Can AI Agents Actually Use Your Website? What Agentic Browsing Means for Contractors

I'm Ryan Goering, founder of BaaDigi. We build and fix websites for contractors, and this year a new question started showing up underneath the old one. For twenty years the only thing that mattered was whether a customer could find your website. Now there's a second question: when something shows up that can't see your logo, can't see your truck, and can't see a single thing you paid a designer for — can it still book you the job? We built a free checker for exactly this, and after running it across our own client book and a pile of prospect sites, the pattern was consistent enough to write down.

What is agentic browsing?

Agentic browsing is when an AI agent operates a website on a person's behalf — clicking, filling out forms, comparing options, requesting quotes — instead of just reading the page and summarizing it. ChatGPT Atlas, OpenAI's browser, is built on Chromium, but rather than analyzing the pixels on your page it queries the accessibility tree, hunting for elements with a role of button or link and an actual accessible name attached. Microsoft's Playwright MCP, the official tool for letting AI drive a browser, does the same thing: it hands the model an accessibility snapshot, not a screenshot.

The industry has quietly converged on the same approach. The way a screen reader reads your website is now the way a machine reads your website. Which gives you one rule that decides everything else:

If it is not in the accessibility tree, it does not exist.

How much does a broken button actually cost you?

In 2023, researchers published WebArena, a benchmark that turns AI agents loose on real websites across 812 tasks — book the thing, find the thing, fill out the form. The best agent at the time completed 14.41% of them. A human completed 78%.

That was the running joke about AI agents for a while. Two years later the top agents score above 70%, a few points off human performance. That gap closed fast.

But look at where they still fail. Break the scores down by task type and agents do great on browsing and reading. They fall apart on the transactional work — precise form filling and checkout flows. One recent agent scored 86% on forum-style tasks and 59% on shopping.

Read that again, because that gap is your business. Reading an article is easy. Filling out a quote form correctly, with the right value in the right field, on a page that isn't labeled properly, is the hard part. The exact moment a stranger turns into a lead is the exact moment these things are weakest. If your form is a mess, you don't lose a page view — you lose the job.

What does an agentic browsing audit check?

Google added an Agentic Browsing category to Chrome Lighthouse that looks at four things:

1. Accessibility tree — can an agent identify your buttons, links, and form fields by name and role? This is the one nearly every contractor site fails, and everything else depends on it.
2. Cumulative Layout Shift (CLS) — does your page hold still while it loads? Agents move through a page far faster than a person. If your header drops down after a banner loads, the agent's click lands on nothing, it reads that as an error, and it leaves. There is no second attempt.
3. llms.txt — a plain markdown file at your site root that summarizes your business and links your key pages. Proposed by Jeremy Howard of Answer.AI in September 2024; Anthropic, Vercel, and Stripe have published one. In a study of 300,000 domains, roughly 10% had adopted it.
4. WebMCP — instead of an agent guessing its way around your page, your site hands it a contract: here's how to book, here's how to request a quote. Built by engineers at Google and Microsoft, it became a W3C draft standard in February 2026 and is already shipping behind a flag in Chrome Canary.

One honest caveat, because most articles on this get it wrong: WebMCP can't be tested from the outside. It needs browser-level inspection, so Lighthouse marks it "not applicable" rather than failing you, and it isn't counted in your score. Any tool claiming to give you a pass/fail on all four remotely is guessing. Our checker scores the first three on your live page and tells you plainly that the fourth can't be measured yet.

See it run on a real contractor site

Here's the whole thing in about five minutes, including a live scan of a real contractor site that fails — accessibility tree FAIL, layout stability PASS, llms.txt FAIL. We blurred the domain; the point is the pattern, not the business.

Watch on YouTube: "Your Website Has a Button AI Can't See" (5:46)

Which contractor sites fail this, and which pass?

We ran our own client book plus a spread of prospect sites through the checker. The pattern was clean enough to be useful: modern framework builds tend to pass; failures cluster on older and inherited sites. Sites built on page builders and drag-and-drop themes are where this bites hardest, because those tools generate styled <div> elements that look like buttons to a human and read as nothing to a machine.

The most common single failure we see is link-name — a link with no readable text, usually an icon-only social link or a wrapped image, where an agent can't tell where it goes. Second most common is button-name. Neither is a rebuild. Both are typically hours of work, not weeks.

Is agent-readiness a ranking factor?

No — and you should be suspicious of anyone telling you otherwise. Agent-readiness is not a confirmed Google ranking factor, and there is no published study showing it improves AI citations. This is positioning for the moment agents start outnumbering humans on your site, not a shortcut to page one today.

What makes it worth doing anyway is that two of the three scored checks make your site better for real people right now. The accessibility tree is the same structure a blind customer's screen reader uses, so fixing it is genuinely fixing accessibility. And CLS is a Core Web Vitals metric Google has used for years. You are not gambling on a trend — you're doing overdue work that happens to also serve agents.

How do you fix it, in order?

Fix these in this order, because each one depends on the last:

1. Accessibility tree first. Give every interactive element a real name and role. Use actual <button> and <a> elements instead of styled divs. Add aria-label to icon-only links. Label every form field. Typical time to fix: hours to days.
2. Then layout stability. Reserve space for images, ads, and banners so nothing jumps after load. Target a CLS of 0.1 or lower. Typical time to fix: hours.
3. Then llms.txt. A markdown file at yoursite.com/llms.txt with an H1 and links to your key pages. About twenty minutes of work with no downside — just don't expect it to move rankings on its own.
4. WebMCP last, and only when it stops being experimental. Almost nobody has implemented it. There's no penalty for waiting.

How do you check your own site?

Run our free agent-readiness checker. Paste your URL and it runs Google's agentic browsing checks against your live page and tells you which of the three you're failing and why. No email required to see your result.

If you also want to know whether AI names your business when a homeowner asks for a contractor — a different question entirely — that's what our AI search visibility checker is for. Agent-readiness asks "can AI use my site?" Visibility asks "does AI know I exist?" You want both.

Related reading: whether to let AI crawlers index your contractor website, the technical checklist of NAP consistency and schema markup, and whether ChatGPT recommends your business when a homeowner asks for a contractor near them.

Not sure where your site stands? Run the free checker above, or get a free quote and we'll walk through your results with you.