<aside> 🚨 Want more notes like this? Subscribe to weekly strategy emails: ainotebook.com
</aside>
Howdy Notebooker!
A piece by Charles Floate has been making the rounds with a title built to detonate in your feed: "I Reverse Engineered LEAKED System Prompts For AI SEO. Google Lied. Anthropic is Targeting SEOs."
It's from Charles Floate, and he brought his A-game. Confident, provocative, stuffed with quotes pulled from what he says are leaked system prompts of Claude Opus 4.7, GPT 5.5 and Gemini 3.
When Charles puts it on your radar, you know it’s good!
After reading it and really letting it sink in, I’ve decided to break down Charles’ article for you guys today.
Two weeks ago Google published its first official guide to optimizing for generative AI search. The headline: AEO and GEO aren't separate disciplines. It's still SEO.
The guide also "busts" a few myths. You don't need to chunk your content. You don't need llms.txt or special markdown files. Chasing inauthentic mentions won't help you.
Charles's argument is that Google's guide is a polished policy document, while the system prompts that actually run these models tell a different story.
He says the prompts reveal a hidden scoring layer that decides whether your content gets used or ignored. That the models clearly evaluate content in chunks. That they filter hard by source quality and treat SEO content as suspect.
His conclusion: there's a new discipline he calls Retrieval-Layer SEO, Google is hiding it from you.
Here's the part worth your time.
Charles describes three layers inside a frontier model. What it learned in training. What it pulls from the web when you ask a question. And a referee that decides which of those two to trust when they disagree.
He calls that referee the arbitration layer. Think of it like a courtroom.
Your content isn't the judge reading your essay and nodding along. Your content is one witness. The model's training is a second witness. A competitor's page is a third. The referee weighs all three and decides whose story makes it into the final answer.