How to make a website with AI, step by step

Most guides to this show you where to click. The clicking was never the hard part. What decides whether you end up with a site you would send to a customer is a short list of decisions made before and after the generating — the same decisions a good designer would have walked you through, in a different order.

What you need before you open anything

• One sentence saying what you do and who for. Not a slogan — the sentence you would say on the phone.
• The single thing you want a visitor to do: call, message, book, buy, subscribe, or walk in.
• Three to five facts a stranger needs: where you are, what it costs, what you make it from, who you are not for.
• Proof you already have: photographs of real work, a client name, a number of years, a qualification.
• The language your customers read in, which may not be the language you think in — those are two separate choices.

That is the entire brief. With it written down, most competent tools will get you something usable. Without it, the best tool in the world produces a confident page about nothing, and no amount of regenerating fixes that.

Step one: describe the business, not the layout

The instinct is to specify a layout. Resist it. "Hero image, three columns, testimonial slider" describes every website ever made, so it steers the result towards the average of all of them. Describe the business instead — "I restore vintage watch movements in Bengaluru and most of my customers find me on WhatsApp" — and the layout follows from something specific. Layout instructions cap the outcome at what you could already picture, which is the one thing a generator adds nothing to.

Step two: choose the two languages separately

Describe the work in whichever supported language you think fastest in, then choose the website's language from who is reading it. Detail is what makes a page specific and detail is the first casualty of writing in a language you are merely competent in. Casel keeps these as two questions for that reason; the language questions are answered here.

Step three: read the first result as a draft

Judge the argument, not the pixels. Read the first version as a stranger who needs something: can you tell within ten seconds what is sold and how to ask for it? Is anything on the page untrue, or true but unimpressive? Is your best fact buried in paragraph four? Those are the notes that improve a site. "I don't like the green" is a note you can leave until last, and often stops mattering once the page says the right things.

Step four: fix it in sentences

Refinements are requested in ordinary language from the same project — "the price should be visible before the gallery", "this reads too formal for a bakery". Direct code editing and export are not part of V1, which is a real limit and worth knowing before you start rather than after. You can see the sort of result this produces in the examples on the home page, and how the interview works alongside them.

Step five: check it the way your customer will

• On a phone, held in one hand, at the size you actually use — not a desktop window made narrow.
• With the contact route reachable by thumb from wherever the decision gets made, not only at the bottom.
• Reading the second and third lines of paragraphs, where typesetting problems live.
• Out loud, once. Anything you would not say to a customer's face should not be on the page.

A generated site fails in the same place a hand-built one does: a visitor who cannot tell in ten seconds what you sell, and cannot find how to ask.

Where AI is genuinely better than a template

• Getting off a blank page, which is where most websites actually die.
• Writing fluently in a language you read well and write slowly.
• Holding typography, spacing and tone consistent across a whole page, which is where hand-edited templates come apart.

And where it is not

• Knowing what you are actually good at, or which of your services makes money.
• Setting your prices, or taking the eight photographs that do more for the page than any copy.
• Deciding that one page is enough, which is usually the correct answer.
• Making you quotable to answer engines by accident — that takes writing for extraction on purpose.

The budget question deserves its own answer: what a small business website actually costs, including the lines nobody quotes for. Ours is $2.99 once for one build, listed in full on the pricing page.