Rebuilding My Firm's Website With an AI Pair

What was broken

My consulting firm's website looked busy and said too much.

The first screen had 116 animated dots and pulses, plus a bouncing arrow. It pulled a stock photo from another site. A popup jumped out after five seconds. Two floating buttons sat on top of each other.

The words had a bigger problem. The home page showed nine company names. A visitor would read them as clients. They were not. They were places where my partners and I had worked as employees. The case studies had the same fault, in a stronger form, because a case study tells a story.

Some service pages had no real work behind them at all. If a buyer asked "show me one you did", we had no answer.

Search told the same story. In six months the site earned 112 clicks. 107 of them went to the home page. Every article had zero clicks, ever. The case studies page sat third in results, and only 1.4 percent of people who saw it clicked. The site was being shown. Nobody wanted to open it.

What I built

I rebuilt the site over four days, from 24 to 27 August, with an AI coding partner. I made 51 changes to the code in that time. I made the calls. The AI did the heavy typing and the measuring.

I started by taking things out. The animation, the stock photos, the popup and the second button all went. Then I set one simple layout for every section. A thin line, a small label on the left, the content on the right. Nothing centred, no boxes. The eye starts in the same place every time.

Then came the honesty pass. I checked every claim against our real work records. The nine names are now labelled as past employers. Every figure says whose it is. One savings claim did not add up, so it went. All eight case studies now carry a line saying the work was done in an operating role, not for a client. I cut five service pages that nothing on our records could back. The old links now point to the nearest real page.

Then the parts a visitor never sees. Only 3 pages carried the markup that helps search engines read a page. Now 47 do. The sitemap was written by hand and had fallen behind. 15 pages added through the site editor were missing from it, so they had never been sent to search engines. Now the sitemap is built from what actually ships, and it went from 30 pages to 47.

Last, I rewrote page titles from real search data. Every title led with our own words instead of the reader's problem. The worst one opened with "Not client engagements." True, and it stays on the page. But as the first words in a search result, it gave nobody a reason to click.

What changed

The new site was merged on 27 August, after my partner signed off.

The most useful finds came from looking at the live pages, not the code. Three layout faults were invisible in the code and obvious on screen. One button had been squeezed into four stacked words.

A bigger one: when a visitor clicked from one page to the next, the browser tab kept the old page's title. Our traffic reports recorded the wrong page name on every click. My first fix looked like it worked. Then I checked again and found two copies of every hidden tag, the old and the new. That looked fixed while still being wrong. The second fix cleared it.

By the evening of 30 August, 0 of 48 pages had a title over 60 characters. That morning it had been 11. The update that shipped them ran about five hours late. I waited instead of forcing it, and that told me the job was slow, not broken.

Not everything is done. On 30 August, 3 of 17 site checks still failed. One of them was 8 pages with no visitor in 28 days. That is a demand problem, not a code problem, and now it has a number.

What I would do differently

I would check every claim before I design anything. The honesty pass changed more than the new look did.

I would look at every change on the live page, not just in the code. Most of the real faults only showed up there.

I would never keep a sitemap by hand. A list kept by hand drifts.

And I would test a fix twice. The fix that looks right is the one most likely to hide the next fault.