I Audited a Website Getting Traffic From ChatGPT. The Results Made No Sense.

Sixty percent of a website's traffic was coming from ChatGPT, Perplexity and Claude. The website's technical setup was far from perfect. That broke my model.

Last week, while investigating the online visibility of a local aesthetics clinic in Buckinghamshire, I came across an analytics result that made very little sense.

Approximately 60% of its traffic appeared to be coming from AI platforms.

Then I audited the website. Based on how AI visibility is usually discussed, I expected to find an exceptionally strong technical setup. Clean architecture. Comprehensive metadata. Perfect machine-readable information.

That wasn't what I found.

The website worked, but its technical setup was far from perfect. It had the kind of gaps that a conventional SEO audit would normally identify as obstacles. Yet the AI referral traffic appeared to be arriving anyway.

Meanwhile, I had seen technically stronger websites with cleaner structures receive little or no obvious traffic from AI platforms.

The contradiction forced me to question one of the assumptions behind the product I'm building: Are we overestimating how much technical optimization actually explains AI visibility?

We approached AI visibility as a technical problem

When we started building Chad, a large part of the problem appeared technical.

Can an AI crawler access the website? Does the important content exist in the raw HTML, or does it only appear after JavaScript runs? Are the pages structured clearly? Do they have useful titles and metadata?

These questions matter. We have tested websites that look completely normal in a browser but return almost no meaningful content in their initial HTML. A human sees a polished website. A crawler may receive an empty shell.

If a system cannot access or understand the information, that creates an obvious ceiling on visibility.

But the Buckinghamshire clinic suggested that technical accessibility may only explain one part of the outcome. The website did not resemble the technically pristine setup I expected. It was still being found.

Readable does not mean recommendable

Technical optimization can make a business easier for machines to understand. It cannot force an AI system to recommend that business.

That distinction sounds obvious, but much of the AI-visibility industry still behaves as though the two are nearly the same thing.

Add the correct schema. Improve the metadata. Publish more question-and-answer content. Make the website easier to crawl. These actions may improve accessibility and comprehension. They do not automatically create the evidence required for a recommendation.

When someone asks ChatGPT to recommend an aesthetics clinic near them, the system is not trying to reward whichever clinic has the cleanest structured data. It is trying to produce a useful, defensible answer.

That decision may depend on signals extending far beyond the clinic's website: its reviews, location, reputation, practitioner expertise, treatment specialisms, directory listings, press mentions and the consistency with which other sources describe the business.

A technically imperfect clinic may have accumulated years of evidence across the wider web. A technically perfect competitor may have excellent infrastructure but almost no independent evidence that it should be trusted.

The first business has credibility. The second has excellent plumbing. AI systems may care more about the credibility.

The recommendation may have been earned elsewhere

This is the possibility I find most interesting.

We tend to treat AI visibility as something that happens primarily on a company's website. Improve the pages, improve the structure and improve the probability of being mentioned.

But the recommendation may have been earned somewhere else.

Perhaps customers repeatedly named the clinic in reviews. Perhaps it was listed in trusted local directories. Perhaps its practitioners had been discussed in articles or community conversations.

Its website may simply have been where people landed after the recommendation had already been made.

If that is true, a website-only audit can never explain the complete result. It can tell us whether a crawler can access the clinic's treatment pages. It cannot tell us why an AI system trusted that clinic over another one.

The more valuable question becomes: What evidence exists across the web that would cause an AI system to select and recommend this business?

That requires a different kind of investigation.

The 60% figure needs context

Before drawing any conclusions, the analytics figure needs to be verified properly.

Sixty percent of 20 visits means something very different from sixty percent of 20,000. The percentage could also be unusually high because the website receives little traffic from conventional sources.

Referral attribution is imperfect as well. A visit from ChatGPT tells us that someone clicked a link. It does not reveal the original prompt, how many competing businesses were included or how frequently the clinic appears across representative questions.

A high referral share does not independently prove that the clinic dominates AI recommendations.

