How to Build Your Website With AI (And What It Still Can't Do)

Can you actually build a website with AI?

Yes, and around 900 million people a week are already using these tools for something, most of them for a fraction of what the tools can do. Building a real one-page site for a business is well inside that gap.

What decides whether it works isn't which model you use. It's the order you do things in. Almost every guide starts at "prompt it to build a website," which is step four of seven, and it's why so much of the output looks the same and gets found by nobody.

What's the right order to build a website with AI?

Seven moves. Skipping any of the first three shows up in the last one.

One, decide. Before any code, get the model to interview you about your business. What you do in the words a customer would use, who actually buys, how they find you now, the objection you hear most, and what people get wrong about your work. Have it write a one-page brief from your answers.

This is the step everyone skips and it's the one that decides whether the site says something a competitor couldn't paste onto theirs. If the brief that comes back could describe any business in your trade, your answers were too general. That's fixable by answering better, not by prompting better.

Two, structure. One page. The contact action appears three times, near the top, in the middle, and at the end. Every section either moves someone toward that action or removes a reason not to take it. If it does neither, cut it.

Three, write the copy. Tell the model to lead each section with the point, use words a normal person says out loud, and put a specific behind every claim. Add one rule that matters more than the rest: if a claim needs a fact you haven't given it, write [CONFIRM] instead of inventing one.

Four, build it. Semantic HTML, mobile first at 390 pixels wide, phone numbers as tel: links, real alt text, 16px minimum body text, and contrast of at least 4.5 to 1, which is the WCAG minimum for normal text.

Five, the invisible layer. Title tag, meta description, LocalBusiness structured data, sitemap, robots file, llms.txt, canonical tag. One prompt, and it's the part that decides whether anyone finds any of the work above.

Six, check it. Hand the model a list and make it answer pass or fail on each item. Never ask "is this good."

Seven, hand it off. Have it write a README covering what the site is for, the decisions you made, the ones you reversed, and what shouldn't change. Paste that into a new conversation in six months and you pick up where you left off instead of rebuilding.

Ask a model whether its work is good and it agrees with you. Give it a checklist and it finds its own mistakes.

Why does the interview step matter so much?

Because the model has read a million websites for businesses like yours, and left alone it will average them.

A real example of the difference. First draft for a tile business: "a premier provider of quality tile solutions serving the greater Vancouver area." Every word of that could be any tile company in America.

After the interview surfaced what customers actually get wrong, the same section became "most shower failures start behind the tile." That's a sentence a competitor can't paste onto their own site, and it came out of an answer, not a prompt.

What does AI get wrong when building a website?

Four things, over and over.

It invents facts that fit the shape of a sentence. "Over 500 satisfied customers since 2009" appears in a first draft when nobody said 500 and nobody said 2009. The [CONFIRM] rule catches this, and you search for that word before publishing.

It invents addresses in structured data because the schema shape it learned usually has one. A made-up address is worse than none, since directories scrape it and then you're correcting it in four places.

It defaults to light gray text at 14 pixels, which fails contrast for anyone over about 45, which is most people buying home services.

And it adds things nobody asked for. Carousels, animated gradients, testimonial cards with invented names. Delete anything with a made-up person in it.

None of these are reasons to skip the tool. They're reasons the checking step exists.

How long does it take?

About two hours for a single page, and most of that is step one. The build itself runs around twenty minutes and the invisible layer is one prompt.

The output won't win a design award. It'll load fast, say something true, dial when someone taps the number, and be readable by search engines. For a business that runs on word of mouth, that's the whole assignment.

Which AI tool should I use to build a website?

Whichever one you already have open. The sequence works in Claude, ChatGPT, Cursor, Lovable, or v0.

What differs is what each does badly. Browser chats drift on long builds and quietly drop sections they wrote earlier, so save the file after every change. Editor-based tools keep the whole project in view but will happily create fourteen files for a one-page site. The prompt-to-app builders look impressive in four minutes and generate a lot of visual noise you didn't ask for.

If you're choosing a platform rather than a tool, the deciding questions are export and domain ownership, and that's covered here.

What happens after the site is live?

This is where most DIY builds stall, and it's not a design problem.

Point the domain and confirm the site loads in a private window. Add it to Google Search Console and submit the sitemap. Then open URL Inspection, paste your page, and click Request Indexing.

That last one isn't a formality. A set of pages on this very site once read "Discovered, currently not indexed. Last crawled: Never." The assumption was that the domain was too new to be worth Google's attention and nothing on the page would change that. Wrong. One click, and ten of twelve URLs were indexed inside 48 hours.

If your site is already up and invisible, start here instead.

FAQ

Do I need to know how to code to build a website with AI?

No. The sequence above produces a single HTML file you never have to edit by hand. You do need to be able to save a file and open it in a browser.

Can AI write my website copy?

It can draft it. It can't know what your customers say on the phone, which is why the interview step exists. The best copy in an AI build is usually your own words, cleaned up.

Is a one-page site enough for a small business?

For most local businesses starting out, yes. One page that answers what you do, who it's for, and how to reach you beats five thin pages nobody scrolls.

What is llms.txt and do I need one?

A short plain-text file at the root of your site describing what your business does, for AI search tools that read it. Not universally used yet, costs one file, and it's a cheap bet as more searching happens inside assistants.

How do I stop AI from forgetting my project?

Keep a README with the decisions in it and paste it into any new conversation before you ask for changes. That single habit is most of what people mean by AI "memory."

The whole sequence, written out

This post is the shape of it. The Site Starter Pack is the actual thing: all eight prompts in order, setup notes for each tool, the worked example including the drafts that came out wrong, and the eighteen-item checklist for the invisible layer. $39, and prompt one is readable on the page before you buy.

Or start free. The Lead-Path Map shows where your current site loses people, whether or not you ever hire anyone.