What a vibe coded website gets right, and what breaks later
The screenshot is not the problem. AI builds attractive pages, and it builds them faster than any person can. The problems are underneath, they are invisible from the front page, and they surface in the order that costs you the most.
Start here
It will look finished. That is the problem, not the proof.
Everything below assumes the site looks fine, because it almost always does. Judging an AI-built website by the screenshot is like judging a car by the photo: the parts that decide whether it gets you anywhere are the parts you cannot see from there.
- The design will usually be competent. Models have read more well-designed pages than any of us.
- The copy will be fluent and say nothing a competitor could not also say.
- It will be responsive, because everything generated now is.
- It will be live on a URL today, which feels like the finish line and is the starting line.
- None of that tells you whether a single enquiry will ever reach your inbox.
The part no screenshot shows
It looks like every other site built the same way.
This is the failure nobody lists, because the page looks good. A model answers with the average of the millions of websites it has read. An average is, by definition, the thing nobody remembers, and a website whose whole job is to make a business look like the obvious choice cannot afford to be forgettable.
- Ask three different tools for a site in your industry and you will get three versions of the same page, because they are all drawing on the same body of work.
- The layouts are competent and the words are fluent, and both belong to nobody. Your competitor can generate the identical thing this afternoon.
- What makes a site persuasive is the specific claim only you can make, in the words your customers actually use, in the order that answers their real objection. None of that is in the training data, because it is in your head and your clients' heads.
- Five years of building sites for other businesses is what tells us which of those claims will land and which sound good in a meeting and convert nobody.
- That is why we start with a half day of strategy rather than a prompt. Get that wrong and a beautiful site sells nothing; get it right and a plain one outperforms.
What is usually missing
Eight more things, in the order they cost you money.
None of these are visible in a screenshot. All of them are checkable, and most of them are the reason a site quietly underperforms for months before anyone works out why.
- 01
The form does not reliably arrive
The single most expensive failure, because nothing looks wrong. The form submits, the visitor sees a thank you, and the message goes nowhere, or into a spam folder, or to an address nobody reads. Every enquiry form we build gets a real delivery path and a fallback that cannot silently swallow a lead.
- 02
Nothing is measured
No analytics, no conversion tracking, no behaviour data. You cannot tell whether the site works, which means you cannot improve it, which means the next redesign will be based on opinion again.
- 03
It is invisible to search
Missing or duplicated titles and descriptions, no sitemap, no structured data, headings used for size rather than meaning, and sometimes content that only appears after JavaScript runs. The page looks fine and ranks for nothing.
- 04
It is slow where it counts
Full-size images served to phones, fonts that block the first paint, scripts that are not needed. Desktop on office broadband hides all of it. Your customer on a phone in a car park does not.
- 05
It is not accessible
Images with no alternative text, contrast that fails, a form nobody can complete with a keyboard, controls a screen reader cannot name. Beyond shutting people out, this is a legal exposure in Australia, and it is the single most common thing generated code gets wrong.
- 06
Nobody can safely change it
Generated code often repeats itself heavily, mixes concerns and has no tests. It is not that it is unreadable; it is that no one can tell which parts matter, so every change risks breaking something invisible.
- 07
The dependencies are frozen in time
It was built against whatever versions existed on the day. A year on, something needs a security patch, the patch changes behaviour, and there are no tests to tell you what broke.
- 08
You may not actually own it
If it lives on a platform, check what you can export, what the site needs to keep running, and whose account the domain and the hosting sit in. Find out before you need to move, not during.
Audit your own
Six checks you can run this afternoon, free.
If you already have an AI-built site, do not rebuild it on our say-so. Check it. Every one of these takes minutes and none of them needs a developer.
- 01
Send yourself a real enquiry
Fill in your own form from your phone, on mobile data, with an address you do not normally use. Did it arrive? How fast? Where did it land? Do this monthly forever, whoever built your site.
- 02
Run Lighthouse
Built into Chrome: right-click, Inspect, Lighthouse, run it on mobile. You get performance, accessibility, best practice and SEO scores with specific fixes. Anything under about 90 is worth understanding.
- 03
Try it with the keyboard only
Put the mouse down. Tab through the page. Can you reach every link, open the menu, complete the form, and always see where you are? If not, neither can a chunk of your visitors.
- 04
Check what Google actually sees
Search your own domain with a site: query. Are the pages there? Do the titles and descriptions read like something a person wrote? In Search Console, does the page render with its content present?
- 05
Look at it on a real phone, on mobile data
Not the desktop browser squeezed narrow. An actual phone, off wifi. Count the seconds before you can read something. Then try to tap the smallest button.
- 06
Ask where it lives
Which account owns the domain, the hosting and the code? Can you export it and hand it to someone else? If you cannot answer all three, that is the first thing to fix, before any design decision.
FAQ
Questions people ask about this
Can AI build a website?
Yes, and quickly. AI tools can produce a complete, responsive, reasonable-looking website from a description in well under an hour, and hosted builders will deploy it for you. The question is not whether it can be built but whether what was built holds up: whether the forms deliver, the pages are fast on a phone, the markup is accessible and indexable, and someone can safely change it later.
Is a vibe coded website good enough for a small business?
Sometimes, genuinely. If the site is a simple presence, the stakes of a mistake are low and you are willing to check it yourself, a well-made AI-built site beats no site and beats a neglected one. It stops being good enough when the site is how you get customers, when it handles their data, or when its failure costs you more in a week than the build cost in total.
What are the risks of an AI-generated website?
In the order they usually cost money: forms that fail silently so enquiries never arrive, no measurement so problems go unnoticed, missing SEO fundamentals so nothing is found, poor mobile performance, accessibility failures that exclude people and carry legal exposure in Australia, and code nobody can safely change later. None of these are visible in a screenshot, which is why they survive so long.
How can I tell if my website was vibe coded badly?
Run four checks. Submit your own enquiry form from a phone on mobile data and confirm it arrives. Run Lighthouse in Chrome on mobile and read the accessibility and performance scores. Tab through the site with the keyboard only and see if you can complete the form. Search your domain with a site: query and see whether the pages are indexed with sensible titles. Failures in more than one of those mean the site needs work, not necessarily a rebuild.
Should I rebuild my AI-built website?
Not automatically. Audit it first. Many AI-built sites need a handful of fixes rather than a rebuild: wire the form up properly, add analytics, fix the titles and the image sizes, correct the accessibility failures. A rebuild is worth it when the structure itself is the problem, when you cannot get at the code, or when the site has to do meaningfully more than it does today.
How much does it cost to fix a vibe coded website?
Less than you would fear if the bones are sound, and more than you would hope if they are not. A round of fixes to forms, tracking, metadata, images and accessibility is a small job. If the code cannot be safely changed, or the site is locked inside a platform you cannot export from, the honest answer is usually that rebuilding is cheaper than repairing, and we will tell you which one you are looking at before you commit.
Start your brief
Start your brief
Tell us what you're launching and what it needs to do. We reply within one business day with a fixed price and the first available Monday.
- Websites from $15,000 ex GST, fixed. The exact figure comes with the reply.
- A reply within one business day, from Michael, not a form robot.
- A thirty-minute call to hear what the site has to do.
- A fixed-price scope, sent as a web page with a version number.
- Kick-off on the next available Monday, held for you. We take four studio builds a month.