What we have is an observation worth investigating — not evidence that technical SEO is irrelevant and not a controlled experiment establishing why the traffic occurred.

The responsible next step is to verify the date range, total session count and source breakdown. Then we need to work backwards. Which questions might have produced those visits? Which pages were selected? Which external sources mention the clinic?

Only then can we start explaining the result.

AI visibility happens in layers

This experience reinforced a distinction I have been thinking about while building Chad.

AI visibility is not one event. A company can publish a page without a crawler discovering it. A crawler can discover it without the page entering a useful search index. A page can be indexed without being retrieved for a relevant question. It can be retrieved without being cited. It can be cited without the company being recommended.

Even a recommendation does not necessarily produce a visit, enquiry or customer.

The complete chain looks more like this:

Published → crawlable → indexed → retrievable → cited → recommended → visited → converted

Most AI-visibility tools measure one or two points in that chain. A technical audit inspects the beginning. A prompt tracker takes a snapshot somewhere in the middle. Analytics records some of what happens near the end.

None of them independently explains why the outcome occurred.

What an AI-visibility audit should do

A useful audit cannot stop at producing a technical score.

The technical layer remains foundational. If a company's website cannot be crawled or understood, that needs to be fixed. But technical readiness should not be confused with recommendation readiness.

An audit also needs to examine the environment around the business.

Which competitors appear for the questions customers actually ask? Which third-party sources influence those answers? What does the wider web consistently say about the company? Is the business absent because its pages cannot be read, or because there is insufficient independent evidence to support a recommendation?

Each problem requires a different intervention.

Adding schema will not solve an authority problem. Publishing more articles will not necessarily repair inconsistent information across directories. Improving page speed will not create customer proof.

A technically strong website can remain invisible if the rest of the evidence does not exist.

What we are changing in Chad

This discovery made one thing very clear: A better technical audit is not enough.

Chad still needs to inspect crawlability, rendering, metadata and structure. That is the foundation. But it also needs to understand the competitive and recommendation environment surrounding a business.

Which prompts produce recommendations? Which competitors are selected? Which sources are being cited? Where does the company already possess useful evidence, and where is that evidence missing?

The product should not simply produce another score or an enormous list of possible improvements. It should determine where the visibility chain is breaking, choose the most valuable intervention and measure whether that intervention changes the outcome.

That is a much more difficult product to build than an AI SEO auditor. It is also much closer to what businesses actually need.

Technical SEO still matters

The wrong conclusion would be that technical SEO is irrelevant. It isn't.

A technically inaccessible website creates a real disadvantage. Clean architecture, server-rendered content, clear metadata and understandable page structures make it easier for search engines and AI systems to process what a business publishes.

But those things appear to be enabling conditions rather than the complete strategy. Technical quality helps a machine read the evidence. It does not create the evidence.

The clinic may have been receiving AI traffic because of its reputation, reviews, relevance or third-party mentions. Its technical imperfections did not necessarily help it. They simply did not prevent it from being discovered in the way I expected.

That distinction matters.

What we are investigating next

I want to understand whether the Buckinghamshire clinic represents a broader pattern or an interesting anomaly.

The next step is to verify the analytics and repeat the investigation across more businesses receiving identifiable AI referral traffic.

Do technically imperfect businesses regularly outperform technically stronger competitors? If they do, which external signals explain the difference? Do the same businesses appear consistently across ChatGPT, Claude and Perplexity?

I don't have those answers yet. That is precisely why the discovery is valuable.

Good product decisions should survive contact with evidence that challenges the original thesis. I began this investigation expecting to confirm that better technical optimization leads to greater AI visibility.

Instead, I found a business apparently attracting AI traffic despite a website that did not meet that expectation. It changed the question I am trying to answer.

The goal is no longer simply to make a website readable by AI. It is to understand what makes a business worth recommending — and then identify which intervention can measurably improve its chances.

Sources

• Alexander Shoucair— First-party observation / founder's note. Local aesthetics clinic audit findings, Buckinghamshire UK, October 2026. Primary research.