Study: Google and OpenAI API Miss Most ChatGPT Citations

Researchers at the National University of Singapore and Xi'an Jiaotong University compared ChatGPT's actual cited sources against results from Google search and OpenAI's own search API. Search Engine Journal covered the paper on October 18. The paper first appeared on arXiv on October 11.

The researchers tested four GPT versions, from 5.2 to 5.5, collecting data mostly between March 29 and July 29 of this year. They ran 80 questions in English and Chinese, covering recommendation, fact-lookup, synthesis, and analysis types, each repeated up to 20 times to capture ChatGPT's actual citations.

They ran the same questions through Google search and OpenAI's search API to see how closely each matched what ChatGPT actually cited. Google search pulled up to five pages per question, averaging 43 results, while the API ran with web search enabled the same number of times. Both fell well short of matching ChatGPT's real citations.

Google search alone matched only 9.4% of cited pages for GPT-5.3 and 6.4% for GPT-5.5. Combining Google and the API results still only reached 28.3% and 26.2%. Even measured at the site level rather than exact page, matches stayed around half.

Citation instability was substantial too. When the same question was run 20 times, 44.5% of pages cited (for GPT-5.5) and 40.8% (for GPT-5.3) appeared in just one of those runs. A page being cited once was no guarantee it would show up again.

The researchers also flagged that monitoring needs to track which underlying model produced a given answer. Some collection sessions switched models mid-run, and ChatGPT rarely searched using the question's exact original wording.

The researchers recommended revisiting any monitoring strategy whenever the model behind ChatGPT changes. These findings predate GPT-5.6 and GPT-6, so current numbers could look different.

The study didn't test any commercial rank-tracking or GEO measurement tool directly. But by quantifying the gap built into using search results or API responses as a stand-in, it's a useful data point for any tool that works the same way.

In Korea

If you've been estimating your ChatGPT citation visibility from Google rankings or a search API alone, this study is evidence that approach can diverge substantially from reality. Running the same query multiple times yourself is worth adding alongside whatever tool you use.

Source: Search Engine Journal, October 18, 2026. Checked October 21, 2026. Read the original