The Dark Side of GEO Metrics: Investments vs. Fake Numbers

Generative Engine Optimization (GEO) and AI Visibility technologies are fast becoming the most critical metrics companies report to the board, yet they operate in a dark reality: measurements are unreliable. An arXiv study found that identical queries to the same assistant often return different brands, with only 40-50% overlap in cited sources. Worse, Perplexity showed even lower consistency at 33%. Even within the same product, ChatGPT's different modes sometimes recommended different brands for the same prompt, as revealed by Semrush research. A brand's citation share could swing from 34% to 90% depending on whether the buyer described themselves as a freelancer or an enterprise. The instability doesn't stop there. Assistants decompose questions into hundreds of sub-queries, with Google's Deep Search potentially firing off hundreds of parallel sub-queries for a single question. This means GEO is more like a continuous experiment than a fixed ranking. A brand's standing shifts weekly without intervention, as the model changes, competitors update, or the customer's history evolves. The noise floor is so high that bootstrap confidence intervals show typical variations spanning three to seven percentage points. A 3% to 4% increase in citation share could easily be random noise.

The biggest flaw in GEO measurement is that it only captures a fraction of the actual influence, with referral logs undercounting the true impact. This makes it nearly impossible for CFOs to determine whether GEO investments are paying off. The industry is now caught between waiting for better metrics and risking misallocated budgets. This is becoming the sector's most critical risk.