AI platforms cite sources differently, impacting transparency for users. The number of references can change with each query. Understanding these differences is key for publishers and SEOs.
SEO expert Den Petrovich has analyzed how leading AI platforms—Google Gemini, OpenAI ChatGPT, and Anthropic Claude—manage source attribution and grounding. His findings reveal that the way these systems cite sources can vary significantly, even for identical queries.
For example, when users ask to see which sources an AI referenced, the platforms may alter their responses. The same prompt can yield seven, two, or nine cited sources, depending on the system. This means that the phrase “AI mentioned us” only makes sense when considering the specific citation behavior of each platform.
OpenAI’s approach separates broadly recognized publications from narrowly cited ones, creating a visibility trap for publishers. Being among the 37 pages OpenAI has read is very different from being one of the two actually cited. Both metrics matter, but they are not equivalent.
Google’s redirection technology adds another layer of complexity. Any tool that reads data from Gemini Grounding must first resolve URLs before matching them to publisher domains, making direct attribution less straightforward.
Anthropic Claude offers the most comprehensive “considered set” of sources, including both cited and rejected fragments. However, it hides the actual text of these fragments, requiring a second pass to reconstruct what was reviewed. This process consumes additional time and tokens, complicating real-time analysis.
