News — On August 10, 2026, Jeremy Howard (Answer.AI) published version 2 of the llms.txt specification: the file no longer just lists your important pages, it now asks you to serve a Markdown version of every page and to declare it explicitly through the rel="alternate" type="text/markdown" and rel="describedby" link relations (llmstxt.org, spec v2, August 10, 2026).
This is no longer a file, it is a discovery protocol. V1 rested on one simple idea: drop a Markdown index at the domain root listing the pages an AI might need. V2 shifts the centre of gravity onto each individual page, which must exist as clean Markdown and point to its own machine-readable version. Eight days after publication, we wanted to know who had actually implemented v2. The measurement says more than the announcement.
What llms.txt v2 actually adds
The specification asks for two new things. First, a Markdown twin for every page that matters, reachable either by appending .md to the original URL (page.html.md) or by replacing the extension with .md. Second, and this is the real contribution, an explicit declaration of that twin: rel="alternate" type="text/markdown" points to the Markdown version of a page, rel="describedby" points to the llms.txt file describing it. These relations can be set as <link> elements or as HTTP headers.
One detail changes the scope of the whole thing: an llms.txt file describes every page under its path, so a /docs/llms.txt covers all of /docs/. V1 forced an agent to guess the file existed by trying its luck at the root. V2 lets each page name it directly. That is the difference between hoping to be found and signalling yourself.
Our August 18 probe: the Markdown is there, the declaration is not
We queried a set of documentation sites known to be ahead on this topic, checking three signals: the presence of a /llms.txt, the existence of a .md twin genuinely returned as Content-Type: text/markdown, and the declaration of the link relations introduced by v2.
| Site tested on August 18, 2026 | .md twin | v2 relations declared |
|---|---|---|
| llmstxt.org | not applicable | Yes, alternate + describedby |
| docs.anthropic.com | Yes, text/markdown | No |
| docs.stripe.com | Yes, text/markdown | No |
| developers.cloudflare.com | No on the page tested | Partial, type text/markdown only |
The pattern is clear. The large documentation publishers have already done the expensive work, producing and maintaining a clean Markdown version of every page, but none of them have done the free work, adding two attributes to the <head>. In other words, the part of v2 that takes ten minutes is the part nobody has done yet.
A secondary signal is worth noting: docs.stripe.com returns a Link: </.well-known/skills/index.json>; rel="service-meta" header, a competing agent-oriented discovery mechanism that has nothing to do with llms.txt. The standardisation of machine readability is not settled.
What llms.txt v2 changes for your site and AI visibility
For your Google rankings, nothing. Google's public position has not moved: llms.txt is not used by Search, including its generative features, and maintaining the file neither improves nor harms visibility. We documented this gap between market enthusiasm and Google's silence in our analysis of Google's official position on llms.txt, and it still holds word for word.
For agents, it is a different story. A coding assistant fetching a library's documentation, a browsing model reading a product page before answering, an internal tool plugged into your knowledge base: all of them consume clean Markdown better than a page loaded with scripts and menus. That is the logic we detailed in our piece on optimising content for AI agents.
So the useful question is not whether you need an llms.txt, but whether agents read your site in order to act. For a restaurant, a tradesperson or a local practice, the answer is almost always no, and the topic can wait. For a SaaS vendor, a technical documentation site, a marketplace or a product catalogue queried by assistants, the answer is increasingly yes.
llms.txt v2: what to do this week
If you already have an llms.txt, add the declaration. It is the best effort-to-result move available right now: a <link rel="describedby"> element pointing to your file in the <head> of your pages. You will be, literally, ahead of Anthropic and Stripe.
If you already serve Markdown, check the Content-Type. A .md twin returned as text/html defeats the purpose. The test is one command: curl -sI https://yoursite.com/page.md.
If you have nothing, do not start with the file. Start by checking that your pages are readable without executing JavaScript. Client-side injected content remains the first obstacle, far ahead of a missing llms.txt, and it is the same diagnosis we reached in our reading of Cloudflare's agent readiness score.
The limits of this llms.txt probe
Our probe is a spot check, not a study. We tested a handful of documentation sites picked for their presumed lead, on one specific date. This is not an adoption rate measured on a representative sample, and we are not publishing a percentage because we do not have one.
Eight days is very short. The absence of declarations today says nothing about adoption in three months. It only says that announcing a specification and applying it are two distinct events, separated by a delay better measured than assumed.
No AI operator has announced that it uses the v2 relations. Neither OpenAI, nor Anthropic, nor Perplexity has documented crawl behaviour relying on rel="describedby". Adding these attributes is a reasonable, cheap bet, not an optimisation with an established return.
This article does not cover writing the file itself. The expected structure, an H1 title, a blockquote summary and H2 sections of links, did not change between v1 and v2: we detailed it in our 2026 GEO guide.
llms.txt v2: frequently asked questions
What changed between llms.txt v1 and v2?
V1 was limited to a Markdown file placed at the domain root, listing important pages. V2, published on August 10, 2026, adds two elements: a Markdown twin per page, reachable by appending .md to the URL or replacing the extension, and standard link relations that declare it, rel="alternate" type="text/markdown" pointing to the Markdown version and rel="describedby" pointing to the llms.txt file. The internal structure of the file did not change.
Does llms.txt improve Google rankings?
No. Google publicly states that Search does not use llms.txt, including for its generative features, and that maintaining the file neither helps nor harms visibility. The expected benefit concerns agents and assistants that read the site in order to act, not ranking in search results.
Should you implement v2 right now?
It depends on your site. If agents or assistants already consult your documentation or catalogue, adding the two link relations takes a few minutes and puts you ahead of most large publishers, who serve Markdown without declaring it. If your site is a local storefront, this is not a priority: whether your pages are readable without JavaScript matters more.
The Cicero take
V2 of llms.txt mostly tells you one thing about this market: everyone comments on specifications, almost nobody implements them. The sites already serving clean Markdown have invested weeks of engineering, and are skipping two attributes that would take a morning.
Our position has not changed since May: llms.txt is not a ranking lever, it is a courtesy extended to machines. It costs almost nothing, it will return nothing on a site that agents do not read, and it will never replace the only work that counts, having pages someone has a reason to cite.
Sources
- → llmstxt.org (primary source): the “The /llms.txt file, v2” specification, authored by Jeremy Howard, published September 3, 2024 and modified August 10, 2026, defining the
.mdURL patterns and thealternateanddescribedbyrelations. - → AnswerDotAI/llms-txt repository (primary source): the “v2” commit (86154ea) and the “discoverability” commit published on August 10, 2026, which date the specification update.
- → Search Engine Journal: coverage of v2 and a reminder of Google's position, August 17, 2026.
- → Cicero probe, August 18, 2026: direct HTTP queries against llmstxt.org, docs.anthropic.com, docs.stripe.com, developers.cloudflare.com, docs.perplexity.ai, developers.openai.com, docs.github.com, vercel.com, plus the anthropic.com, openai.com and ai.google.dev roots.
Growth and SEO & GEO content strategist, I founded Cicéro to help businesses build lasting organic visibility — on Google and in AI-generated answers alike. Every piece of content we produce is designed to convert, not just to exist.
LinkedIn