Run your website through Google Lighthouse and Cloudflare’s Agent Readiness checker, and you may get two very different results. Lighthouse might pass your page on its experimental agent checks while Cloudflare gives your website a low readiness score.

Neither result is necessarily wrong. The tools measure different requirements, using different methods and different definitions of what makes a website ready for AI agents.

Once you see where each one stands, the disagreement stops being confusing and starts being useful.

Key Takeaways

Why do Google and Cloudflare give the same website different results?

Because each tool tests a different layer of your website. Lighthouse loads a single URL in Chrome, waits for the page to render, then audits the result: speed, layout, accessibility and agent-related checks.

Cloudflare takes a different approach. Its checker primarily tests what your website exposes through HTTP requests, files such as /robots.txt and those under /.well-known/, response headers and agent-related protocols. Rather than focusing on the rendered experience of one page, it evaluates whether the site supports standards that help AI agents discover content and interact with services.

So a well-built page can pass Lighthouse’s agent checks while the server behind it publishes none of the files Cloudflare looks for. Neither result is a malfunction.

What does Google Lighthouse actually test?

Lighthouse audits the experience of loading and using one page. It has four long-standing categories, Performance, Accessibility, Best Practices and SEO, and since Lighthouse 13.3 shipped on May 7, 2026, a fifth called Agentic Browsing. Each category produces its own result, so a strong Performance score says nothing about Agentic Browsing, and the reverse.

The SEO category is narrower than its name. When it launched in 2018, Google described it as a check of SEO basics with no guarantee for rankings: a crawlable page, a valid canonical tag, a title and description. It does not tell you whether anyone, or any AI system, will choose your content.

What does Lighthouse’s Agentic Browsing category check?

It checks whether an AI agent working inside a browser can read and operate your page. Chrome’s documentation⁠ lists four areas: WebMCP integration, accessibility for agents, llms.txt and layout stability.

The accessibility audits are a subset of the ones Lighthouse already ran: does every button and form field have a programmatic name, are roles and nesting valid. Layout stability reuses Cumulative Layout Shift, because an agent that finds a button and then clicks it may miss if the target moves. The llms.txt audit checks for a machine-readable summary file at the domain root.

The category is marked experimental and shows a fraction of passed checks, not a 0 to 100 score, because Google says agent standards are still forming. WebMCP audits also require experimental browser support. In DebugBear’s testing⁠, example.com passed 2 out of 2 applicable checks without any special AI features.

What does Cloudflare’s Agent Readiness checker test?

It checks whether your domain supports a list of emerging standards that help agents find, read, access and use a site. Cloudflare launched isitagentready.com on April 17, 2026⁠ and scores four areas.

Discoverability: robots.txt, an XML sitemap, and Link headers in the HTTP response.

Content: whether your server returns Markdown when an agent asks for it.

Bot access control: rules for named AI crawlers, Content Signals declaring how your content may be used, and Web Bot Auth for verifying a bot’s identity.

Capabilities: API catalogs, OAuth discovery, MCP Server Cards, Agent Skills and WebMCP.

Not every check shown in the report counts toward the score: commerce protocols such as x402 and the Universal Commerce Protocol are examined but currently excluded, and llms.txt is off by default.

Why can one site score 33 and 67 on the same checker?

Because the score depends on which requirements the checker treats as relevant. Slobodan Manic, writing in Search Engine Journal⁠, scanned his content site and got 33 out of 100 on the default “All Checks” setting.

He then ran the “Content Site” preset on the same site, the same day, and got 67. Nothing on the website changed between the two scans.

The Content Site preset excludes API and service discovery checks, authentication-related capabilities and Web Bot Auth, which aren’t relevant to many informational websites. If one checker can give a single website two scores that far apart, comparing your number against anyone else’s means nothing until you know which settings produced both.

Why can Cloudflare give websites low readiness scores?

Because it scores adoption of standards that very few sites have adopted yet. Cloudflare analysed a filtered dataset drawn from 200,000 of the most visited domains, excluding categories such as redirects and ad servers.

Of the sites evaluated, 3.9% supported Markdown content negotiation and 4% declared Content Signals. MCP Server Cards and API catalogs were even rarer, appearing on fewer than 15 sites combined.

