Preprint proposes defense against manipulated generative-search citations

A new arXiv paper describes GEO Defender, a two-stage defense for generative-search systems facing documents rewritten to influence citation preferences. The authors say its reranker demotes manipulated documents while a training-free generation layer guides source use without fine-tuning the target model. In experiments across seven attacks and five closed- and open-source LLMs, the paper reports average attack success falling from 50.32% to 6.20% while retaining 94.12% of benign-evidence use. These are preprint results.