Grounding metadata consists of structured API response payloads—including search queries, retrieved web chunks, and character offset mappings—that reveal which web sources an AI model accessed and cited.