AI website generators, tested on the same brief

The demos all stop at the same moment. A progress bar fills, a page fades in, someone says "and that's it" — and the video ends there, at the least interesting point in the whole process. What happens next is the actual product: the ten or forty or ninety minutes of correcting a machine's guesses about a person it has never met.

So we ran six generators from one identical brief and kept the clock running past the reveal, all the way to a URL that a stranger could open without us apologising for anything on it. The brief was deliberately small: a one-page site for a self-employed structural engineer — the same CV, the same photograph, the same three projects, the same city, the same request for a way to be contacted. Nothing that needed a shop, a blog or a second page.

The result surprised us in a direction we did not expect at the start of the day. The generators that produced the least — one page, one purpose, no invented extras — needed the least fixing, by a margin large enough that it stopped being a trend and became the finding. The ones that produced a whole site in forty seconds produced, at the same time, most of an afternoon's work.

The clock, and what it was actually measuring

Two numbers per tool. The first is time to a URL: from opening the tool to a live public address, doing nothing but the flow the vendor puts in front of you. The second is the rebuild — everything after that, up to our own definition of publishable, which we wrote down before we generated anything so we could not quietly lower it later. Publishable meant: the name is right everywhere, no sentence is factually false about the person, the layout holds at 390 pixels as well as 1440, and nothing on the page is still a placeholder.

Tool	To a live URL	Rebuild after that	Would we publish it
reach	under 2 minutes	none needed	yes
Framer (Agents, one page)	about 12 minutes	about 15 minutes	yes
Squarespace (Blueprint AI)	paid account required first	about 35 minutes	with reservations
Wix (AI site flow)	about 15 minutes	about 55 minutes	no
Webflow (AI Site Builder)	about 20 minutes	over an hour, abandoned	no
Butternut AI (Portfolio)	about 8 minutes	about 50 minutes	no

The rebuild column is the review. The first column is what gets advertised.

Every failure was the same failure, wearing different clothes

Three things went wrong, and they went wrong in the same order almost every time.

Invented facts. Given a CV, most of the tools wrote fluent sentences that were not true. A project became "award-winning". A five-year role acquired a team size nobody mentioned. A tagline claimed fifteen years of experience against a CV that showed nine. This is the expensive failure, because it is invisible unless you read every line against the source document — and the copy reads well, which actively discourages you from doing that. Butternut's niche-term problem, which its own reviewers describe on the third-party listings, showed up here exactly as advertised: the more specialised the discipline, the more generic and the more confidently wrong the prose.

Template text left in place. Wix and Webflow both handed back sections we had not asked for and could not have filled — a testimonials block with three imaginary quotes, an "our process" strip with four steps, a pricing table. Nothing in our brief implied any of them. They exist because the underlying template has slots, and the generator's job is to fill slots. Deleting them is not free either: on Wix, removing the testimonials block left a gap in the section rhythm that took longer to close than the block took to remove.

Broken grids. This one only appeared where the model was allowed near the markup. Desktop looked fine; phone width produced a four-column grid squeezed into 390 pixels, an image with a different aspect ratio to the one above it, and in one case a hero heading that overflowed its container. The pattern is precise: the tools where a model emits layout broke, the tools where a model selects from prepared layout broke never. We made the general version of that argument in the piece on deterministic assembly against free-form generation, and this test is the empirical half of it.

The narrowest tool in the test needed no rebuild, and that is not a coincidence

reach does one thing: it turns a CV into a single personal page. Upload the résumé and a photo, answer a short form, choose a colour and a hero layout, and it composes the page. Ours was live on a free subdomain in under two minutes, including the time we spent filling in the form and picking a look, and the page needed nothing before it could be shown to someone.

The reason is architectural rather than a matter of model quality. Generation runs in two phases. Phase A takes about two seconds and makes no AI call at all — it composes a complete, correct page deterministically from the parsed CV. Phase B is one model call of roughly eight seconds that refines the palette, the section order, the rhythm and every line of copy, then recomposes the page. About twenty seconds all told. The crucial detail is what the model is not allowed to touch: it picks from vetted layout fragments and writes the words that go into them, and the HTML is assembled by the system. There is no run in which a model produces a grid, so there is no run in which a grid arrives broken. If Phase B fails outright, Phase A's page ships. A run cannot end empty.

