01What Self-Improving AI SEO Agents Are and How They Work

Self-improving AI SEO agents are AI systems that write, grade, and fix their own content without you in the middle. You set the standard once, and the agents loop through drafts until the work passes your bar. Julian Goldie calls this the self-improving SEO engine.

A builder model writes a draft. A separate judge model grades it out of 100 and lists every flaw. If the score is below your pass threshold (like 90%), the notes go back to the builder and it fixes them.

This loop repeats until the content scores high enough. Then it publishes. Because the judge is a different model and is told to be adversarial, it finds real problems.

The builder never grades its own homework.

In the episode, Julian shows a live blog post that was created this way. The agents even wrote about the system itself inside the post. The result?

A site went from nothing to 222 clicks per day. The key idea is simple: you stop being the loop, and the machine becomes the loop.

02Main Features of the Self-Improving SEO Engine

The system has several important features that make it powerful for SEO:

  • Two separate models: A builder model writes content, and a different judge model grades it. Using the same model for both would be like rubber-stamping your own work.
  • Definition of done: You write what a high-quality result looks like, for example "a high quality, SEO optimized, humanized piece of content for this keyword."
  • Maximum rounds: You set how many iterations are allowed before the system gives up or publishes.
  • Multiple API options: You can use free APIs like N2 Step 3.7 Flash, cheap frontier models like GLM 5.2, or premium judges like Fusion Council (five models judging together).
  • Memory and logging: Every round is saved to an Obsidian vault. You can review logs later to see what worked and what didn't.
  • Auto-publish: With a Netlify API key, agents can publish directly to your website. You never need to log in.

These features let you turn an hour of manual quality control into a five-minute setup.

03How to Set Up Self-Improving AI SEO Agents

Here is how to set up the loop inside the agent operating system (AgentOS):

  1. Go to the agent operating system inside the AI Profitable Boardroom (aiprofitableboardroom.com).
  2. Find the loop section. It is a text box, not code. Type your definition of done, for example: "a high quality, SEO optimized, humanized piece of content for this keyword."
  3. Optionally, give a starting point if you already have a draft to improve. Leave it empty to start from scratch.
  4. Select the builder model. You can use free APIs like N2 Step 3.7 Flash, or Claude, or GLM 5.2.
  5. Select the judge model. Pick a different model, like GLM 5.2 for a sharp judge, or Fusion Council for premium judging.
  6. Set the maximum number of rounds (iterations). The judge will grade until the score passes your bar or you hit the limit.
  7. Click "Run Loop" and close the tab. The agents work on their own.

For the team method, use a Kanban board instead. Drop a goal into triage. The orchestrator breaks it into tasks, assigns them across profiles (researcher, writer, video producer), and the judge passes work before it moves to done.

04How to Use the Agents for SEO Content Creation

Julian shows two ways to run the system. First, the single loop method. You define what done means.

A cheap builder drafts the content. A separate adversarial judge checks it. In the episode, the judge was GLM 5.2.

The content needed to score 90% or higher to pass. Round one was not good enough. The loop improved step by step.

Eventually it scored 92 and passed. That draft became a live blog post.

Second, the Kanban board method. This scales the same idea to a full content team. The board has columns: triage, ready, running, blocked, done.

The orchestrator takes a goal from triage and splits it into tasks. A researcher profiles keywords. A writer drafts the post.

A video producer creates an explainer video. The judge profile reviews everything before it reaches done. Nothing ships until the judge passes it.

In the episode, Julian shows a blog post and a video created this way.

You can also use the system for videos. An AI avatar script is written by one agent, the video is produced by another, and a judge grades the final video. If it is not good, it loops again.

05Results, Tests and Comparisons Shown in the Episode

Julian shows a live website that used this system. The rankings trajectory went from basically nothing to 222 clicks per day, and it was growing. One blog post created by the agents ranks number one inside Google AI reviews for the keyword "best AI community."

He compares the old way vs. the new way. The old way: you ask AI for a draft, read the whole thing, spot weaknesses, type every fix, paste again, read again, repeat five times. It takes an hour of quality control, and you lose energy and focus.

You settle for "good enough" because you are tired. The new way: you write what done means once, pick a builder and a separate judge, hit run. The judge grades and the builder fixes.

It takes five minutes. Every round is saved, so you can learn from past work.

Julian emphasizes that the win is not a better prompt. A better prompt still puts you in the loop. The win is not being in the loop at all.

06Tips, Limits and Mistakes to Avoid

Julian shares several tips and warnings from his experience:

  • Never use the same model for builder and judge. If the same model writes and grades, it will rubber-stamp its own work. The judge must be a different model and told to be adversarial.
  • You define the bar, not the machine. Some people think their work is too nuanced for a machine to judge. But the machine does not decide what is good. You write the standard. The judge just checks the work against your bar every round.
  • Free APIs work fine. You can use free models like N2 Step 3.7 Flash for the builder and a free model for the judge. No expensive tokens needed.
  • Set a reasonable maximum round count. The system loops until it passes or hits the limit. If you set too few rounds, it may not finish. Too many might waste API credits.
  • Do not overcomplicate the definition of done. Keep it in plain English. For example: "Beautifully formatted, reads like a human, SEO optimized."
  • Log everything. The memory system (Obsidian vault) saves every round. Use it to review what works and what does not.

Avoid the belief that you just need to prompt better. A better prompt still puts you in the loop. The real time saver is removing yourself entirely.

07Who Self-Improving AI SEO Agents Are For and Why It Matters

This system is for anyone who creates SEO content regularly: bloggers, marketers, agency owners, and business owners. If you are tired of reading and editing every draft, this lets you set the standard once and walk away. Julian says even beginners can do it because the agent operating system uses text boxes, not code.

It matters because you stop being the bottleneck. The old way takes an hour of quality control per post. The new way takes five minutes.

You can scale content production without burning out. The agents can also learn from each other because the memory system saves past logs. Over time, the system gets smarter.

Julian also notes that the community at the AI Profitable Boardroom has 3,600 members building similar systems. You get coaching calls, daily tutorials, and a 30-day roadmap. The point is that self-improving loops are not just for SEO.

You can apply them to videos, code, or any project a judge can check. But for SEO specifically, it solves the biggest problem: quality control without your personal time.