Google Lighthouse starts auditing sites for AI agent readiness
When Google's own auditing tool starts scoring you on agent readiness, llms.txt stops being optional and starts appearing in engineering roadmaps.
Key takeaways
- Google Lighthouse now includes an experimental Agentic Browsing audit that checks for llms.txt and agent compatibility.
- llms.txt has moved from community proposal to a signal Google's own diagnostics measure.
- Financial services and multilateral sites buried behind JavaScript or PDFs are invisible to AI agents citing sources.
- Adding an llms.txt file is the cheapest single intervention to improve LLM citation odds for canonical brand content.
- Agent readiness is now an infrastructure problem owned by engineering, not just a content problem owned by marketing.
What happened
Per The Decoder, Google has added an experimental "Agentic Browsing" category to Lighthouse, the open-source auditing tool that millions of developers already use to score sites on performance, accessibility, and SEO. The new audits check whether a site is legible to AI agents, including whether it serves an llms.txt file at the root.
The Decoder reports that the category is experimental and currently behind a flag, but the signal is unambiguous: Google is now formally testing for agent compatibility inside the same tool that defines what "good" looks like on the open web. llms.txt, the Markdown-based convention proposed by Jeremy Howard in 2024 to give language models a clean, structured summary of a site's content, has gone from community experiment to something Google's own diagnostics check for.
The audit also looks at machine-readability signals beyond llms.txt, including structured content and how cleanly pages render without JavaScript execution, the conditions most agents operate under today.
Why it matters for your brand
This is the moment llms.txt stops being optional. When Google bakes a check into Lighthouse, it propagates into every CI/CD pipeline, every agency audit deck, and every procurement checklist within twelve months. CMOs at large enterprises will start seeing "Agentic Browsing" scores in the same quarterly reports that currently flag Core Web Vitals. If your score is red, someone will ask why.
For financial services brands, the implication is specific. Asset managers, banks, and insurers have spent the last decade locking content behind JavaScript-heavy interactive tools, gated PDFs, and login walls. None of that is legible to an agent. A wealth manager whose fund commentary lives inside a React-rendered carousel is invisible to a ChatGPT agent answering "what does BlackRock say about private credit in 2025." A clean llms.txt that points to plain-text versions of the firm's published views is now the difference between being cited and being skipped.
For multilaterals and policy institutions, the stakes are higher than they appear. UN agencies, the World Bank, and OECD publish vast quantities of authoritative content that should be the default citation source for any model answering questions about development, climate, or disaster risk. In practice, much of it is buried in 200-page PDFs with no machine-readable summary. llms.txt is the cheapest possible intervention: a single Markdown file that tells an agent "here is our position on X, here is the canonical link." Institutions that ignore this will continue to watch secondary commentary outrank their primary research in LLM answers.
For industrial groups, the agent-readiness audit exposes a different gap. Product specification pages, sustainability disclosures, and investor materials are typically scattered across regional subdomains with inconsistent structures. An agent trying to answer "what is Holcim's Scope 3 target" should not have to crawl twelve PDFs. A consolidated llms.txt that surfaces the canonical answer with a citation link is now table stakes. Procurement teams running AI-assisted vendor evaluations will route around suppliers whose sites their agents cannot parse.