OpenAI API and Google results found only a quarter of ChatGPT's cited pages, preprint finds
A preprint from researchers at the National University of Singapore and Xi'an Jiaotong University has found that neither OpenAI's Responses API nor Google's organic results reproduce the pages ChatGPT cites. Combined, the two stand-ins surfaced 26.2% of the pages cited by GPT-5.5 Instant and 48.7% of the sites behind those citations.
The team recorded the data flow into ChatGPT's web app across signed-in sessions, capturing the searches it ran and the pages it selected as sources. They put 80 questions from the Search Arena dataset to four model versions, 40 in English and 40 in Chinese, running each 20 times on GPT-5.3 Instant and GPT-5.5 Instant. Google results alone accounted for 6.4% of cited pages and 22.8% of cited sites on GPT-5.5 Instant.
Results varied sharply between runs: 44.5% of the distinct pages returned for a given GPT-5.5 Instant question appeared in only one of its 20 runs. ChatGPT also rewrote queries rather than searching the original wording, and for Chinese questions it searched in English alone in 69.6% of visible GPT-5.5 Instant runs.
The authors recommend tracking at both page and site level, repeating the same prompt, and checking whether a tool's figures come from the API or the ChatGPT app. Every model tested predates GPT-6, so the numbers describe the versions current from March to July 2026.
Reported by Search Engine Journal ↗
Topics: ChatGPT, Citations, Measurement