Chrome shipped a Lighthouse audit in May that checks whether you publish an llms.txt. Google Search’s own AI features guide tells you Search ignores it. Both pages are still up.
That’s the post. The rest is how I got there.
What landed
Chrome 13.2 put the code in on 1 May. 13.3 turned it on by default six days later. PageSpeed Insights now shows a fifth card next to Performance, Accessibility, Best Practices and SEO. Six audits in the new category: WebMCP tool registration, agent-centric accessibility (axe rules in a fresh hat), CLS, and llms.txt. Two of those audits — CLS and the accessibility rules — already existed under different names. WebMCP and llms.txt are new.
The scoring is a fraction, not a 0–100 number. You see something like 3/4, which looks meaningful until you read the code that produces it.
How the score is actually built
calculateCategoryFraction() skips anything marked notApplicable, manual or informative before it counts. Not enrolled in the WebMCP origin trial? Those three audits get skipped, and you’re scored out of 3, not out of 6. The number changes based on which experimental flags are flipped on. That’s not a property of the category — it’s a property of how the audit was bolted together.
Then there’s the llms.txt audit itself. No file at /llms.txt returns 404, which Lighthouse marks as Not Applicable, and your score doesn’t move. A malformed file fails. llms-txt.js wants an H1, at least one real Markdown link, and 50 characters of content. A tidy plain-text list of URLs fails. Twenty guides out there explain how to hit 3/3. The denominator isn’t 3, and anyone who wrote those guides didn’t open the source.
What crawlers are actually doing
Agents are showing up. HUMAN put browser-based agents at roughly 71% of agent traffic in April, with SaaS growing 41.5% month over month. Real number, real direction.
Not for this stuff. Seekio watched ~900 domains for 191 days and logged 1,227 requests for llms.txt against 44,996,657 AI-bot requests in the same window. None came from GPTBot, ClaudeBot, PerplexityBot or Google-Extended. Ahrefs found 97% of llms.txt files across 137,210 domains were never fetched at all. SE Ranking modeled 300,000 domains and their AI-citation predictions got better when they deleted the feature.
WebMCP is starker. A scan of 111,076 of the top 200,000 sites found zero implementations. No mainstream agent calls modelContext yet. They all still parse the DOM.
Two of the new audits (CLS, accessibility) are things you should be doing anyway. The other two — WebMCP and llms.txt — audit for the file nobody requests and the API nobody calls.
The caveats
This can flip. Google says Gemini in Chrome will be the first consumer of WebMCP tools, and the origin trial runs through Chrome 156, so registering tools today is a cheap forward bet. Buys you nothing now, isn’t stupid.
The category is experimental, doesn’t touch your other scores, isn’t a ranking factor. Issue #17082 has PageSpeed Insights running Chrome 146 against docs that require 150, open since June. Take the number accordingly.
What to do
Fix the CLS. Fix the accessibility tree. If you publish an llms.txt, write it with real Markdown links — because a broken one costs you a point an absent one doesn’t. Then move on to the parts of your site that real agents are actually hitting today.
Links:
- Lighthouse agentic browsing audits (source)
core/audits/agentic/llms-txt.js- Google’s AI features guide
- AI bots ignore llms.txt — Seekio
- llms.txt study — Ahrefs
- llms.txt and AI citations — SE Ranking
- A Developer’s Guide to WebMCP — freeCodeCamp
- State of Agentic Traffic, April 2026 — HUMAN
- Google Lighthouse Has A New Agentic Browsing Category — DebugBear
