When the AI-built website doesn't work

// pre-launch// field-notes6 min read

AI made your website. The website does not work. You go to look at it, and either it looks weird, or it loads slow, or a button does nothing, or the form silently swallows submissions, or the whole thing falls apart on a phone. You are not crazy. You are also not the only one. The failure has a shape, and once you can name the shape, you can usually decide whether you are an afternoon away from a fix or whether something deeper is going on.

I have built a working production site this way and I have pulled apart sites that other people sent me broken. The patterns repeat. Three shapes show up over and over.

Shape one: it looks finished, but it isn't actually finished

This is the most common one and it's the meanest because everything LOOKS done. The pages exist. The nav has the right items. The hero section has copy. The font feels close to what you wanted. You scroll and it scrolls. You click and pages load.

Then a visitor tries to do the thing the site was actually built for. Submit the form. Buy the product. Schedule a call. Read the blog post you wrote yesterday. And one of those user paths just dead-ends.

The reason it looks finished is that AI is excellent at producing surface evidence of completion. Pages render. Headlines write. CSS applies. The fastest path from "instruction" to "done feeling" is rendering the surface that a human would judge as complete. But "the user can complete the thing the site is for" is a deeper test than "the page paints." Those two tests can pass and fail independently.

I shipped one of these myself on the page you're probably reading parts of right now. The thing scrolled fine, looked clean on every static check, and yet the mouse wheel was dead because of a coordination bug between my smooth-scroll library and my animation framework. The methodology that came out of that arc is documented in the-gate-that-caught-us. The short version is that "complete" means "I ran the actual user path end to end and saw it work," not "the file is updated and the build is green."

If your AI-built website is in this shape, the fix usually isn't a rebuild. It's a discovery pass. Sit down, open the site, and try to do the thing a visitor would actually do. Note where it dead-ends. That dead-end IS the work.

Shape two: it works in preview but breaks live

This one is sneaky because the AI legitimately tested the site. It loaded it. It saw it work. Then you pushed it to your domain and now the same site is broken. This is the gap between "preview environment" and "production environment," and it shows up as some subset of:

Environment variables that worked locally but aren't set in production
A third-party service the site embeds that gets blocked by ad blockers or privacy browsers (the modal that won't open, the form iframe that's invisible)
Cache or CDN behavior that serves a stale build for the first hours after deploy
Mobile-specific breakage that didn't surface in desktop preview
Form submissions that worked against a test endpoint but fail against the real one

I hit one of these myself a week ago. The site had an intake form built as a third-party embed. It worked perfectly when I tested it in the browser I was using to build it. Then I tested it in Brave with shields up, which is what a meaningful slice of my actual audience uses, and the entire form was hidden. Default-blocked as a third-party tracker. The fix was to rebuild the form as native code instead of a third-party iframe. Useful work, just not work I had scoped at the time.

If your AI-built website is in this shape, the diagnostic question is: where was it tested? If "it works" means "it worked in the AI's preview pane," that is a different test from "it works for a real visitor on a real device with a real browser and real privacy settings." You probably need a second pass that runs the site against conditions closer to production.

Shape three: it works, but it's not the site you actually wanted

This is the most quietly demoralizing version. The site works. Pages load. Forms submit. Nothing's technically broken. And yet when you look at it, it doesn't feel like yours. The copy sounds generic. The layout looks like every other AI-built site. The hero image could be from anywhere. The fonts are fine but not distinctive. The colors are tasteful but interchangeable.

This isn't a bug. It's an output of the way AI tooling works. AI is excellent at producing the median output of its training data, which means by definition AI-built sites tend to drift toward the average. If your business is the average, great. If your business is anything that needs to signal trust, taste, expertise, or differentiation, the median output is the problem itself.

The fix for shape three is not a code fix. It's a design and voice decision. You have to know what you want the site to feel like, and you have to be willing to specify that clearly enough that AI can hold the brief. That work can absolutely be done with AI in the loop, but it requires the operator (you, or someone you hire) to bring the taste, the constraints, and the discipline to enforce them across iterations. The methodology is laid out in how-to-vibe-code-a-production-landing-page.

How to tell which shape you're staring at

If a real visitor can't complete a real user path: shape one. Fix the path.

If it works in some environments and breaks in others: shape two. Test in production-shaped conditions.

If everything works but the site feels generic, you don't trust it for your business, or it doesn't differentiate you in the way you need: shape three. The work is design and voice, not code.

Shape one is usually a single-afternoon fix once you know where the dead-end is. Shape two is usually a single-day fix if you can reproduce the failure on demand. Shape three is the heaviest because it's a redesign pass, but it's also the most strategically valuable because the differentiation you build there is durable.

The trap that turns this into a rebuild

The trap is treating every shape as if it's the same shape. Most "my AI-built website doesn't work" stories that escalate into a full rebuild went through three rounds of asking the AI to fix the wrong shape. The AI patched a surface issue when the actual problem was a third-party embed being blocked. Or the AI rewrote copy when the actual problem was a broken form submission path. Each unfocused patch adds debt without fixing the underlying shape, and after three rounds the codebase is worse than the original draft.

If you can name the shape in one sentence, you can usually fix it in an afternoon to a day. If you can't, the project doesn't need more AI. It needs someone to look at the whole thing and figure out what's actually broken before more code gets written.

If you're staring at an AI-built website that isn't holding up, send the site URL, the failure description, what you've already tried, and your timeline. VibeKoded can scope the build, ship the prototype, or hand off the production site. → Work with VibeKoded