Retrieval Contamination Worsens in AI Answer Engines as More Than Half of Gemini 3’s Correct Answers Lack Source Support

AI answer engines integrate online information through real-time retrieval, but errors can repeatedly spread between search results and responses when the systems cite other AI-generated content, creating “retrieval contamination.” Oumi tested Google AI Overviews using OpenAI’s SimpleQA benchmark, underscoring that a correct answer does not necessarily mean its cited material can be verified.

The Inference reviewed the research on April 21, 2026. After testing all 4,326 SimpleQA questions, Oumi found that Gemini 3’s accuracy rose to 91% from Gemini 2’s 85%. However, the share of correct answers lacking source support also increased to 56% from 37%, indicating a marked decline in traceability.