Publish fake stats, get cited in LLMs
Fabricating Data to Manipulate LLMs: A Case Study in GEO
By Metehan Yesilyurt | GEO Researcher at Peec AI | October 7, 2026
What happens when you feed the internet 50,000 completely fabricated statistics? I decided to find out. The result was a startling revelation about the current state of AI search: ChatGPT will cite your data even if Google has completely erased you from the map.
📉 The Experiment: High Impact, Low Effort
About seven months ago, I launched a project to test the boundaries of Generative Engine Optimization (GEO). While most industry experts focus on how to get a brand mentioned in an AI response, I wanted to see what happens after the response is generated—specifically, how that "hallucinated" or fake data propagates through the real world.
The Setup
I didn't spend months on this. In fact, the entire infrastructure was built in a single day.
Project Checklist:
- Purchase a new domain:
stateglobe.com - Generate 50,000 fake statistics using
GPT-4.1 - Target niches: SEO, e-commerce, social media, content marketing, and web tech
- Scope: Data spanning 200 different countries
- Deploy via a basic AI web design template
The Financial Breakdown:
| Item | Cost |
|---|---|
| Domain & Hosting | ~$30 |
| AI Template/Tools | ~$20 |
| Total Investment | ~$50 |
🌪️ The Divergence: Google vs. ChatGPT
The aftermath of the launch provided a perfect case study in the difference between traditional search indexing and LLM retrieval.
1. The Google Collapse
Initially, the site performed well. I used indexer tools to get nearly every page live within a week. However, the honeymoon was short-lived.
Google Visibility: High Google Visibility: Zero
Two consecutive Google algorithm updates decimated the site's rankings. I even tried creating a subdomain to claw back some impressions, but it was futile. For all intents and purposes, Google "forgot" the site existed.
2. The ChatGPT Persistence
While Google turned its back, ChatGPT doubled down. According to Cloudflare logs, the site received over 600,000 bot requests in the last 30 days alone. These weren't random crawlers; they were the agents ChatGPT uses to ground its answers in real-time.
"Each request represents ChatGPT fetching a fake statistic to answer a real user's question."
🚀 The Ripple Effect: From Fake Stats to Real Citations
The most alarming part of this experiment wasn't just the AI traffic, but the human trust that followed. Because ChatGPT cited stateglobe.com, humans trusted the output and shared it.
The Results:
- Traffic: 10,000+ sessions originating directly from ChatGPT (via GA4).
- Backlinks: 330+ referring domains.
- High-Authority Citations: Links from industry giants including Shopify, DHL, and Bluehost.
- Real-World Integration: These fabricated numbers migrated into academic papers and investment research from banks.
Visibility Comparison
By October 7, 2026, I tracked the site's visibility for specific prompts using Peec's internal tools.
🧠 The Core Lesson: Retrieval Ranking
This experiment proves that ranking on Google and being retrievable by an LLM are two entirely different technical challenges.
How LLM Retrieval Works
Unlike a keyword-based search, an LLM follows a complex pipeline to ground its answers:
As shown in the diagram, a page can be invisible to a standard Google search but still be a prime candidate during the Candidate Retrieval and Reranking phases of an LLM's process.
Technical Summary
If we define the probability of being cited as , the formula for GEO is not:
P(cite) = P(Google Rank #1)
Instead, it is:
P(cite) = P(Retrieval) * P(Reranking) * P(Grounding)
🖼️ Evidence Gallery
Below are the visual data points from the experiment tracking the bot traffic and the resulting citations.
Final Thought: If you are only tracking your Google Search Console, you are blind to a massive channel of visibility. The AI index is a different beast entirely—and as this experiment shows, it is dangerously trusting.