Google and the OpenAI API together recover only about a quarter of ChatGPT's cited pages
A preprint that five researchers posted to arXiv on October 11, 2026 compared the pages ChatGPT cites in web search with what Google results and the OpenAI Responses API return. Used together, the two outside signals recovered roughly a quarter of the cited pages. The work has not been peer reviewed.
Outside tools did not come close to predicting what ChatGPT cites
The paper, titled "Understanding ChatGPT Citations", went onto arXiv on October 11, 2026. Its authors are Yunxuan Jiang, Shiqi Zhang, Anqi Xing, Xiangyu Chang and Xiaokui Xiao. They studied four ChatGPT Web settings running from ChatGPT 5.2 to 5.5, spanning both Instant and Thinking modes, and looked at how ChatGPT rewrites a question, which pages it pulls in as candidates and which ones it ends up citing. Google Search and the OpenAI Responses API served as the outside yardsticks.
The abstract's verdict is blunt: even combined, the authors write, "they recover only about one quarter of cited URLs". The detailed percentages below come from Search Engine Journal's October 18 coverage of the paper.
Page-level matches are low, site-level matches approach half
According to Search Engine Journal, the study drew 80 questions from Search Arena, split evenly between English and Chinese, and put every question through GPT-5.3 Instant and GPT-5.5 Instant twenty times each. Taken together, Google results and the API located 28.3% of cited pages for GPT-5.3 Instant and 26.2% for GPT-5.5 Instant. Counted by site, those figures rise to 50.3% and 48.7%. Google alone matched just 9.4% and 6.4% of pages, while the API's full source list alone matched 22.9% and 22.5%. The paper adds that both proxies sit closer to ChatGPT at the site and entity level than at the exact-address level.
The same report highlights two behaviors. In visible searches, ChatGPT never used the exact wording of the original question in GPT-5.3 or GPT-5.5 runs. For Chinese questions with visible searches, 67.0% of GPT-5.3 runs and 69.6% of GPT-5.5 runs searched only in English. Results also shift between runs: for a given question, 44.5% of distinct pages (GPT-5.5) and 40.8% (GPT-5.3) showed up in only one of the 20 runs. Most of the data was gathered between March 29 and July 29, 2026.
Retrieval visibility and citation visibility are separate things to measure
The authors conclude that generative engine optimization (GEO) should treat being retrieved and being cited as two different kinds of visibility. Across repeated runs, frequently retrieved candidates tended to be cited more, though SEJ notes the authors do not read this as proof that frequency drives citation; and showing up in Google or in an API source list does not by itself explain which address ChatGPT will name. The paper also suggests accounting for rewritten queries and run-to-run variation, and watching site-level and entity-level presence alongside exact addresses. It adds that strategies deserve a fresh look whenever the deployed model or serving environment changes.
There are limits. This is a preprint, and arXiv does not run peer review. Every tested model predates GPT-5.6 and GPT-6, so the results speak only to the versions in use between March and July. Our October 12 issue covered Ahrefs' report on ChatGPT shifting toward official and primary sources, a different angle on the same subject.
The sources say nothing about Türkiye, so this section is commentary
Neither source mentions Türkiye or Turkish. The questions were prepared only in English and Chinese, so there is no test with Turkish queries. This section is not a fact drawn from the sources; it is explicitly UNALSOFT commentary.
Our reading: since Chinese questions often ended up searched in English, it is reasonable to suspect language choice can shape the pool of pages ChatGPT draws on. Whether that holds for Turkish queries cannot be said from this study. For a business in Türkiye, the one cautious takeaway is to test your own questions in ChatGPT by hand rather than guessing visibility from Google rankings.
UNALSOFT's take
We read this as a lesson in measurement. A Google ranking or an API output explains only part of what ChatGPT cites, so "visibility in AI answers" deserves its own goal, separate from rank tracking. In practice that means a site that is clear, citable and crawlable, both page by page and as a whole. In web design we structure content, page titles and core facts with that in mind. Because the sources contain nothing on Türkiye, this piece does not suggest a local conclusion.
Are your pages ready to be visible in ChatGPT?
A short conversation is enough to review your site's content structure together.