SCI-Defense: Detecting Persuasive Structured Manipulation of LLM Ranking in Generative Engines

SCI-Defense uses an LLM to review five types of persuasion tactics and detect GEO attacks that manipulate LLM rankings. On 1,152 CORE attacks across six Amazon product categories and six MS MARCO web domains, it detected 89.6% of reasoning-type and 96.2% of review-type attacks at a 1.2% false positive rate. Flagged entries are downranked rather than rewritten; on a subset of 120 queries, only 0.6% of honest entries fell out of the top three, versus 4.0-4.5% for rewriting and truncation.