We Scanned 97 Business Websites. Here's What AI Agents Couldn't Use.

How we did it

The sample is 97 live business websites across home services, trades, professional services and hospitality — real operating businesses, not a curated list.

• Agent readiness: Google Lighthouse agentic audits via the PageSpeed Insights API — accessibility tree, layout stability (CLS) and llms.txt validity.
• AI visibility: we asked ChatGPT, Perplexity and Google AI Overviews who the best provider was in each business's trade and city, then checked whether the business was named.
• 6 sites could not be scanned — dead domains, redirect loops, or no real website. Those are excluded from percentages and discussed separately below.

One scoring note: Lighthouse defines four agentic checks, but WebMCP requires browser-level inspection that cannot be done remotely. Every score here is out of three.

Result 1: 43% of sites had a defect an agent trips on

Result 2: the failures cluster tightly

Within the accessibility failures, one defect dominated: 20 sites had links with no readable text. An icon, a bare arrow, an image with no alternative text. A person infers the destination from position and context. An agent cannot.

Second most common was misapplied ARIA attributes — several sites carried between four and seven affected elements, almost always inherited from a theme rather than hand-written.

The one that should worry a business owner most was rarer but worse: a contact form with four unlabelled fields. An agent cannot tell which box takes the phone number. That is a lead-capture failure, not a technical nicety.

Result 3: 77% of these businesses are already named by AI

This is where our expectations were wrong. We went in assuming widespread invisibility.

• 70 of 91 businesses were named by at least one AI assistant.
• ChatGPT named 59. Perplexity named 54.
• Only 21 were not named at all.

If you have been told your business is invisible to AI, check before you buy a solution. For most of this sample it was not true.

Result 4: site quality and AI visibility are mostly independent

Crossing the two datasets is the finding that changed how we talk about this.

Read the corners. Thirty businesses are named by AI despite real site defects. Twelve have technically perfect sites and are named by nobody.

So a website rebuild would not have fixed those twelve. Their problem sits off-site — third-party presence, reviews, citations. And the thirty with defects are being recommended by AI and then handing visitors a page an agent struggles with.

These are two separate problems that are routinely sold as one package. If someone proposes a rebuild to fix your AI visibility, ask them which of these two things they are actually addressing.

Result 5: six businesses had no working website at all

Not slow. Not broken. Gone.

• Three had no DNS record — domains parked at registrar holding pages.
• One had no nameservers, consistent with an expired domain.
• One resolved but redirect-looped until it timed out.
• One had a domain that had been lost to a reseller and now displayed a "for sale" page.

That last one is worth dwelling on, because our scanner scored it. It returned a result, because there was a functioning page at that address — it just belonged to somebody else. Any automated audit will happily grade a squatter's page. If you run scans at scale, check that the site you scored is still the site you think it is.

What we would take from this

1. Check before you conclude. Most of these businesses were already named by AI. Assumed invisibility is often just assumed.
2. Unreadable links are the cheapest widespread fix. One in five sites, usually twenty minutes of work.
3. Test your contact form specifically. Unlabelled fields turn an agent-assisted enquiry into nothing at all.
4. Do not confuse the two problems. A clean site does not get you named; getting named does not mean your site works.
5. Verify the domain still resolves. Six of 97 did not, and one was scored anyway.

Limits of this research

Stated plainly, because the numbers above are only as good as the method.

• The sample is not random — these are businesses in one agency's client base, skewing toward home services in the US.
• AI answers vary between runs. We saw different names returned minutes apart on identical questions. Treat the 77% as a snapshot, not a constant.
• Three prompts per business, phrased as a customer would ask. Different phrasing surfaces different businesses.
• Three checks, not four — WebMCP is not remotely testable.
• One day. Everything here is 13 August 2026.

If you want the same two checks against your own site, both are free: the agent-readiness check and the AI search visibility checker. Related reading: what agentic browsing is and how to get cited by AI.