Free Tool · Agentic Browsing
AI assistants no longer just answer questions: they browse, compare, fill in forms and buy on behalf of their users. Google has been measuring this since May 2026 with Lighthouse's Agentic Browsing category, visible in PageSpeed Insights. Answer 10 questions and find out whether an agent can read your website, understand it and operate on it, and what to fix if it can't.
Indicative result based on your answers. For a precise measurement, run Lighthouse on your URL or ask us for the audit.
Google PageSpeed Insights for kiwop.com, mobile report from 13 July 2026. The Agentic Browsing category is experimental and shows a ratio of passed checks, not a weighted score.
Check it live on PageSpeed InsightsLighthouse's Agentic Browsing category audits four areas. None of them are about keywords: they're all about whether a machine can understand your page and act on it.
Names, labels and the accessibility tree
An agent doesn't see pixels: it navigates the accessibility tree, the same one a screen reader uses. Lighthouse checks that interactive elements have an accessible name, that the tree is intact and that what's visible matches what's exposed.
Labels on every field, real <button> elements (not clickable divs), ARIA roles only where they belong and a clean heading hierarchy.
CLS: nothing should jump
If the layout shifts while it loads (images without dimensions, banners that push content, fonts that reflow), the agent clicks where the button used to be. Visual stability stops being just a UX metric: it becomes operability for machines.
Explicit dimensions on images and embeds, reserved space for injected content, and font-display that doesn't reflow the text.
Discoverability for models
A Markdown file at the root of your domain that summarizes what your site is and where everything lives, aimed at language models. Lighthouse checks that it exists. To be honest: no AI provider confirms today that it actually consumes it, but it takes minutes and Google already audits it.
Publish /llms.txt with your site's structure and key pages in plain Markdown.
Declared actions for agents (experimental)
A proposed standard that exposes your website's actions (search, book, buy) as tools an agent can call directly, without guessing where to click. Lighthouse audits forms without a WebMCP declaration and the validity of the schema. Requires Chrome 150+ and an origin trial.
Declare your business's key forms and actions as WebMCP tools. This is pioneer territory: almost nobody has it yet.
The category is experimental and Google moves it fast: the ratio (e.g. 3/3) counts the checks applicable to your page that you pass, not a score out of 100. Our test below goes further than Lighthouse: it covers everything an agent needs to read you, understand you and operate on you, including what Lighthouse doesn't measure yet.
Ten questions about how your website is built. Each one comes with a hint to check it in two minutes. At the end: your level, your concrete gaps and how to close them.
Your result
An AI agent can read your content, understand what you sell and complete actions. You're in the leading group: most websites don't even get past the reading stage.
Part of your website is visible to machines, but there's friction that makes tasks fail: where a human improvises, an agent gives up. The gaps below are your task list, in order.
Today an AI agent can't reliably read your content, let alone operate on it. Every task an assistant attempts in your sector will end up on a website that lets it through.
Serve server-rendered content (SSR or static generation). Modern frameworks (Next, Nuxt, Astro) do this out of the box; if your website renders everything client-side, agents and much of the AI crawler traffic read an empty page.
Review robots.txt and explicitly allow GPTBot, ClaudeBot, PerplexityBot and Google-Extended if you want to exist in AI answers. Blocking them is voluntary invisibility in the fastest-growing channel.
Publish /llms.txt: a Markdown file describing what your business is and your key pages. Lighthouse checks it within Agentic Browsing and it takes minutes to write.
Add JSON-LD for Organization and for your services or products (Service, Product, FAQPage, LocalBusiness). It's the format models use to extract reliable data: name, prices, address, what you do.
Restructure the page: one h1 per page and a hierarchy without gaps, landmarks (nav, main, footer), alt text on meaningful images. The accessibility tree you generate is literally what the agent reads.
Associate a label with every field, use real button and a elements instead of divs with onclick, and name controls by their function. This is Lighthouse's names and labels check, and it decides whether an agent can fill in your contact or purchase form.
Set width and height (or aspect-ratio) on images and embeds, reserve space for banners and widgets that get injected, and use font-display so text doesn't reflow. Target: CLS below 0.1.
Replace overlays that cover content with discreet banners that don't (cookie banners included). An agent doesn't always find the close button; a human on mobile sometimes doesn't either.
Optimize your main image or text: modern formats (AVIF/WebP), preload the LCP resource, a cached server and no blocking JavaScript. Target: LCP under 2.5s on mobile.
Offer a path to conversion without mandatory sign-up and without a preventive CAPTCHA (use server-side fraud detection if abuse is a concern). An agent that can't complete the task takes it to another website.
Indicative result based on your answers. For a precise measurement, run Lighthouse on your URL or ask us for the audit.
Result of the Agentic Browsing Test at www.kiwop.com/en/agentic-browsing-test. Indicative: run Lighthouse for the exact figure.
Agentic browsing turns your website into another machine's interface.
In May 2026 Google added Agentic Browsing to Lighthouse (13.3) and turned it on by default; PageSpeed Insights has shown it ever since, alongside Performance, Accessibility, Best Practices and SEO. When Google dedicates an entire category to something, it's telling you where it's looking.
Browser-based assistants (ChatGPT, Claude, Perplexity, Copilot and the agents built into Chrome) already visit websites to complete tasks: comparing prices, filling in contact forms, booking and buying. If your website isn't operable by an agent, that customer doesn't bounce: they buy from your competitor's website, and the visit doesn't even show up in your classic analytics.
This is the second half of GEO. The first was getting AI to cite you when it answers; this one is getting its agents to act on your website when the user hands them the task. Both rest on the same foundation: machine-readable content, clean semantics and zero friction.
What people ask us when we show them the 3/3.
It's web browsing carried out by AI agents on behalf of a person: the user hands over a task (compare these prices, book me a table, request a quote) and the agent visits websites, reads their content and executes the steps. The website stops being just an interface for humans: it's also the interface your customers' machines use.
It's a Lighthouse category (from version 13.3, May 2026) that audits whether an agent can use your page: agent-oriented accessibility (names, labels, tree integrity), layout stability (CLS), presence of llms.txt and WebMCP integration. It shows a ratio of passed checks (e.g. 3/3), not a score out of 100, and Google marks it as experimental: the checks will keep changing.
A proposed standard (based on the Model Context Protocol) that lets a website declare its actions as tools an agent can invoke: search, filter, book, buy. Instead of the agent guessing where to click, your website tells it what it knows how to do. Today it requires Chrome 150 or higher and origin trial registration; it's the most pioneering part of the category.
It's a Markdown file at the root of your domain that summarizes your site for language models. To be honest: no AI provider confirms today that it uses it to generate answers, so don't expect a visibility jump just from publishing it. But Lighthouse already checks its presence within Agentic Browsing, it takes minutes and has no downside: do it and move on to the next item.
They share the same foundations. Almost everything an agent needs (semantic HTML, content that doesn't depend on JavaScript, a stable layout, speed, structured data) is also what Google has been rewarding in search for years: Core Web Vitals, accessibility, indexability. Preparing your website for agents doesn't compete with your SEO: it reinforces it.
This very website scores 3/3 on Agentic Browsing and 100 across all four PageSpeed categories (mobile report, July 2026), and we apply the same standard to our clients' websites. The AI audit tells you what agents and models see on your website today and in what order to fix it; the GEO service covers the other half: getting AI to cite and recommend you.
We tell you what agents see on your website today and leave it exactly like this one: 100 across the board and operable by machines. No spin, PageSpeed report included.
Talk to an architect