Study: OpenAI's Responses API captures only a quarter of ChatGPT's citations

Researchers from Xi’an Jiaotong University and the National University of Singapore observed four ChatGPT Web configurations spanning GPT-5.2 through GPT-5.5, in both Instant and Thinking modes, and found that pages retrieved more often across repeated runs were also more likely to be cited. Replaying the same queries through Google Search and OpenAI’s own Responses API as external proxies, the study found the two combined recovered only about one-quarter of the exact URLs ChatGPT Web cited, though a model combining both signals matched citations more closely at the hostname and entity level, reaching AUCs above 0.74 for GPT-5.3 and GPT-5.5.

Instant settings stayed mostly single-round, while GPT-5.4 Thinking frequently ran multiple retrieval rounds — a difference the authors say means GEO tracking should be re-checked whenever a deployed ChatGPT setting changes.