You can spot AI-written content in two seconds. Not because AI is bad at writing. Because every AI tool produces the same voice.
Same sentence openers. Same transition phrases. Same structure. "It's important to note." "Let's dive in." "In today's rapidly evolving landscape." You've read that article a thousand times. It's on every SaaS blog, every niche site, every content farm that discovered ChatGPT in 2023.
The problem isn't that AI writes badly. The problem is that AI writes like every other AI.
The AI Slop Problem
Feed ChatGPT a keyword and tell it to write a blog post. You'll get something that's grammatically correct, factually acceptable, and completely forgettable. Every paragraph sounds like it was written by the same intern at the same content agency.
Here's what AI defaults to:
Generic openers. "In today's digital landscape..." "As technology continues to evolve..." "When it comes to [topic]..."
Filler transitions. "It's worth noting that..." "Another important aspect is..." "Additionally, it's crucial to understand..."
Weak conclusions. "In conclusion, [topic] is an important consideration for..." "By implementing these strategies, you can..."
Passive voice everywhere. "The content can be optimized by..." instead of "Optimize your content by..."
Zero specificity. "Improve your meta descriptions" instead of "Your title is bleeding 4,800 impressions because the keyword isn't in the first 30 characters."
Google's helpful content system was built to detect this. Not because Google hates AI. Because Google hates content that adds nothing. And when every AI tool produces the same structure with the same phrases targeting the same keywords, none of it adds anything.
Why This Happens
LLMs are trained on the internet. The internet is full of mediocre content. The model learns to produce the average of what it's seen. That average is corporate blog voice — safe, generic, forgettable.
When you prompt "write a blog post about SEO," the model reaches for the most common patterns associated with SEO blog posts. Those patterns are the filler phrases, the predictable structure, the hedge-everything tone.
It's not a bug. It's what you asked for. You asked for "a blog post" and got the statistical average of all blog posts.
The Fix Isn't Better Prompts
The prompting advice is always the same. "Be more specific." "Give examples of your tone." "Tell it to write casually."
This helps a little. It doesn't solve the problem.
Even with detailed prompts, you get output that sounds 30% like you and 70% like ChatGPT pretending to be you. The model still reaches for its defaults. It still adds filler. It still hedges. One prompt can't override thousands of hours of training on generic content.
The actual fix is a system, not a prompt.
What Actually Works: A Writing Style System
When I built my SEO agent, the content it produced was good but generic. Same problem everyone has. So I built a style system that actually learns your voice from your existing content.
Here's how it works.
Step 1: Analyze existing content.
The system reads your published articles. Not one or two — all of them. It extracts patterns across tone, sentence length, vocabulary, structure, paragraph cadence, and how you use (or don't use) certain phrases.
Step 2: Generate 6 style files.
From that analysis, it creates:
- Tone profile. Are you formal? Direct? Technical? Conversational? What's the ratio?
- Structure patterns. Do you use short paragraphs? Do you lead with conclusions? How do you transition?
- Sentence patterns. Average sentence length. Variation rhythm. Short-long-short patterns.
- Vocabulary. Words you use often. Words you never use. Technical terms you prefer.
- Examples. Actual snippets from your writing that demonstrate your voice.
- Banned words list. 50+ phrases the AI is not allowed to use.
Step 3: Enforce it.
When the agent writes content, these files aren't suggestions. They're rules. The banned words list alone kills the most obvious AI tells instantly.
The Banned Words List
This is the highest-impact, lowest-effort change you can make. Here's a sample of what gets banned:
# AI Slop — Never Use These
"It's important to note"
"In today's rapidly evolving"
"Let's dive in"
"When it comes to"
"In this article, we will"
"As we all know"
"At the end of the day"
"Without a doubt"
"Game-changer"
"Cutting-edge"
"Revolutionary"
"Best-in-class"
"Synergy"
"Leverage" (as a verb)
"Ecosystem"
"Holistic approach"
"Moving forward"
"It goes without saying"
"Needless to say"
"First and foremost"
"Last but not least"
"In conclusion"
"To sum up"
"Delve"
"Landscape"
"Paradigm"
"Streamline"
"Harness"
"Utilize" (just say "use")
"Facilitate"
"Robust"
"Comprehensive" (when meaningless)
"Innovative" (when everything is)
"Excited to announce"
Copy that list. Put it in your system prompt. Your AI output improves immediately.
