Google search results and OpenAI's Responses API together located just 26.2% of the pages ChatGPT cited in tests of the GPT-5.5 Instant setting, according to a preprint posted Oct. 11 by researchers at the National University of Singapore and Xi'an Jiaotong University. The two proxies fared better at site level, finding 48.7% of the sites behind those citations.

For marketers who track brand visibility in AI answers, the gap is the story: neither stand-in reproduced ChatGPT's citations exactly, which means common tracking methods built on Google rankings or OpenAI's API undercount what users actually see. The paper is an arXiv preprint, and arXiv does not peer-review submissions — a caveat worth holding against every figure below.

How did the researchers test ChatGPT?

The team recorded the data flow inside their own signed-in ChatGPT sessions, capturing the searches the model performed, the pages those searches returned, and the pages ChatGPT selected as sources. A page counted as cited if ChatGPT chose it as a source in at least one of 20 runs per question.

The test set was narrow:

• 80 questions drawn from Search Arena, a public database of real user interactions with search-enabled chatbots
• 40 questions in English, 40 in Chinese
• Four request types: recommendation, fact lookup, information synthesis and analysis
• Each question run 20 times on GPT-5.3 Instant and GPT-5.5 Instant

Data collection ran from March 29 through July 29, 2026. GPT-5.2 Instant and GPT-5.4 Thinking received fewer runs because OpenAI decommissioned them before collection finished — itself evidence of how fast the ground shifts under any monitoring methodology.

To build the stand-ins, the researchers pulled up to five pages of Google organic results per question through SerpAPI, averaging about 43 results per question. They also sent 20 queries per question to gpt-5.3-chat-latest and gpt-5.5 through OpenAI's Responses API (the interface that lets developers query models directly) with web search enabled, all within three days of the matching ChatGPT runs.

What do the numbers show?

Combined, Google and the API found 28.3% of the pages cited by GPT-5.3 Instant and 26.2% of the pages cited by GPT-5.5 Instant. At site level — with subdomains counted as separate sites — the figures rose to 50.3% and 48.7%.

The components tell different stories. Google's results alone accounted for only 9.4% and 6.4% of cited pages, and 26.1% and 22.8% of cited sites. The API's full source list performed better, surfacing 22.9% and 22.5% of all cited pages. Notably, the overlap tightened between newer models: GPT-5.5 Instant cited 23.3% of the pages the API itself cited, against 16.4% for GPT-5.3 Instant.

Site-level matches ran consistently higher than page-level matches. In plain terms: a tracking tool will find your domain in AI answers far more often than it will find the specific URL users actually see.

Did ChatGPT search the way users asked?

No. In all runs with visible searches, GPT-5.3 Instant and GPT-5.5 Instant never queried using the exact wording of the original question. GPT-5.2 Instant did so in only 5.5% of runs. OpenAI's own help page explains that ChatGPT typically transforms a query into one or more specific search terms before searching.

Language compounded the distortion. For questions written in Chinese, GPT-5.3 Instant used only English in 67.0% of runs with visible searches, and GPT-5.5 Instant did so in 69.6%. The paper's authors echo earlier findings from Peec AI, reported in February, that 43% of fan-out queries for non-English prompts ran in English.

Run-to-run instability was severe. Of the distinct pages returned for a given question, 44.5% appeared in only one of the 20 GPT-5.5 Instant runs; the figure was 40.8% for GPT-5.3 Instant. Pages appearing in more runs had a higher likelihood of being cited, though the authors caution this does not prove citation is driven by frequency of appearance.

A separate CiteChoice preprint, covered previously, found that rerunning identical inputs flipped 15% of cite-or-not decisions. Two collection sessions in this study also switched models mid-stream after 14 and 16 runs, leading the authors to recommend checking which model generated each response.

What should monitoring look like?

The authors' guidance is concrete: track at both page and site level, run the same prompt multiple times, and account for ChatGPT rewriting prompts into its own searches. They also advise verifying whether any tracking tool compares webpage-level results against ChatGPT data collected via the API or directly through the web app — the two do not align. The paper tested no commercial tracking tool, so vendors' accuracy claims remain unverified here.

Do the findings still apply?

Possibly not. Every setting in the paper 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, while GPT-5.6 Luna becomes the default for Free and Go users. Around the GPT-5.5 change in May, Sistrix data showed day-to-day variation in German-language ChatGPT citations jump to 47%, against a normal 1% to 2%.

The authors say monitoring strategies should be re-evaluated whenever the model behind ChatGPT changes. Until someone repeats this comparison on GPT-6, these figures describe only the models that were current from March to July.