llms.txt v2 lands with Lighthouse auditing and WordPress support as Google disputes its value

The llms.txt proposal, a markdown file placed at a site's root to explain its content to AI agents, reached version 2 on August 10, 2026, two years after Jeremy Howard published the original. According to a dev.to writeup dated September 24, the revision arrives as the surrounding infrastructure catches up: Chrome's Lighthouse now checks the file, WordPress.org hosts one itself, and three of the four biggest WordPress SEO plugins can generate one in their free builds.

What version 2 changes

The file format itself is untouched. A valid document still needs only a single line — an H1 carrying the site or project name — after which a summary blockquote, free-form paragraphs and H2 sections of markdown links are all optional. The revision instead formalises discovery. The llmstxt.org changelog now defines what a file in a subpath means: it governs the pages beneath that path, and the most specific file wins. A page can point to its file through a rel="describedby" link element or an HTTP Link header set at the CDN, and per-page markdown twins may now live at a short page.md address in addition to the older page.html.md form.

Adoption is visible at the top of the ecosystem. On September 24, both developers.openai.com and docs.claude.com served llms.txt files with successful responses, as did wordpress.org — a 75-line document that opens by crediting WordPress with powering more than 43% of all websites. Howard counts thousands of publishing sites, AI labs among them, the dev.to post reports.

Lighthouse audits it but does not score it

Lighthouse 13.3 introduced an Agentic Browsing category with four audits, one of which — coverage reported by DebugBear in May 2026 — is named for the file not following recommendations. It fails a file with no H1, a file that is too short, or a file without links. A site that publishes nothing is not penalised: a 404 response marks the audit as not applicable, and the whole category deliberately carries no 0 to 100 score while the field is still young, per Chrome documentation cited by dev.to.

The audit is also currently buggy. Lighthouse issue 17082, opened June 18, 2026, records a spec-compliant file that returns a successful response yet still triggers a fetch failure in PageSpeed Insights, which was apparently running an older Chrome build than the audit requires.

Google's products disagree

Google Search Advocate John Mueller wrote on Bluesky in June 2025 that no AI system currently uses llms.txt, and that server logs make this plain: consumer chatbots fetch pages for training and grounding but never request the file. Search Engine Journal, as quoted by dev.to, framed the split as Google giving different answers depending on which product you ask — Search dismisses the file while Chrome audits it. The llmstxt.org text stakes out narrower ground, saying the heaviest real use is in software documentation, where coding agents follow the file toward API references and tutorials. No vendor has claimed that its shopping or search crawler reads it.

WordPress support is wide but shallow

A directory search for llms.txt returned 521 WordPress plugins on September 24, according to the dev.to comparison of free plugin builds. Yoast SEO 28.5 ships its generator switched off, then writes a physical file weekly through WP-Cron with five links per post type, drops posts older than twelve months, and refuses to overwrite a hand-written file. Rank Math 1.0.279 also starts off, then serves the file through a rewrite rule with nothing on disk, 50 links per post type and a noindex header. All in One SEO 5.0.2 writes its file daily plus ten seconds after any save, defaulting to 1,000 links per post type. SEOPress 10.2 is the outlier: its readme lists the feature among free ones, but the free build holds only a toggle and a documentation link, with the actual generator in the paid version. Prime SEO 1.2.1 — a plugin built by the dev.to post's author, a disclosed conflict of interest — serves the file immediately on activation and is alone in optionally producing a full-text llms-full.txt, off by default.

The notable gap cuts across all five: none of the free builds implements any of v2's discoverability additions. No describedby link, no Link header, no markdown twins.

Why it matters

llms.txt has crossed from proposal to plumbing. A mainstream browser tool now audits it, generators sit inside the dominant CMS's SEO stack, and the spec's own site, WordPress.org and AI vendor docs all publish files. Yet the consumer side has not moved: Google says no chatbot fetches the file, and crawler access to actual pages is still decided by robots.txt, which llms.txt does not replace. For site owners the cost is a few kilobytes and a toggle; the payoff today is mostly readiness for standards — such as v2's discovery hooks — that no free WordPress tool has implemented yet. Publishing one is cheap and harmless. Treating it as a lever for AI traffic is not supported by the evidence so far.