Giving an AI-built clinic site the structure and speed it never had

Jouvonna came to us with a finished website. She had researched the content with Perplexity, built the site with Claude, and shipped it herself. She wanted it converted to Next.js, pixel for pixel, with a CMS if one was needed.

We could have done exactly that. Instead we crawled all 18 routes first and measured every claim against the live site rather than assuming it.

What the audit found underneath was a different story: six critical issues, five high-priority ones, and a root cause that no Lighthouse score would ever surface. We wrote it up, showed her the evidence, and proposed the rebuild the findings actually justified. She agreed, and that became the project.

An audit tells you what is broken. It does not tell you what good looks like in your category. So before proposing anything, we studied the sites she is actually losing patients to, and read them for structure rather than style.

Ten competitors across the three groups that matter to her:

• The national urgent-care operators, CityMD, Carbon Health, One Medical and Village Medical, for how a multi-location clinic organises services, locations and providers at scale.
• Her direct local rivals, Next Level Urgent Care, which runs 45+ Texas clinics including one in her own Atascocita neighbourhood, and Clinic 45, the beatable Houston incumbent.
• The telehealth weight-loss players she names on her own site, Hims, Ro and Found, plus ZoomCare as the closest structural analog to what she is building.

Read side by side, the same patterns showed up in every single one of them, and in none of hers.

1. 1
   Services live under one prefix

   CityMD nests /services/illnesses/cold-flu. ZoomCare nests /services/weight-loss. Every operator in the set treats the URL as the hierarchy. Hers were flat root slugs with no hub, which is where the whole IA problem starts.

2. 2
   Locations are a two-tier tree

   One Medical and Next Level both run a locations index down to a city page. Hers sat at the root with no index for Google, or a Google Business Profile, to point at.

3. 3
   Homepages are short

   CityMD runs about five sections, Carbon and One Medical about seven, ZoomCare about eight. Hers ran fifteen, at 18.7 screens tall.

4. 4
   Many services get compressed, not listed

   CityMD collapses everything into four intent buckets: I am sick, I am injured, I need a test, I need immunizations. Nobody in the set puts nine full service cards, two pricing tables and five treatment monographs on a homepage.

5. 5
   Trust is a named human

   Hims leads with a named physician. One Medical and Village Medical run provider directories. Next Level leads with its founder's story. Her founder, a credentialed FNP-C with eight press features, appeared only on the press page.

The research also told us what to leave alone, which mattered just as much:

Everything else on that list was fixable with better tooling. This was not.

The site had 18 pages and a navigation bar that misdescribed how they related. Its own structured data declared a hierarchy of hubs that had never been built, with the phantom crumbs quietly pointed at the homepage to paper over the gap. A single URL was doing the job of a category index and three separate services, so search engines were being told one page was about colds, and diabetes management, and strep testing, and the services hub. It ranks well for none of them.

The fix was to make the URL the hierarchy: if a parent page exists, its children live under it.

The hubs only pay off if the navigation and the first screen point at them. Both were rebuilt around what a patient is actually trying to do.

The navbar carried ten interactive items, and two of the six real ones did not resolve to a page. Services dropped you on one arbitrary service. Locations was a scroll anchor, which from any interior page meant a full navigation back to the homepage to land at a scroll position. New Patient was a label that excluded returning patients and led to a menu of billable products.

It is now six items, and every one is a real page:

• Services and Locations are proper dropdowns that list every child page, so all eight services and all five clinics are one hover from anywhere on the site. Those five location pages previously had zero inbound links from the header, on a site whose whole business is people searching for a clinic nearby.
• New Patient became Book online, which is what it always was.
• Press kept its slot, and the rest of the clutter came out.

The hero had one primary button, "Walk in today", pointing at a scheduler. The rebuild splits the intent the way the strongest operators in the benchmark set do, into two equal paths: find a location, or browse services. Someone who needs care today gets an address. Someone still deciding gets the menu.

Underneath it sits the thing the research said was her real weapon: three headline prices, in the first screen, where the transparent pricing that beats almost every competitor is the first thing a visitor sees instead of something buried on page four.

The homepage behind it went from fifteen sections and 18.7 screens to ten focused sections, roughly 43% shorter, with the weight-loss depth moved back onto the weight-loss page where it stops competing with itself.

Restructuring URLs is the highest-risk part of a rebuild like this. Her site was young, so the ranking cost of moving was low, but her press coverage linked directly to the old flat paths and those links are hard-won.

1. 1
   Every legacy path redirects permanently

   Ten redirects carry each old URL to its new home, including the ones a prefix rule could not cover because the service was renamed as well as moved. All verified on the production domain, not just in config.

2. 2
   The sitemap cannot drift

   Both the sitemap and the robots file are generated from the same slug exports that generate the routes, so a new service page appears in the sitemap by existing rather than by someone remembering.

3. 3
   Previews stay out of the index

   Only the real production domain is crawlable. Vercel previews and demo deployments are disallowed, so no staging copy ever competes with the live site.

Three findings had nothing to do with design and everything to do with the business:

• She had no analytics at all. Not on any of the 18 pages. She was running a press campaign, paying for a website, and driving traffic to seven separate booking links, and could not answer how many people visited, which location page got traffic, or how many actually booked. The rebuild ships her own Google tag with booking clicks tracked as events.
• Most of the page vanished if one script failed. Thirty-one homepage elements started at opacity: 0 and were revealed only by a single inline script, with no <noscript> anywhere. An ad-blocker or a corporate proxy meant a blank page. Content is now server-rendered and visible by default, with the reveal applied on top only after the page has already loaded.
• Every mistyped URL hit a blank 79-byte response. No title, no branding, no way back. Now a branded 404 that routes people to the pages they were probably looking for.