Google search results and OpenAI's Responses API together surfaced only 26.2% of the pages ChatGPT cited in tests of the GPT-5.5 Instant setting, according to an Oct. 11 arXiv preprint from researchers at the National University of Singapore and Xi'an Jiaotong University. At the site level, the same combination found 48.7% of cited sites — meaning any rank-tracking setup built on Google results or the API sees roughly half the domains and a quarter of the URLs ChatGPT actually references.

The study covered four ChatGPT versions, from GPT-5.2 Instant to GPT-5.5 Instant, with most data collected between March 29 and July 29, 2026. Note the caveat: this is an arXiv preprint, and arXiv does not peer-review papers.

How did the researchers observe ChatGPT?

They recorded the data flow inside their own signed-in ChatGPT web sessions, capturing the searches the model ran, the pages those searches returned, and the pages ChatGPT then selected as sources. A page counted as cited if ChatGPT picked it at least once across 20 runs of the same question.

The test set was 80 questions drawn from Search Arena, a public database of real user interactions with search-enabled chatbots — 40 in English, 40 in Chinese, spanning four request types: recommendation, fact lookup, information synthesis and analysis. Each question ran 20 times on GPT-5.3 Instant and GPT-5.5 Instant. GPT-5.2 Instant and GPT-5.4 Thinking got fewer runs because OpenAI decommissioned them before collection finished.

For the stand-ins, the researchers pulled up to five pages of Google organic results per question via SerpAPI — roughly 43 results per question on average — and sent 20 queries per question to gpt-5.3-chat-latest and gpt-5.5 through the Responses API with web search enabled, all within three days of the matching ChatGPT runs.

What did the numbers show?

The combined Google-plus-API overlap was 28.3% of cited pages for GPT-5.3 Instant and 26.2% for GPT-5.5 Instant, rising to 50.3% and 48.7% at site level (subdomains counted as separate sites). The breakdown:

• Google results alone: 9.4% and 6.4% of cited pages; 26.1% and 22.8% of cited sites.
• The API's full source list: 22.9% and 22.5% of cited pages.
• The API's own answer citations overlapped more closely: GPT-5.5 Instant cited 23.3% of the pages the API cited, versus 16.4% for GPT-5.3 Instant.

The practical takeaway for SEOs: an API- or Google-ranking-based visibility test will match sites far more often than exact pages, so page-level comparisons against ChatGPT's reported citations understate true overlap.

Did ChatGPT rewrite queries and vary its sources?

Yes, on both counts. OpenAI's help page confirms ChatGPT typically transforms a query into one or more specific search terms when it uses an external provider. In observed runs, GPT-5.3 Instant and GPT-5.5 Instant never searched using the exact original question wording; GPT-5.2 Instant did so in only 5.5% of runs.

Language handling matters too. For Chinese-language questions with visible searches, GPT-5.3 Instant (67.0%) and GPT-5.5 Instant (69.6%) searched in English only. That aligns with earlier Peec AI testing, reported in February, which found 43% of fan-out queries for non-English prompts ran in English.

Run-to-run instability was substantial: 44.5% of distinct pages returned for a question appeared in only one of 20 GPT-5.5 Instant runs; the figure was 40.8% for GPT-5.3 Instant. Overlap was stronger at site level than page level. The authors caution that pages appearing in more runs were more likely to be cited, but that doesn't prove citation frequency drives selection. Two collection sessions also switched models mid-stream after 14 and 16 runs, so the authors advise verifying which model produced each response — echoing the earlier CiteChoice preprint, which found rerunning identical inputs flipped 15% of cite-or-not decisions.

What should trackers do now?

The authors recommend monitoring at both page and site level, running the same prompt multiple times, and accounting for ChatGPT rewriting prompts into its own searches. They also advise checking whether any tracking tool comparing webpage-level results to ChatGPT citations collects data through the API or directly through the ChatGPT app — the paper tested no commercial tool.

Every setting tested predates GPT-5.6 and GPT-6. OpenAI's model help page says GPT-6 is rolling out in Chat to eligible Plus, Pro, Business and Enterprise users, with GPT-5.6 Luna becoming the default for Free and Go users. Around the GPT-5.5 change in May, Sistrix measured day-to-day variation in German-language ChatGPT citations jumping to 47%, against a normal 1% to 2%. Until someone repeats this comparison on GPT-6, the paper's figures describe only the models current from March to July — and the authors say monitoring strategies should be re-evaluated each time the underlying model changes.