How AI website builders work

An AI website builder takes a short description of your business, has a language model turn it into a site plan, text and image choices, and then renders that plan through the builder’s own page components or as generated code. Hosting, domains, forms and payments are handled by conventional software that was there before the AI.

Vendors do not publish their internals, so this page describes what the products visibly do and what their public pages say. The stages are the same across the market even when the names differ.

The short version

• The model’s work is the draft: plan, words, pictures, styling choices and search fields.
• Hosted builders assemble; code tools write. One fills the vendor’s building blocks, the other produces source files.
• Publishing is not AI. A generated site is served by the same infrastructure as a hand-made one.
• Every stage has a review you should do, and the copy review is the one people skip.

1. Intake: a questionnaire, a prompt or an existing site

The builder first has to learn what the site is for. There are three common ways in.

• A fixed questionnaire. Durable’s flow begins with “Answer 3 questions about your business” (Durable).
• A free-text prompt. Wix describes its current builder as replacing the “rigid questions” of its retired ADI tool with a description in your own words (Wix).
• Something you already have. 10Web accepts “a prompt, a link or your files” (10Web), and Hostinger can start from the address of an existing site (Hostinger).

Whatever the form, your input is short. The builder wraps it in a much longer set of its own instructions, often called a system prompt, before the model sees it. To produce a page the builder can display, those instructions have to tell the model which sections exist, what the design rules are and what form the answer must take. This is why two sentences are enough to get a result, and also why the result reflects the builder’s taste more than yours.

2. Planning: a site map before any page exists

The model’s first output is usually a plan: which pages the site needs, which sections go on each, and what tone to write in. A dentist gets Services, Team, Insurance and Booking; a photographer gets Portfolio, Pricing and Contact.

Some products show you this step. 10Web advertises an “AI planner that shows every page before it builds”. If your builder shows a plan, correct it there. Removing a page from a list is quicker than deleting it after it has been written and styled.

3. Assembly: filled building blocks, or generated code

This is the stage where products differ most, and the difference decides what you can do later.

Assembled sites. In a classic hosted builder, the model does not write a web page. It picks sections from the vendor’s library (a hero, a three-column feature list, a gallery, a contact form), chooses a colour and font theme, and fills each section’s text and image slots. In effect its answer is a list of choices, not markup. The builder’s existing rendering engine turns those choices into the page. The evidence is in what happens next: the result opens in the same drag-and-drop editor as a hand-built site, and vendors stress that you can then edit it by hand.

Generated code. A prompt-to-code tool has the model write source files, typically a JavaScript framework project with HTML and CSS, then builds and previews them. Nothing confines it to a section library, so it can produce layouts and behaviour that no template offers. The price is that small visual changes are requests to rewrite code, and mistakes can be real bugs.

Hostinger shows both approaches in one product. Its page describes a mode where you “get a ready-made site instantly, and fine-tune it visually from there”, and a second, “Agentic mode, also known as vibe coding”, where building and customising both happen in a chat. Types of AI website builders sorts the market along this line.

4. Copy: fluent, generic and unverified

The model writes every headline and paragraph from your few sentences plus what it knows about businesses like yours. The second source is much bigger than the first. So the text is grammatical, well organised and largely interchangeable with any competitor’s.

It can also be wrong. Google’s guidance for site owners notes that generative models “don’t retrieve facts, but predict a likely sequence of words based on their training data”, and asks you to fact-check AI-generated content before it goes live (Google Search Central). Read every generated sentence that states a fact about you. Delete the ones you cannot stand behind.

5. Images: matched stock or new pictures

Images arrive one of two ways. The builder searches a stock photo library for pictures matching each section, or it calls an image-generation model. Squarespace’s page, for instance, says you can “generate a new visual, transform an existing photo, or build a cohesive image set” (Squarespace).

Either way the pictures are placeholders for a real business. A generated photo of “your team” is not your team, and a stock kitchen is not a kitchen you fitted. Replace them before launch, and check the vendor’s terms for how generated or stock images may be used.

6. Search fields: titles, descriptions, alt text

Because the model has just written the page, the builder can ask it for the fields search engines read: a title and meta description per page, alt text for images, sometimes structured data. 10Web states that each site comes with “titles, meta descriptions, schema, a sitemap and redirects set up”.

Two cautions apply. The sitemap and redirects are produced by code, not by a model, and are either correct or not. The titles and descriptions are generated text, and they are what a searcher sees first. Google recommends descriptions that “accurately describe the specific page” (Google’s snippet guidance), so edit them like any other copy.

7. Publishing: conventional infrastructure

Pressing Publish hands the site to systems that involve no language model: web hosting, a content delivery network, an SSL certificate, and DNS for your domain. Forms, bookings, shops and payments are likewise existing product features that the generator switches on.

This matters when you compare products. Two builders with similar generators can differ widely in how fast their pages load and how accessible their components are, because those qualities come from the platform and not from the model. What to check before you choose shows how to measure both on a generated site.

8. Editing: where the time actually goes

After generation you change the site in one of two ways: directly, by clicking and typing in a visual editor, or by asking the AI in a chat box. Squarespace’s wording is typical: “make hands-on edits, ask your AI assistant to help, or move between both approaches”. WordPress.com says of its two modes that “neither mode locks you out of the other” (WordPress.com).

The two kinds of edit work differently underneath. A direct edit changes one stored value, such as a heading’s text, and nothing else. A chat edit, in outline, sends your request and the current state of the page to the model, which returns a new version of the section or the code. That is why a chat edit can alter things you did not mention, and why each one uses model capacity that the builder has to pay for.

Where the model is, stage by stage

Stage	Language model involved?	Your review
Intake	No, it only collects your answers	Give specifics: location, services, audience
Planning	Yes	Check the list of pages
Assembly	Yes, choosing sections or writing code	Check structure on a phone
Copy	Yes	Fact-check every claim
Images	Sometimes, or a stock search	Replace with real photos
Search fields	Yes for text, no for sitemaps	Rewrite titles and descriptions
Publishing	No	Compare platforms, not generators
Editing	Only when you use the chat box	Learn the manual editor too

If the vocabulary here is new, the glossary defines each term in a sentence.

Frequently asked questions

Do AI website builders use templates?

Most hosted builders do, in the sense that the model chooses and fills pre-built sections from the vendor's own design system instead of writing a page from nothing. That is why the result opens in a normal drag-and-drop editor. Prompt-to-code tools work differently and have the model write the source files directly.

Which AI model does a website builder use?

Vendors rarely say, and it changes. Builders generally call large language models from the major labs and add their own instructions, design system and checks around them. Framer is one that lets you pick a model for its agent. For a buyer, the editor and the publishing platform matter more than the model name.

Why do AI-generated websites look alike?

Two reasons. A language model produces the most typical answer for your kind of business, and a hosted builder assembles that answer from a fixed library of sections. Sites from the same builder therefore share a family resemblance until someone replaces the stock text, photos and section order with their own.

Is the AI involved after the site is published?

Only when you ask it to change something. Serving pages to visitors is done by ordinary hosting, a content delivery network and a security certificate. No model runs when someone loads your home page, unless you have added an AI feature such as a chatbot to the site yourself.