Can you rebuild a website in 6 hours with AI agents? Aurem’s build log

Why rebuild a website that worked?

Our old site displayed properly and had visitors, and people could reach us through it. It also had flaws you cannot see with the naked eye. We listed them on 27 September 2026, before touching anything:

• the heading font, Cormorant Garamond, was declared but never loaded, so every large heading displayed in Times;
• the page measured 1,692 pixels wide on a 1,440 pixel screen, and 408 pixels on a 390 pixel phone: it overflowed sideways;
• it contained 15 main headings (H1), including the FAQ questions, when a page should have only one;
• the copy switched between informal and formal address from one section to the next, with a few typos in the French (“optimisattion”, “ACCOMPAGENEMENT”);
• the contact form said nothing about how personal data would be used;
• several figures (“3X ROI”, “−80 %”, “x2 leads”) had no source.

That last point settled everything else. An agency that asks its clients never to claim anything without proof cannot display figures it is unable to justify. We removed them all. The only figure left on the home page is a count of coffees shared with business owners: 14 at the time of going live.

How was the work split between humans and agents?

We broke the build into phases. Each one was handed to a specialised agent or kept by a human. The durations come from our build log.

Total: just under 6 hours. That total includes neither the two rounds of back-and-forth on the mock-up nor the next day's work on loading speed. We would rather say so than show a flattering stopwatch.

Scoping and acceptance testing together account for half the time. They are the two moments when a human makes the calls and checks the result.

What do the agents actually do, and what stays human?

Before launching the first agent, we wrote a brief. It sets the scope (one home page, in two languages, fully editable), lists what is forbidden, and states the criteria that say when the work is done.

The prohibitions were precise: no figure without a source, no fake “our team”, no AI-generated images, no dashes in headings. For the agent producing the page, one rule mattered more than the others: use only blocks and settings that already existed on the site, without inventing a single one. A script then checked every block, the number of main headings and how faithfully the copy matched the approved file.

The acceptance criteria can be checked without debate: a single H1 per language, no error when opening the page in the editor, nothing overflowing at 1,440 or at 390 pixels.

What stayed human:

• Scoping. Four multiple-choice questions on voice, the extent of the change, languages and visuals, then sign-off on the brief.
• Proofreading of every piece of copy and of the mock-up.
• Positioning. On the evening of the launch, we rewrote the opening of the home page around one idea: understand before you build. No agent would have suggested it on its own, because it comes from conversations with clients.
• Going live. When everything was ready, the agent tried to lift maintenance mode and the safeguard in its environment stopped it. Clément did it himself, in two clicks. It is the rule we apply for our clients: the agent builds, and the human presses the button that commits the business.

Can the site still be edited without the agency?

That was the starting condition. Aurem's site runs on WordPress with the Divi 5 builder. A custom-coded theme would have meant calling a developer for every change. So the agent generated the page directly in the format Divi uses for its own modules, and then we imported it.

The result has 173 modules per language, which open without errors in the visual editor, exactly as if they had been placed by hand. You change a piece of text, a photo or a button with the mouse.

The rest of the set-up is deliberately simple. The site is hosted by Hostinger®, and Aurem is a Hostinger partner. The English version runs on the free Polylang plugin. The cookie banner gives “Accept” and “Reject” equal weight, and no analytics cookie is set before the visitor agrees.

What broke along the way?

A build log without the breakdowns would be of no use to anyone. Here are the main ones.

The heading font fell back to Georgia. The plugin that serves fonts from our server named some files with a space (“italic-300 700.woff2”). The browser rejected the rule and displayed a fallback font. We rewrote those declarations with encoded URLs.

The mouse wheel was slow. A “smooth” scrolling option in the theme reduced each notch to 56 pixels. We turned it off.

An animated counter stayed frozen, because the builder stripped out a CSS instruction when compiling the stylesheet. A small script replaced it.

Speed on mobile was the biggest job. At the start, Lighthouse gave 59 out of 100 for mobile performance, and the main content took 9 seconds to appear. There were several causes: a loading screen, around twenty unused plugins loading their code on every page, a cache with no optimisation at all, and a 9 second animation on the opening photo. We dealt with them one by one on 30 September.

Two issues remain open, and we are publishing those too: there is no navigation menu on mobile yet, and a 528 KB stylesheet still slows down mobile rendering.

What were the results, and how were they measured?

The starting point (59 for mobile performance) was measured with Lighthouse, the end point with PageSpeed Insights. Both tools use the same engine, but under different conditions: the gap gives an order of magnitude and should not be read as a point-by-point comparison.

We are not giving any figures for enquiries or sales. The site has been live for too short a time to measure anything meaningful, and we will publish those figures once they exist.

What we would do for you

Speed alone is a poor reason to rebuild a website. A website that loses you enquiries is a good one. For you, we would therefore start with an audit of your website, your advertising, how your enquiries are followed up and what could be automated, ending with three or four priorities at most. Only then would we apply the same method as on our own site: a brief written before building, agents to move quickly, acceptance testing done by a human, and a website in your name that you can edit yourself.

The audit is free: book 30 minutes.

Frequently asked questions