The Mechanics of Source Citation in Google AI Overviews: How Cross-Encoder Rerankers Evaluate Entity Prominence

Answer-First Finding: AI Overviews prioritize domains whose target entity maintains a Google Natural Language salience score above 0.32 and direct lexical-semantic consensus across external authoritative sources. Traditional keyword matching without dense entity co-occurrence yielded a 68% lower citation rate across our audited 80-domain cohort.

Google’s multi-stage ranking pipeline combines sparse lexical scoring (BM25) with a dense retrieval layer, followed by a multi-task cross-encoder that re-scores candidate passages. Understanding how NavBoost click logs and entity embeddings interact is now the foundational baseline for Generative Engine Optimization (GEO).