I advise companies on SEO and AI search, so in September I ran Google’s new AI agent audit on my own website, nirlevi.com. It passed every check. Then I tested the one thing that actually matters on that site, the contact form, and found a bug no score had caught.
That gap between a passing score and a finished task is where many businesses will lose customers. Here is what I tested, what broke and how to check your own site in an afternoon.
Why business owners should care now
AI is already a sales channel. Adobe reported that traffic from AI sources to U.S. retail sites grew 393% year over year in the first quarter of 2026. By March, those visits converted 42% better than other traffic. A year earlier, they converted 38% worse. Adobe also measured a 12% higher engagement rate for these visitors.
The same report found a weak spot. Adobe scored how much of each page machines could actually read. Retail homepages averaged 75%, and product pages averaged 66%. In other words, about a third of a typical product page couldn’t be read by machines.
The next step is already arriving. Google’s guide to AI search describes agents as “autonomous systems that can perform tasks on behalf of people, such as booking a reservation or comparing product specifications.” Browser agents can read your site through screenshots, the page code and the accessibility tree, the same structure screen readers use. They don’t just visit. They try to finish the job.
If the job is a demo request, a booking or a checkout, your website is now being tested by a new kind of customer.
How an agent sees your page
An agent doesn’t look at your site the way a visitor does. Google’s web.dev team describes three main inputs. Screenshots let the agent see the page, but they are slow and expensive to process, so they work best as a fallback. The page code shows how elements are nested and which controls belong to which content. The accessibility tree is a stripped-down summary of every button, link and field, with its name, role and state.
Most agents combine all three, which is why small details matter. A “button” built from a styled box can look fine in a screenshot and still have no role in the accessibility tree. A form field without a label is a blank box with no instructions. An error shown only as red text, and not tied to its field, gives an agent no clear signal about what went wrong.
What the AI agent audits measure
Two tools gave me two very different numbers for the same homepage on September 10, 2026.
Lighthouse Agentic Browsing: 2 out of 2
Lighthouse is Google’s free auditing tool built into Chrome. Since June 2026, it has included an Agentic Browsing category. It covers three main areas:
- Accessibility basics that machines rely on, such as button names, field labels and valid roles. In my run, this rolled up 33 separate rules.
- Layout stability, because a page that shifts while loading makes an agent click the wrong thing.
- WebMCP, a proposed standard that lets a site describe its actions to agents directly. My site doesn’t use it, so these checks showed as not applicable.
Chrome calls the category informational and unbenchmarked. The result is a count of passed checks, not a 0 to 100 grade.
Cloudflare Agent Readiness: 20 out of 100
Cloudflare’s check asks different questions. Can agents discover your content? Does the site serve a Markdown version of a page when an agent asks for one? Does it publish rules for AI bots? Does it expose an API or tool catalog?
My site passed the robots.txt, sitemap and basic bot-rule checks and little else. That sounds alarming until you notice that 8 of the checks look for API, sign-in and agent-tool interfaces, which most small businesses don’t run.
The two numbers measure different things, and you can’t convert one into the other. More importantly, neither one tries to complete your actual task. The same Lighthouse report also gave the site 97 for accessibility, 100 for best practices and 100 for SEO. Not one of those numbers tried to send a message.
The bug no score caught
So I tested the task. In a controlled test with made-up details, I sent the contact form a three-letter message padded with a space on each side.
The browser’s check required at least five characters. It counted the spaces, so the message passed. The server trimmed the spaces first, counted three characters and rejected it. Rejecting it was correct. The problem was what the visitor saw next: a generic message saying the inquiry might not have been sent, with advice to retry or send an email.
Retrying would fail again, because nothing told the visitor what to change. A person might give up and write an email. An agent working through a task could simply retry, fail and report that your form doesn’t work. Either way, a lead is gone, and every audit still shows a pass.
To be fair to the tools, I found this edge case by reading the code, so the test proves the defect exists, not how often real visitors hit it. That is the point, though. No score would have told me to look.
The fix took four changes:
- The browser and the server now use one shared validation rule, so they can’t disagree.
- The error is attached to the field itself, using standard accessibility attributes that screen readers and agents can read.
- The cursor returns to the field, and everything already typed stays in place.
- The success message appears only after the server confirms it accepted the inquiry.
Then I verified it. With the same bad input, the form showed a field-level error and sent nothing. With a corrected message, it sent exactly one request, the server accepted it, and exactly one inquiry arrived in the test inbox. The Lighthouse result stayed at 2 out of 2 before and after. The score didn’t change. Whether the task could be finished did.
I wrote up the full test, settings and fix for anyone who wants the technical detail.
How to test your own site in an afternoon
You don’t need a new tool. You need a clear finish line and a few bad inputs.
- Pick the one path that makes you money. A demo request, a quote form, a booking or a checkout.
- Define the finish line in the receiving system. “Someone clicked submit” is an action. “The correct inquiry is sitting in our CRM” is a result you can check.
- Run the normal path, then the messy ones. Add extra spaces, a phone number with a country code, a pasted address or a skipped optional field. Agents and busy humans both produce input like this.
- Read every error message as a machine would. Is it attached to the right field? Does it say exactly what to change? “Something went wrong” fails both tests.
- Fix the basics. Google’s guide to building agent-friendly websites recommends real buttons and links instead of styled boxes, labels connected to every form field, no invisible overlays on top of controls and layouts that stay put. As the guide notes, everything that makes a site agent-ready also makes it better for humans.
- Retest after every change. New consent banners, layout updates and form tweaks are exactly where these bugs come back.
Where these bugs usually hide
- Browser and server checks that disagree, as in my case
- Error messages that appear as a brief pop-up and then vanish
- Cookie banners or chat widgets that sit on top of the submit button
- Multi-step forms that wipe earlier answers after an error
- Phone, date and postcode fields that accept only one format without saying so
Give each step an owner. Someone on the business side defines what success means, a developer fixes validation, and someone checks the result in the receiving system.
What not to chase
Audit tools come with long recommendation lists. Cloudflare gave me 12. Resist turning them into an automatic to-do list. An API catalog adds little if you don’t offer a public API, and Google has said its search systems ignore llms.txt files. Add those things when they support a real customer action, not to move a number.
Treat scores as a smoke alarm, not a certificate. A failing check deserves a look, and fixes like a missing label or a shifting layout help every visitor. A passing check proves only that the check passed.
The real test
AI agents will judge your website the way your most impatient customer already does: by whether the task gets done. The businesses that win this shift won’t have the highest audit scores. They’ll be the ones whose forms, bookings and checkouts work on the first try, even when the input is messy.
Pick your most valuable path this week and try to break it. If you can, so can an agent.
About the author
Nir Levi is an SEO, GEO and AEO consultant who helps SaaS companies, startups and large organizations get found in Google and AI search. He is the founder of Madrank and Godrank and has spent more than 18 years in search. He publishes his tests, guides and free tools at nirlevi.com.
