I Retooled My AI Content Pipeline After an Expert Audit: What Changed and Why

You know how sometimes you think you’re doing everything right, only to find out that half of it isn’t working? That’s exactly what happened when I had a technical SEO expert take a look at my AI article pipeline. Out of 12 components, only 5 were really doing their job. Ouch, right? Three were just taking up space, and one was actually messing things up!

Let’s dive into what I changed and why it matters.

The Original Pipeline

After cranking out over 70 articles, my system had become a bit of a tangled mess. Here’s what I had:

• A made-up Information Density formula ((entities + facts) / word count) that I thought measured article quality
• Em dash policing that flagged every dash as an AI giveaway
• Forced question headings - every H2 had to be a question
• Automatic speakable schema injected everywhere
• A rigid 20-40 word answer requirement after each heading
• An Obsidian RAG limited to just my vault notes
• An Information Gain tool that pulled entities and facts

I really thought this was a solid setup. The expert kindly pointed out how misguided I was.

What Got Cut

The Information Density Formula (Rating: 5/10)

That ID = (E + F) / W formula? Yeah, it turns out that’s just wishful thinking. Google never mentioned ranking by entity density. The expert called it out as speculation, and honestly, they were spot on.

I swapped it out for a Rich Quality Score that checks what really matters:

• Real-world examples
• Actual commands and code snippets
• Performance benchmarks
• Experimental results
• Documented failures
• Honest tradeoffs
• Visual aids like screenshots

Now, an LLM analyzes articles based on these evidence types. If it scores 7 or above, it’s high quality. Below 3’ Well, that needs some work.

Em Dash Cleanup (Rating: 1/10)

When I mentioned the em dash policing, the expert chuckled. Turns out no search engine or AI cares about how often you use an em dash. So, I just deleted that metric from content_auditor.py. What a relief!

Forced Question Headings (Removed)

Making every H2 a question felt unnatural. The expert suggested I aim for 30-50% questions instead of 100%. After all, “What Are Docker Layers - sounds pretty robotic compared to the straightforward “Docker Layers.”

I updated enforce_inverted_pyramid in apply_fixes.py to only flag extreme cases (below 20% or above 70% questions).

20-40 Word Answers (Replaced)

Instead of sticking to rigid answer blocks, the new flow is: heading → brief intro → detailed explanation. Because let’s be real, that’s how people actually read!

Speakable Schema (Deprioritized)

I had automated schema markup for voice search, but the expert told me it’s barely supported by Google and not useful for tech content. So, I removed it from check_meta_faq.py.

What Improved

Obsidian RAG → Full Knowledge Search (Rating: 10/10)

Previously, my system only searched research/vault/. The expert encouraged me to “Use everything you’ve got.”

Now, the new obsidian_rag.py searches through seven knowledge bases:

• research/vault/ - Core notes
• research/experiments/ - Test results
• research/terminal-logs/ - Command history
• research/benchmarks/ - Performance data
• research/bookmarks/ - Saved resources
• research/github-notes/ - PR discussions
• research/project-logs/ - Postmortems

This means every article kicks off with relevant context from across my digital brain.

Information Gain → Evidence Focus (Kept, refined)

The core idea of building on my own experiments and data still stands strong. That’s something AI farms can’t easily replicate.

I kept the tool but changed its scoring. Now, it tracks original evidence instead of just key phrases.

The Results

After implementing these changes, my system is running smoothly:

• 315 pages generated
• 0 errors
• ~45 second builds
• Pagefind indexed 315 pages (~8k words)

The article count jumped from 69 to 70, but honestly, the real victory is that every new piece now has:

1. A Rich Quality Score based on solid evidence
2. Context from my entire knowledge base
3. A Quick Summary for both humans and machines
4. Natural headings - questions where they fit

Key Takeaways

The biggest lesson I learned’ Stop optimizing for imaginary signals.

That density formula, em dash rule, forced questions, and voice markup? All of that was chasing phantoms. They just added complexity without any real benefit.

What really worked was simpler:

• Use actual experimental data
• Let the structure flow from the content
• Cut out the unnecessary complexity

The expert really hit the nail on the head: “Your edge isn’t more content - it’s more original evidence.”

That quote now sits proudly above my monitor.

What do you think? Have you ever had to rethink your own processes like this? I’d love to hear your thoughts!

References & Further Reading

• Wikipedia: Content marketing
• OpenAI: ChatGPT