But a banned words list alone isn't enough. It removes the worst offenders. It doesn't add your voice. That's what the full style system does — it doesn't just stop the AI from sounding like AI. It makes the AI sound like you.
Numbers Beat Adjectives
The biggest difference between generic AI content and content that actually ranks is specificity.
AI defaults to adjectives. "Fast performance." "Affordable pricing." "Great results."
Humans (good ones) use numbers. "15 tokens per second." "$29/month." "68,000 impressions in 9 days."
Every time the AI reaches for an adjective, the style system pushes it toward a number. Not "the Mac Mini is affordable" but "the Mac Mini M4 Pro costs $2,000." Not "the analysis is fast" but "the analysis takes 2 minutes."
Numbers do three things adjectives don't:
- They're credible. Anyone can say "fast." A specific number means you measured it.
- They're memorable. People remember "$2,000" but forget "affordable."
- They're comparable. "15 tokens/second" tells a developer exactly where this sits relative to alternatives.
What This Looks Like in Practice
Without style system:
In today's competitive digital landscape, it's important to have a comprehensive SEO strategy. By leveraging cutting-edge AI tools, you can streamline your content creation process and achieve better results. Let's dive into how you can harness the power of AI to revolutionize your SEO workflow.
With style system:
Your SEO workflow is 5 tabs and 30 minutes of copy-pasting. GSC to Sheets to ChatGPT to your CMS. The AI you're pasting into has never seen your site. Here's what happens when you connect an agent directly to your data.
Same topic. Same AI model. The difference is the style system rejecting every generic pattern and replacing it with voice, specificity, and directness.
How to Build Your Own (Without My Tool)
If you want to do this manually, here's the process:
1. Collect your best 10 articles. Not all your content. Your best content. The posts that got engagement, that you're proud of, that sound like you.
2. Analyze patterns. Read them with a fresh eye. How long are your sentences? How long are your paragraphs? Do you use questions? Do you address the reader as "you"? What words appear often?
3. Write a style document. Not "be conversational and engaging." That means nothing. Write specific rules:
## My Writing Rules
- Max 3 sentences per paragraph
- Lead with the conclusion, then explain
- Use "you" not "users" or "one"
- Numbers instead of adjectives
- Short sentence. Short sentence. Longer sentence with detail. Short.
- Never use passive voice
- Never start with "In this article"
- Technical terms in English, even in German text
4. Create your banned words list. Go through AI-generated content and highlight every phrase that makes you cringe. Add them to the list. Be aggressive. You can always remove something later.
5. Test it. Put the style document in your system prompt. Generate content. Compare it to your actual writing. Iterate.
This process takes an afternoon. The output quality improvement is permanent.
The Verdict
AI content doesn't have to sound like AI. It sounds like AI because nobody tells it not to.
A banned words list kills the worst offenders in 5 minutes. A full style system — tone, structure, patterns, vocabulary, examples — makes AI output genuinely sound like your writing. Not 100%. Maybe 80%. But 80% your voice in 2 minutes beats starting from scratch every time.
The tools that will win the AI content game in 2026 aren't the ones that generate the most words per minute. They're the ones that generate words that sound like they came from a real person with a real perspective.
I built this system into Agentic SEO because I was tired of publishing content that sounded like every other AI blog. Connect it to your site, let it analyze your existing writing, and the agent writes in your voice. Not in ChatGPT's voice wearing your hat.
*Writing about AI, SEO, and building in public at marc0.dev.
Top comments (0)