The AI Website Mistakes That Look Right At A Glance

Your page looks done. The layout is clean, the colours work, and the button says exactly what you wanted it to say.

Then a friend opens it and asks who Maria is. You don't know a Maria. She's in your testimonials section, praising a garden she's never visited, because the AI needed a testimonial and invented one.

These are the AI website mistakes that actually hurt. I don't mean the layout breaking on a phone. You'd spot that in five seconds. I mean the ones that look right: a review from nobody, a statistic nobody measured, a phone number that rings nowhere, and a thank-you message that confirms nothing.

They get past you because you're checking the design, and the design is fine.

Below is the list I check every generated page against, why each one slips through, a typical v0 page as an example, and a ten-minute proofread that catches nearly all of them. The technical failures (mobile, images, links between pages, a form with no handler, accessibility) live in the overview of building a website with AI. This post is about the words and the promises.

What Does AI Get Wrong On Almost Every Page?

It invents what you didn't give it, leaves placeholders it meant you to replace, and promises features nothing on the page can deliver. Here are the eight I find most often, roughly in order of how often they show up:

1. Invented social proof. Testimonials with first names, star ratings, "trusted by 500 local families". If you didn't paste it in, it's fiction.
2. Invented numbers. "Founded in 1998", "40 raised beds", "over 3,000 kg of produce a year". They sound like research, and none of it came from you.
3. Placeholder details left in. 123 Main Street, a 555 phone number, hello@example.com, a footer that still says "Your Name".
4. Buttons that go nowhere. A link pointing at #, which just jumps back to the top of the page.
5. Social icons pointing at the platform. The Instagram icon opens instagram.com, not your account. Sometimes it's an icon for a platform you don't even use.
6. A success message that fakes success. You press submit, the page says "Thanks, you're signed up!", and nothing was sent anywhere.
7. Promises nothing keeps. "We'll email you a reminder the day before." A newsletter box. A chat bubble. None of it is wired to anything.
8. Your facts, improved. Opening hours shifted by an hour, prices rounded, a date paired with the wrong weekday, a copyright year that will be stale in January.

Number six is the one that costs people real money, so the example below spends the most time on it.

Why Do These AI Website Mistakes Get Past You?

Because each one is the most plausible thing that could sit in that spot. A model writes the likely next sentence, and the likely next sentence on a community garden page is "Join over 200 neighbours who already volunteer". It reads right and fits the design. It only looks wrong if you know the real number.

You're also checking the wrong layer. Most people judge a generated page the way they'd judge a template: does it look good? It does. The wording gets a skim, because the wording sounds like every other website, and that's how the fiction gets through.

Builder tools make it harder still. A chat model hands you one file you can read top to bottom. A builder shows you a polished preview and spreads the words across a dozen component files you never open.

A Community Garden Sign-Up Page, Built In v0

Here's a page shape I see constantly: a volunteer sign-up page for a community garden, built in v0 in one evening.

v0 builds with Next.js, Tailwind and shadcn/ui components, and the preview it shows is good. A hero photo slot, a short "about the garden" section, a grid of weekend shifts, and a sign-up form with name, email and a shift picker. It's a handful of prompts' work.

Then you read it properly, and this is what turns up.

The social proof. Two testimonial cards from "Maria, volunteer since 2019" and "Dev, Saturday regular". A banner reading "Join 200+ neighbours". Say the real garden has 34 volunteers, none of them called Maria. The page now claims six times that, plus a person who doesn't exist.

The shifts. Six shift cards with dates. Expect a date or two with the wrong weekday, because the model writes "Saturday 14th" without checking a calendar.

The promise. Under the form: "We'll text you a reminder the day before your shift." Nothing on the page can send a text. Nothing on the page knows your phone number either, because the form never asked for one.

The form. This is the big one. Press "Sign me up" and a friendly message slides in: "You're on the rota!" But in a generated React page like this, the button often does nothing more than show that message. Nothing is stored, emailed or sent. Every person who signs up gets thanked and then forgotten.

The fixes are plain. Delete both testimonials and the "200+" banner. Correct the two dates. Remove the text reminder line. And connect the form to something that keeps the submissions: a form service, a host that catches forms, or even an embedded Google Form. how a form on a static page actually sends anything explains the options without assuming you write code.

Which Mistakes Can Actually Cost You?

The fake success message and the fake reviews. Everything else on the list is embarrassing, but those two lose you people or put you on the wrong side of consumer law.

The fake success message loses sign-ups silently. Nobody complains, because everybody thinks it worked. You find out weeks later when the rota is empty.

Fake reviews are also a legal problem. In the US, the FTC's rule on consumer reviews and testimonials, in force since October 2024, explicitly covers reviews that claim to come from someone who doesn't exist, "such as AI-generated fake reviews", and lets the agency seek civil penalties against knowing violators. the FTC's announcement of its final rule on fake reviews and testimonials Other countries have consumer rules of their own. The page says it, so you said it.

Changed prices and hours come next. A price on your website reads like a promise, and somebody will turn up at 8am because the page said you open at 8.

Placeholders and dead buttons cost you trust rather than money. A 555 number tells a visitor nobody checked this page, and they start wondering what else nobody checked.

The Ten-Minute Proofread That Catches Them

Run these six steps on every generated page before anyone else sees it. They work on any page from any tool.

1. Ask the model what it invented. Paste your original content back and ask it to list every fact, number, name, review and link on the page that isn't in your content. The check prompt in the six prompts I use for every AI site is written for exactly this.
2. Search the file for the telltale strings in the table below. Use your editor's find, or search the whole downloaded folder if a builder gave you many files.
3. Click every link and button on the live page, not the preview. Note anything that jumps to the top or opens a platform's home page.
4. Submit the form with your own details, then go and find the submission. If you can't find it in an inbox or a dashboard, the form doesn't work, whatever the page said.
5. Read every number and date aloud against your content document. Check each date's weekday against a real calendar.
6. Hand it to someone who knows the business for two minutes. They'll spot "Maria" faster than you will.

It takes about ten minutes, and it's the difference between a page you checked and one you hoped was fine.

What To Do Next

Run the proofread now, before you share the link with anyone. Fix the form first, then the fake reviews, then everything else in the order of the list.

Then decide where the page lives. A v0 project publishes to Vercel in one click, and if you plan to keep building in v0 that's the easiest route by a distance. If you want it somewhere else, it has to be built into plain files first, and what you actually upload from a React project walks through that.

Be honest about the shape of what you built. A static host serves files and nothing more. OxyPages can catch submissions from a plain HTML form once you add one attribute to it, but it can't run a Next.js server, a database or a text-message reminder, so if your page needs any of those, pick an app platform instead. If it's plain pages and a form, put the finished page online and test that form one more time on the live address.