We Published 500 Articles With AI. Here Is What Happened.
We built an AI pipeline that drafted about 500 articles, then measured. Four clicks in ninety days. What went wrong, what experience changed, and the circuit breakers we run now.
What we built
Four Eyes has been building websites since 1998, and for most of that time the articles on a site did a simple job: bring in search traffic, turn some of it into calls. When language models got good enough to draft an article from a brief, we built a pipeline around them. Topic research, a briefing interview, a draft, external links, images, an editorial pass, and a push to WordPress. It could turn a topic into a competent 2,000-word draft in minutes.
We ran it hard. About five hundred articles came out of it. That was deliberate: we wanted to know what happens when a site publishes at the speed the tools allow.
What we measured
On September 23, 2026, we pulled ninety days of Google Search Console data, June 22 to September 20, for every URL on foureyes.com, and joined it to the archive as it stood then: 510 published articles, the AI drafts plus older hand-written pieces.
| Measure | Result |
|---|---|
| Published articles at the time of the audit | 510 |
| Clicks from Google to all of them, 90 days | 4 |
| Articles with zero search impressions in 90 days | 343 (67%) |
| Clicks to the whole site in the same period | 28, almost all to the home page and the services hub |
| Articles competing with another article for the same topic | 28 about choosing an agency, 24 about AI search, 18 about website cost, 12 about redesigns |
| Articles left after human review and consolidation | 296 (42% removed) |
Four clicks. Not four percent, four clicks, across more than a million words.
What volume alone got wrong
The writing was competent. That is the uncomfortable part. Three things went wrong, and none of them was the prose.
The pipeline competed with itself. Ask a tool to write about website cost, and it will happily write eighteen articles about website cost, each with slightly different numbers. Google sees eighteen candidates for the same query, picks none with confidence, and ranks a competitor who wrote one good page. Two of our cost articles disagreed on the price bands. A prospect who read both would trust neither.
Nobody was accountable for any single page. When someone writes an article, they’re embarrassed if it is thin. When a pipeline writes it, nobody reads it after it is published. We found a 44-word page ranking near the top for a branded query, with a form that did not render. It had been live for five months.
Housekeeping did not keep pace with output. 165 articles had been redirected in an earlier cleanup but never unpublished, so they sat on the blog index as cards leading to a redirect. 691 internal links pointed to redirected addresses. 35 old posts redirected to a page that no longer existed. None of that is an AI problem. It is what happens when a site produces content faster than anyone looks at it.
What experience changed
The pages that earned impressions had something in common: they were long, specific, and answered one question a real buyer asks. Which performance monitoring tools are worth it. How to run a redesign without losing rankings. What belongs in a website RFP. Which WordPress host to choose. The pipeline drafted some of those too.
The difference was that a person had decided each one was the only page on the site about its subject, and had made it useful enough to read to the end.
So we did what the pipeline could not do for itself. We read the data, kept the pages that earned their place, merged overlapping ones into one strong page each, and redirected the rest. 510 became 296. We rewrote every internal link so none passes through a redirect. Conversion tracking went in, so we now judge an article on whether it produces an inquiry, not a visit.
The circuit breakers
We did not throw the pipeline away. We use it every week. We added checks that make experience part of the process, not an afterthought.
- Nothing publishes itself. Every draft lands in WordPress as a draft. A person reads, updates, improves, and approves it, or it never goes live.
- One page per topic. Before a draft is written, the pipeline classifies the request against what the site already has: new page, improve an existing page, or consolidate several. New only wins when nothing covers the ground.
- A person sets the brief. The list of what to write comes from client calls and search data, through a briefing interview, not from a tool asked to fill a calendar.
- Every number is ours or cited. Price bands, timelines, results: from our own work, or linked to the source. The editorial pass scores each draft before anyone sees it.
- Consolidation writes its own redirects. When two pages become one, the old addresses redirect to the survivor in the same step, so rankings move with the content.
- The data gets read. Search performance is pulled into the same tool, and anything that earns nothing over a quarter is merged or removed.
Why this matters if you are buying a website
What happened to our blog is what happens to AI-built websites. The tool produces something competent, fast, and cheap. What it cannot do is decide whether the site is the right site, notice when it stops working, or take responsibility when it doesn’t perform. That part is still a person with enough experience to know what to look for. If you want to know whether you need one, we wrote an honest guide: when an AI website builder is enough, and when it is not.
Already built one, and it is not earning? We take those on.