The new if/else agency website is live. Mockups, design, development, content, marketing, SEO, GEO: AI produced or reviewed all of it, and humans who know how to use it ran the whole process.
Our last redesign took months. This one took 5 days.
We're sharing how we did it because the method matters more than the result. Anyone can ask an AI to build a website. Getting one that holds up, keeps its search rankings and actually tells a story is another matter.
Before producing anything, we set up a team of AI agents. Each one had its own role, skills and scope, the way people do on a real project team.
Yes, it burns more tokens than giving one agent the whole job. But it saves us from constantly redoing work, and from spending our days at the screen proofreading every line it writes.
We did our research and set our aesthetic criteria, but we based the design on what our audience wants, not on our own tastes. That's the classic trap with agency websites: designing to please yourself instead of the people who will actually read it.
We started from a child theme of Jingle, our HubSpot theme. That mattered to us: our own website had to prove that Jingle is versatile, fast, customizable and accessible.
The rule was simple: use everything Jingle offers out of the box, and only build new modules when something is missing. It's exactly what we do for our clients.
To move fast without losing consistency, we had Python scripts generate the sections and templates from the Figma mockups. The result is around 50 sections and 14 templates. A block that appears on ten pages lives in a single file.
Finally, we built a local preview that mimics HubL, HubSpot's templating language, so we could check desktop and mobile rendering in Chrome before pushing anything to HubSpot.
A redesign that tanks your SEO is a failed redesign, so we were very methodical about it.
We documented every page of the old site in a file: its URL, copy and images. We first rebuilt the pages as drafts in a staging account, on the same URLs and with their language groups.
Then we had all the content rewritten to fit our tone, our audience, the story we wanted to tell and the keywords we were targeting.
This is the important part: in production, we didn't recreate the pages. We wrote the new version into the draft of each existing page. Same ID, same URL, same analytics, same hreflang tags. We backed up every page before touching it, and an automated audit checked everything (all 56 pages passed) before we hit publish.
Going live didn't mean we were done.
Our job came down to making the strategic calls, fixing errors in the code and passing on our best practices.
Were there misfires? Yes. Saying everything worked the first time would be a lie. AI gets things wrong, which is exactly why you need people who can read what it produces: developers and designers who know what needs to be done and can check the work.
Did it save us time? Massively. Our last redesign took months. This one took 5 days.
Our job is changing fast, and we have to accept it. That doesn't mean doing whatever we like, or treating AI as a solution on its own. It needs experts behind it: people who know their craft and can steer it and check what it does.
When AI can produce just about anything, what counts is the vision we bring to it and the decisions we make. That's the difference between a site generated on autopilot and one that actually works for your business.
Want to talk about your own website? Book a call with Garance.