If an agency brings you a report proudly claiming your brand "shows up in a third of AI answers," don't take their word for it. I dug into a generative engine optimization (GEO) experiment by Cossa, and the numbers are a cold shower.
Over 5 weeks, 53 expert articles were published across 8 platforms. A month later, visibility was tested against 20 target prompts across 6 AI channels with 3 runs per prompt (360 sessions and 1,887 source links in total).
At the domain level, the numbers looked great: 559 links (around 30% visibility). But measuring exact URLs of the published articles revealed the real picture:
- Out of 1,887 links, only 14 pointed directly to the articles (under 1% actual citation rate).
- Those 14 links led to just 5 unique pieces out of 53.
- The remaining 48 articles were completely ignored by the AI models.
The gap between apparent domain-level reach and actual citations was 40x. RAG algorithms split articles into chunks: if your text lacks dense facts and actionable numbers, the search crawler discards it. Without tight structure, your content stays invisible to AI search.
