.txt, and as following the proposal's shape when it opened with an H1; not-found pages, homepage fallbacks and other pages plainly not the file counted as absent, and four sites where a bot check, an error page at the path or an access-denied response prevented a reading were recorded as undetermined, with ambiguous error pages compared against the page the same site returns for an address that does not exist. HTTP status codes, response headers other than the content type the browser reported, and the page-level discovery features of the August 2026 revision were not measured. Sizes, section counts and link counts come from the text received; for four files in the first batch they are estimates from a text-fetch tool, two sites were read from tabs opened by hand when the browser extension could not reach them, a small sample of linked pages was checked for four first-batch files through a separate fetch, and Salesforce's file was read in full from a saved copy and then re-read directly, with the same result.