Lighthouse now measures Agentic Navigation: is your website ready for AI agents?
Lighthouse, the auditing engine behind Google's PageSpeed Insights, has just introduced a new, still experimental category: Agentic Navigation. For the first time, the tool evaluates whether your website can be read, navigated, and operated by artificial intelligence agents, not just by human visitors.
TL;DR: Lighthouse's new Agentic Navigation category measures whether AI agents (such as ChatGPT browsing, Gemini taking action, or Perplexity searching) can understand and operate your website. The score looks at llms.txt, WebMCP, and visual stability, three pillars that turn websites into machine-readable environments.
This is probably the most important change of the year for anyone who maintains a digital presence, and almost nobody noticed.
Why is Google measuring websites for AI agents?
The way people reach brands is changing. It is no longer just traditional search: now we have ChatGPT browsing websites, Gemini acting on the user's behalf, Perplexity citing sources as it builds answers, and agents filling out forms automatically.
The web was built for the human eye: visual structures, clickable interactions, responsive layouts. But AI agents do not "see" the way we do: they read code, follow semantic annotations, and operate tools through APIs or structured standards.
And when Google adds a new category to Lighthouse, the whole market reacts. That is what happened with Core Web Vitals (which became a ranking factor) and with Accessibility (which became a requirement in public tenders and certifications). Now the era of Agentic Navigation begins.
The message is clear: whoever makes life easier for machines will be read, cited, and recommended by them. Whoever ignores it risks disappearing from the map.
What does the Agentic Navigation category evaluate?
The Agentic Navigation score is still in beta, but it already looks at three concrete technical pillars:
1. llms.txt: the site map for agents
The llms.txt file is a proposed standard that describes a website's structure and content in a way AI agents can understand quickly. It works like a map: it says which pages exist, what each one does, how they relate to one another, and which ones take priority.
Without it, an agent has to crawl the entire website, spending tokens, time, and resources, to figure out what is available. With llms.txt, the AI knows in advance where to look for the information it needs.
Practical example: an agent checking hosting prices can go straight to /planos instead of going through the home, about, blog, and contact pages.
2. WebMCP: forms that AI knows how to operate
WebMCP (Web Model Context Protocol) is a layer of semantic annotations that turns ordinary forms into tools that agents can operate. Instead of a generic text field, the agent understands that one field is the "contact email," that another expects a "CNPJ" (Brazilian company tax ID), and that a certain button submits a quote request.
This means that, for the first time, an agent can fill out forms, trigger actions, and complete flows on the user's behalf, without the developer having to build a dedicated API.
The entire web becomes an API, as long as it is annotated correctly.
3. Visual stability: because agents also "look" at the screen
Many AI agents "see" the interface as an image: they take a screenshot, analyze the layout, and try to locate elements. If the layout keeps jumping around (buttons moving, images loading late, blocks repositioning), the agent gets lost.
The Cumulative Layout Shift (CLS) metric, already part of Core Web Vitals, carries double weight here. A website that is unstable for humans is unusable for visual agents.
The combination of llms.txt, WebMCP, and low CLS creates an environment where agents can act autonomously, and that redefines what a "functional website" means in 2026.
How do you check your website's Agentic Navigation score?
The category is already available in the beta version of Lighthouse. To run the audit:
- Go to PageSpeed Insights.
- Enter your website's URL.
- Wait for the full analysis.
- Look for the Agentic Navigation section (its label may vary depending on the version and language).
The score goes from 0 to 100 and shows how ready your website is to be operated by agents. Values below 50 are critical; above 80 are considered good.
Important: the metric is still experimental and will evolve. Google is collecting data, adjusting weights, and refining criteria. But the strategic direction is already set.
What do website developers need to do now?
The short answer: start treating websites as environments that agents can read and operate, not just humans.
In practice, that means:
- Implementing llms.txt with a clear website structure, page hierarchy, and content descriptions.
- Annotating forms with WebMCP, describing fields, expected types, and possible actions.
- Prioritizing visual stability: avoiding layout shifts, loading critical resources before rendering, and reserving space for images and ads.
- Documenting public APIs, if any, so agents can consume them directly instead of crawling pages.
These adjustments are not "for the future." Agents are already browsing websites today, and every month more AI tools gain the ability to act on the web.
Those who implement now get ahead. Those who wait will have to play catch-up.
Agentic Navigation is more than a technical metric
This new Lighthouse category marks a paradigm shift: the web stops being a space exclusively for human interaction and also becomes a service surface for autonomous agents.
And this goes beyond SEO or performance. Companies that make it easy for agents to interact with their systems can:
- Be cited as a trusted source in AI-generated answers (GEO, Generative Engine Optimization).
- Close deals without human intervention, when the agent fills out forms or queries APIs automatically.
- Reduce friction in the funnel, because the agent hands the end user a ready answer instead of making them click through 5 pages.
Agência Rollin has been building websites and applications on this principle: made to be read and used by agents, not just by people. What Google has just turned into a metric, we already treated as a design principle.
Key takeaways
- Lighthouse now evaluates Agentic Navigation, measuring whether AI agents can operate your website.
- The score considers llms.txt, WebMCP, and visual stability (CLS).
- The metric is still in beta, but it follows Google's pattern: it becomes a market requirement within a few years.
- Websites optimized for agents gain visibility in AI-generated answers and reduce friction in the funnel.
- Implementing now is a competitive advantage; waiting is a risk of obsolescence.
Here is a question for you: have you already run PageSpeed on your website to see your Agentic Navigation score?
If you want to understand how to prepare your website for this new era, or if the audit revealed gaps you do not know how to fix, Agência Rollin is ready to talk. Our team builds websites, applications, and GEO strategies with exactly this in mind: environments that work for both people and agents.
Talk to us and let's build a website that is ready for 2026, and for whatever comes next.
