Google Lighthouse starts auditing sites for AI agent readiness

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.

For philanthropic and policy institutions, this changes how grant-making positions and research findings surface in AI answers. Foundations whose research is technically open but practically unreadable to agents will lose narrative ground to think tanks that publish cleanly. The cost of fixing this is measured in days of engineering time, not quarters.

The signal in context

Google adding agent audits to Lighthouse follows a pattern visible across the major AI platforms in 2025. Anthropic, OpenAI, and Perplexity have all shipped browsing agents that need structured access to the open web. Cloudflare has rolled out infrastructure to let publishers control which AI crawlers see what. The llms.txt proposal has been adopted by Anthropic, Hugging Face, Stripe, and a growing list of developer-first companies. What was missing was a mainstream measurement layer that would force the rest of the web to comply. Lighthouse is that layer. When a tool used by every front-end team on earth starts grading you on agent readiness, the standard becomes de facto regardless of whether any official RFC ever lands.

The deeper shift here is that brand visibility in LLM answers is becoming an infrastructure problem, not a content problem. For the last two years, the playbook for showing up in AI answers has been "publish more, publish on Reddit, get cited by Wikipedia." That still matters. But the next layer is whether your own site, the canonical source of your brand's positions, is even legible to the agents doing the citing. A Lighthouse score gives CMOs the lever they have been missing: a number to put in front of a CTO that translates marketing's AI-visibility problem into engineering's quarterly roadmap.