AI content ranked fine for nine months then dropped 70% overnight — post-mortem

One site, 1,140 articles, roughly 80% AI-drafted with light human editing. Nine months of steady growth to 214k sessions/mo. Then an 11-day slide to 64k. Here is everything I can establish, and what I cannot.

What dropped
Not uniform. Broken down by page type:

• "Best X for Y" listicles: -91%. Effectively gone
• Informational how-tos: -74%
• Pages with original data, photos or a named author with a real profile: -12%
• Pages I had rewritten by hand in the previous six months: -8%

What did not change
No manual action. No technical change. Traffic from the same queries now goes to older, worse-looking pages on sites with a visible author.

What I think happened
The pattern maps almost exactly onto "is there evidence a specific person did this work". It does not map onto word count, publishing velocity, internal links, or which model wrote it.

What I am doing
Rewriting the top 120 pages by hand with a named author, original photos and first-hand testing. Six weeks in, those pages are up 34% while the rest keep sliding. Expensive — about $18k of writer and photography time — and it is working.

The lesson I would actually take
AI made publishing cheap. It did not make the thing being rewarded cheap. I spent nine months confusing the two and the correction cost more than the content did.

The -8% on hand-rewritten pages against -91% on listicles is about as clean a natural experiment as you will get on a live site. That is a genuinely valuable dataset.

This is the thread I point people at when they tell me the pipeline is the business. The pipeline is a cost reduction. It was never a moat.

Were the listicles monetised with affiliate links at a higher rate than the how-tos? Asking because commercial intent pages have been getting hit independently of how they were written.

Yes, and that is a real confound I cannot separate. The listicles were the money pages. Some of that -91% may be commercial-intent filtering rather than anything about authorship.

Appreciate you naming the confound instead of letting the neat story stand.

The $18k rewrite cost is the number people should sit with. Cheap content has a deferred bill and this thread is the invoice.

As someone about to start publishing: should I just not use AI at all?

Use it for research, outlines, and editing your own writing. Do not use it to produce the thing a reader reads unless you are adding something only you have. On day one, what only you have is your own effort.

Same curve showed up on social about eighteen months earlier. Cheap production floods the channel, the platform raises the bar, the people who were only cheap get removed.