I built a medical practice's website with AI. Eight months on, here is how it held up.

The job

I have built a lot of websites over the years. Last November I got to build one that mattered more than most: the website for Legacy OB/GYN, a women's health practice in Frisco, Texas. For a lot of new patients, it is the first thing they see of the practice.

The first change went in on November 4, 2025. The site went live on the practice's own address on January 27, 2026. It has grown to 67 pages: every service, every provider, the areas the practice serves, and a journal of articles, each reviewed by a physician before it is published.

How AI built it

Most of the typing was done by Claude, an AI coding assistant. I decided what each page had to do, described it, read what came back, and sent it back again until it was right. The site's own history keeps score: of the 207 changes made to it so far, 175 name the AI as co-author.

The speed is real. In the first two days the site went from an empty folder to a homepage, provider profiles, a FAQ, and the hidden search information that tells Google who the practice is. That was 34 changes in two days.

What the AI does not do is know anything about this practice. It knows what a medical practice website usually looks like. That difference is where every mistake in this note came from.

The mistake we caught before launch

Early drafts had a section of five-star patient reviews from names like "Sarah M." and "Jennifer K." They read beautifully. No patient had written them. The AI had filled the space with what a review section usually says.

On January 3, three weeks before launch, I took them all out. The site now shows only real reviews from the practice's Google listing. A medical practice publishing invented reviews is not a style problem. It is the kind of thing that ends trust, and it is the clearest example I have of why nothing an AI writes goes live without a person reading it.

Eight months live: what held up

In October we ran our new search health check against the live site. Most of it held up well:

• 95.5 out of 100 on a 62-point checklist covering indexing, titles, headings, images, share previews, links, structured data and speed.
• Fast for real visitors. On phones, the main content appeared in 2.2 seconds for real visitors over the last 28 days, under Google's 2.5 second bar. On computers, 1.2 seconds.
• Every provider page was in Google. All five.
• It kept moving. 93 changes since launch day, including the physician-reviewed journal and more detailed provider information.

What we both missed

The same check found what neither the AI nor I had caught. Google had 54 of the 67 pages in its index. Thirteen were not there.

Every cause was the kind of thing that looks right to a person and an AI reading the code, and wrong to Google reading the page:

• Answers that only existed after a click. The FAQ answers and the header menus were built to appear when someone tapped them. Patients never noticed. Google, which reads the page as it arrives, never saw them, so the pages those menus linked to had almost nothing pointing at them.
• A confident ID card with 31 errors. The hidden information that tells search engines who the practice is used words the standard does not have, labeled nurse practitioners as physicians, and gave a practice name that did not match the Google Business Profile. It looked complete. It was written the way these records usually look, not checked against this practice.
• A hidden answer that said more than the visible one. The insurance answer in the search information named plans the answer on the page did not. Google requires the two to match.

Notice the pattern. It is the same one as the fake reviews: plausible, typical, and not true of this practice. AI is very good at typical. A practice is not typical. That gap is what a person has to check.

Fixing it, with the same tools

The fixes were made with the same AI assistant, the same week, and the score went from 95.5 to 100 with all 62 checks passing. The homepage's 31 errors are now 0. The full story of that week, including one fix that broke something for eight minutes and was rolled back the same hour, is in the companion note: The score said 95. Google was missing thirteen pages.

That broken fix belongs in this story too. Automatic image resizing is a sensible change. It looked fine when we checked it, and it stalled under heavy testing traffic. The lesson was not to stop using the tools. It was to load test that kind of change before it goes live, and to write that into the checklist.

What I would tell anyone building with AI

• Let it type, never let it vouch. It will write a review, a credential or a fact about your business that sounds exactly right. Read everything that a customer will read.
• Check the site the way Google sees it, not the way you see it in a browser. Page by page, with Google's own inspection tool.
• Run the check again after launch. Ours found real problems on a site that scored well and looked finished. The next one is in four weeks.

Would I build it this way again? Yes. A practice got a site it owns, and it has held up for eight months in public. The AI did the volume. The judgment, and the checking, still had to come from a person.