Several of those standards are early. MCP Server Cards are a draft proposal, Web Bot Auth is an IETF draft, and Markdown for Agents and Content Signals were introduced by Cloudflare itself. A low score often means “not adopted yet”, not “broken”, though limited adoption does not prove a standard has no value.

How do Google’s and Cloudflare’s definitions of agent-ready differ?

The difference is one of emphasis, not two separate kinds of agent. Both tools check some form of agent integration, including WebMCP, but they assess different parts of the job.

Lighthouse leans toward the rendered page. An agent working in a browser uses screenshots, the page structure and the accessibility tree to decide what to click, as Google’s AI optimization guide⁠ describes, so Lighthouse checks labels, stable layout and WebMCP tools registered on the page.

Cloudflare leans toward what your domain publishes. Most of its checklist covers a Markdown copy of your content, machine-readable lists of your APIs, sign-in discovery and crawler rules, which an agent can use without going through the visual page.

How they inspect your site

Lighthouse renders one page in Chrome. Cloudflare mainly requests domain-level files, headers and selected protocol endpoints.

Page speed, layout and accessibility

Lighthouse checks these. Cloudflare’s Agent Readiness checker does not.

robots.txt and sitemaps

Lighthouse includes limited SEO-related checks, rather than a comprehensive sitemap audit. Cloudflare checks robots.txt and sitemaps, plus rules for AI crawlers.

Markdown and llms.txt

Cloudflare checks Markdown support. Lighthouse does not. Lighthouse checks llms.txt, while Cloudflare makes that check optional and leaves it off by default.

APIs, OAuth and MCP Server Cards

These are part of Cloudflare’s checklist, not Lighthouse’s.

Scoring

Lighthouse gives each core category a score from 0 to 100 and reports a pass ratio for Agentic Browsing. Cloudflare produces a composite score from 0 to 100, plus a level.

Which result should you trust?

Trust the individual checks, not the headline number, and only the checks that match what your site does. A coach’s blog needs readable content, a working sitemap and labelled buttons. It does not need an OAuth discovery endpoint.

Google’s own handling of llms.txt shows why the context matters. Lighthouse includes an llms.txt audit for emerging agent use cases. Google Search, however, explicitly states⁠ that it does not use llms.txt to determine visibility in AI Overviews or AI Mode.

Those statements do not contradict each other. They concern different products doing different jobs, and each tool answers only its own question.

Frequently Asked Questions

Why does Cloudflare give my website a low score when Google Lighthouse passes it?

Because they test different layers. Lighthouse audits one rendered page; Cloudflare mainly checks which agent standards your domain exposes, and most of those standards are still rare.

Is Lighthouse’s Agentic Browsing result a Google ranking factor?

Google has not identified Lighthouse’s experimental Agentic Browsing results as a ranking factor. The category measures technical capabilities for browser-based agents, not search rankings. Google also confirms that Search does not use llms.txt for AI Overviews or AI Mode.

Do I need an MCP server or API catalog for my website to be AI-ready?

Not necessarily, because browser-based AI agents can interact with ordinary website interfaces without MCP. These standards become more relevant when a business wants to expose structured tools, APIs or services for direct agent integration. Most informational websites don’t need them simply to make content accessible.

Why did my Cloudflare score jump when I chose the Content Site option?

The Content Site preset excludes API and service discovery checks, authentication-related capabilities and Web Bot Auth, which aren’t relevant to many informational websites. The default All Checks scan counts them, which pulls content sites down.

Does Google use llms.txt for AI Overviews or AI Mode?

No. Google’s AI optimization guide says Google Search does not use llms.txt and that creating one will neither help nor harm your visibility in Search.

Where can I see Lighthouse’s Agentic Browsing category?

It ships in Lighthouse 13.3 and later and in Chrome 150 DevTools. Lighthouse’s release notes said it would reach PageSpeed Insights within two weeks of the May 7, 2026 release.

⸻

Google and Cloudflare introduced different approaches to testing AI agent readiness within weeks of each other. Neither approach is necessarily wrong. They measure different capabilities.

The important question isn’t whether your website scores 100. It’s whether the technology your business actually needs is working.

Passing an AI readiness check doesn’t guarantee that AI systems will cite or recommend your business. Technical readiness and AI visibility are different challenges.

AI Visibility Studio helps businesses improve how their websites are discovered, understood and accessed by AI systems: aivisibilitystudio.com⁠.

References