Study: repeated GEO optimization homogenizes content

A new arXiv preprint introduces CHASE, a simulation framework testing what happens when content creators repeatedly rewrite documents to match large-language-model ranking signals. Across six domains and 20 rounds of optimization, alignment between ranking position and independently assessed quality declined in every domain tested, and the authors first validated ranking as a stand-in for LLM citation at a rank-citation AUC of 0.853. A control run without the ranking incentive showed no such decline, isolating the effect to optimization pressure rather than rewriting itself.