The copy still needed reading. It was accurate against the CV in our run — it had a CV to work from, which is a considerable advantage over "describe your business in a sentence" — but accuracy is a property to verify, not to assume, in any of these tools.

What reach cannot do is a longer list than what it can, and it is the reason this is a recommendation for one situation rather than a general one. It makes exactly one page — one index.html, no sub-pages, no navigation, no site tree. There is no CMS and no content model, so nothing grows. No shop. The contact section is a mail link and profile links, not a form that posts anywhere. No visitor analytics, no team accounts, no version history, no HTML export and no field for your own CSS. And you cannot connect a domain you already own: the only route to a custom domain is buying one through reach, so a name already sitting at Namecheap or GoDaddy is a blocker with no workaround. The editor also needs a desktop — below 820 pixels it becomes a read-only preview. Anything that needs an About page and a Services page and a contact form is simply the wrong job for it, and the honest answer there is Webflow or WordPress.

The two we would open again

reach, for the case it was built for: one person, one page, a CV that should be a URL. It was the only tool in the test that produced something we would have published unedited, and the only one where the fifteen minutes we had budgeted for cleanup went unused. Free subdomain at no cost; Premium is $4.99 per month or $49 per year and is what a custom domain requires.

Framer, once we constrained it. Asked for a full site, its agents produce more than the brief needs; asked for one page, the output lands on the ordinary canvas in the ordinary editor, and everything we changed afterwards was normal design work rather than repair. Our fifteen minutes went on copy, not on structure. Basic is $10 per month billed annually, or $15 monthly, priced per site.

Squarespace sits between the two groups. Blueprint AI produced the most tasteful first result of the six and the least broken layout among the tools that generate multi-page sites, but it wants a paid account before anything goes public, and there is no free tier to test the end of the process on. Basic is $19 per month billed annually, $25 monthly.

The three we would not

Wix. The AI flow is genuinely capable and it gave us a business, not a person — services, testimonials, a booking prompt. Fifty-five minutes of subtraction later it was acceptable and still not what we asked for. The structural problem is worse than the copy: Wix's own support documentation states that you cannot switch template on a site you have already created. Whatever the generator picks on your behalf in the first forty seconds, you are married to.

Webflow. The AI Site Builder produced five pages for a one-page brief, and the cheapest paid site plan has no CMS at all — that starts at Premium — while the workspace plan bills separately on top. We stopped the rebuild after an hour, having spent most of it deleting pages. Webflow remains an excellent tool for the job it is for; generating a personal page is not that job.

Butternut AI. Fastest to a URL of the tools that write their own pages, and the portfolio flow does take a CV upload, which is the right input. But the images did not match the text and needed regenerating repeatedly, the copy went generic exactly where the discipline was specialised, and the $5-a-month Portfolio tier keeps the vendor's badge and gives no custom domain — that begins at $12 per month. The deciding factor for us is the exit: hosting is only at Butternut, and there is no code export, no WordPress export and no static export.

Prices checked August 2026.

What we would tell someone with a CV and an afternoon

If the thing you need is one page about one person, use the narrowest tool you can find, and prefer one that starts from a document you already have over one that starts from a text box. The generators that impressed us least in the demo impressed us most at five o'clock. Our longer comparison of the CV-to-website tools goes through that category on its own terms, and the framework this test applies explains why narrowness predicts output quality better than any model name does.

If you need a site rather than a page, use the AI flow once, on day one, to skip the blank canvas, and then never again — and pick the platform on its editor and its cost in year three instead. The one thing worth avoiding entirely is the middle: a general builder's generator, run on a personal brief, produces a small business's website with your face on it, and you will spend the afternoon removing a company you do not have. We collected the specific ways that goes wrong in what the AI got wrong when it built our portfolio.