llms.txt v2: What Changed in the August 2026 Spec Update

llms.txt v2 is the first revision to the llms.txt specification since Jeremy Howard and the Answer.AI team introduced the format in 2024. Published on 10 August 2026, it keeps the file's original job unchanged, a plain-text index at the root of a domain that gives an AI agent a curated map of what a site contains, while adding a formal mechanism for linking an HTML page to a Markdown equivalent and quietly demoting one of v1's stricter rules to a naming convention.

For a site that already publishes an llms.txt file, the practical scope of the update is small: a few link relations, an optional HTTP header, and a second URL pattern to consider for Markdown copies of a page. Nothing about how the root llms.txt file itself is structured changes, and nothing in v2 alters the separate, harder question of whether the file measurably affects how often a site gets cited by ChatGPT, Claude, Gemini, or Perplexity.

What actually changed in llms.txt v2

Three changes make up the entire revision. Each is additive rather than a breaking change to the v1 format most sites already have in place.

1. A formal link from HTML to a Markdown version

V2 adds a standard way for a page to point an agent at a Markdown version of itself: a link relation in the HTML head, and an equivalent HTTP header for servers that would rather not modify page markup. An agent browsing the ordinary HTML page can now detect, without guessing at a URL pattern, that a lighter, easier-to-parse Markdown copy exists and fetch that instead. This is the change v2 is really about; the original 2024 spec covered the root index file but said nothing formal about how individual pages should expose machine-readable equivalents of themselves.

2. A second URL pattern for Markdown pages

V1 established the convention of appending .md to a page's path for its Markdown version, so /docs/tutorial would have a counterpart at /docs/tutorial.md. V2 keeps that pattern but adds a second option: appending .md to the full filename rather than the path, so /docs/tutorial.html becomes /docs/tutorial.html.md. The second pattern exists because a growing share of static site generators and CMS platforms serve pages with an explicit .html extension, where the original convention was ambiguous about whether the extension should be replaced or extended.

3. The Optional section loses its mechanical meaning

V1 gave a special, enforceable meaning to a section header named "Optional" inside llms.txt: tools consuming the file were expected to treat links under that heading as safe to skip when trimming content to fit a context window. V2 keeps "Optional" as a useful heading convention for organising a file, but removes the requirement that a consuming tool actually treat it differently from any other section. In practice this was already how most implementations behaved, since there was no way to enforce the distinction, so v2 mainly formalises the gap between the written spec and how the format was actually being used.

Why Howard made these changes now

The gap v2 closes is the one between an index file and the pages it points to. llms.txt v1 solved discovery at the site level: an agent could fetch one file and get a curated list of what mattered. It said nothing about the format of the linked pages themselves, so an agent still had to fetch ordinary HTML, strip navigation and boilerplate, and hope the result was clean enough to reason over. As agentic browsers and AI shopping and research assistants started fetching far more pages per session than a single chat query ever did, that per-page parsing cost became the more expensive part of the round trip, which is the problem the new link relations are aimed at.

Does v2 change whether llms.txt helps you get cited

No, and this is the part worth being precise about. The evidence on llms.txt and AI citation has not moved: sites with an llms.txt file are not cited measurably more often by ChatGPT, Claude, Gemini, or Perplexity than equivalent sites without one, because the major AI crawlers have not confirmed they read the file at scale. Google's own public guidance is unchanged too, that no new machine-readable file is required to appear in AI Overviews or AI Mode. V2 is a specification update aimed at agentic browsing and page-fetch efficiency, not a citation lever, and treating it as one would repeat the mistake covered in sitemap.xml vs robots.txt vs llms.txt: prioritising an unconfirmed file over the two that crawlers are actually known to read.

Should you update your file to v2

For most sites, updating is low effort and low risk, which is a reasonable bar to clear even without a confirmed citation benefit. It becomes a clearer priority for a specific kind of site:

• Documentation sites and developer platforms, where agentic coding tools already fetch Markdown copies of docs pages when one is available
• Sites that publish content in a CMS capable of generating a Markdown export automatically, where the marginal cost of exposing it is close to zero
• Sites already investing in a complete llms.txt implementation as part of a broader agentic-browsing strategy, rather than a one-off file created and forgotten
• Sites that were penalised by the old Lighthouse Agentic Browsing rules for an ambiguous file extension pattern, since v2 directly addresses that ambiguity

A brochure site with a handful of static pages and no plan to generate Markdown copies gets little from migrating beyond keeping the file technically current. That is a reasonable place to stop; nothing in v2 makes an unmigrated v1 file broken or non-compliant.

How to implement the v2 changes

The migration has two independent parts. The root llms.txt file needs no changes if it already validates. The new work, where you choose to do it, is exposing Markdown copies of individual pages and linking to them.

1. Generate a Markdown version of the pages worth exposing, typically documentation, guides, and long-form articles rather than marketing or checkout pages.
2. Publish each Markdown copy at either the path-based pattern (/docs/tutorial.md) or the filename-based pattern (/docs/tutorial.html.md), matching whichever your server or static site generator produces more naturally.
3. Add the link relation to the HTML head of the source page, or the equivalent HTTP header if you would rather not touch page markup, so an agent can discover the Markdown copy from the HTML page itself.
4. Validate the root llms.txt file still parses correctly and, if useful, add the Markdown-equivalent pages under their own heading for discoverability from the index file too.

<!-- In the <head> of /docs/tutorial -->
<link rel="alternate" type="text/markdown" href="/docs/tutorial.md">

<!-- Equivalent as an HTTP response header, for servers
     that would rather not edit page markup -->
Link: </docs/tutorial.md>; rel="alternate"; type="text/markdown"

The bottom line

llms.txt v2 is a real, useful revision aimed at a genuine gap, per-page discoverability of Markdown content for agents that fetch far more pages than a search crawler ever did. It is not evidence that llms.txt has newly become a citation lever, and the site-level guidance stays exactly where it was before the update: get robots.txt right for AI crawlers, make sure Schema.org markup is complete, and treat llms.txt itself as hygiene for agentic tools rather than a growth strategy. If you already have a v1 file, there is no urgency, and no penalty for waiting until you have a concrete agentic-browsing use case before doing the extra work v2 makes possible.