Selected citations are sources explicitly credited via inline annotations or text segment mappings in an AI model's final output, confirming they actively informed the response's assertions. Unselected citations encompass the broader set of candidate web pages and grounding chunks retrieved during search execution that the model evaluated, browsed, or ingested into context but decided not to reference.

By parsing provider-specific grounding metadata across models like Gemini, Claude, and ChatGPT, platforms can monitor this selection funnel to evaluate how search algorithms filter authoritative content:

• Conversion Tracking: Calculating a domain's trust ratio (the ratio of selected citations to total unselected retrievals) measures an AI's confidence in that domain's content.
• Extraction Dynamics: Content that is retrieved but unselected indicates that a page ranks well in underlying web retrieval systems, yet fails to provide the concise, extractive factual answers required by the language model.
• Competitive Analysis: Analyzing unselected competitor URLs provides valuable competitive intelligence, identifying topics where an AI model looks at competing domains but rejects their specific claims.