TLDR
Google Search and OpenAI's Responses API together matched fewer than 28 per cent of exact pages ChatGPT cited in real signed-in sessions, an arXiv preprint published 11 October 2026 found. Matching by website reached around 50 per cent. Results also varied run to run: 44.5 per cent of distinct cited pages appeared in only one of 20 repeated runs.
KEY TAKEAWAYS
The gap between the stand-ins and ChatGPT
Tools that estimate how often ChatGPT cites a brand cannot see inside a user's ChatGPT session, so they lean on stand-ins: Google results, or OpenAI's Responses API with web search switched on. A preprint from researchers at the National University of Singapore and Xi'an Jiaotong University checked both against what ChatGPT actually cited in real signed-in sessions.[2]
Combined, Google and the Responses API matched 28.3 per cent of the pages GPT-5.3 Instant cited and 26.2 per cent for GPT-5.5 Instant. Google on its own matched 9.4 and 6.4 per cent.[2] Roughly three in four cited pages turned up in neither stand-in.
Matching by website rather than by page did better: 50.3 per cent for GPT-5.3 and 48.7 per cent for GPT-5.5.[2] That is the difference between knowing a brand's site was in the mix and knowing which article ChatGPT actually pointed people to.
Share of pages ChatGPT (GPT-5.5 Instant) cited that each stand-in also found
How the study was run
Between 29 March and 29 July 2026 the researchers recorded the data flowing through their own signed-in ChatGPT sessions to see what it searched for, what came back and what it cited. They used 80 questions from Search Arena, 40 in English and 40 in Chinese, and ran GPT-5.3 Instant and GPT-5.5 Instant 20 times on each one.[2]
The answers moved around a lot. For GPT-5.5, 44.5 per cent of the distinct pages cited for a question showed up in only one of the 20 runs.[2] In runs where its searches were visible, GPT-5.3 and GPT-5.5 never searched using the question's exact wording, and on Chinese questions around two-thirds of runs searched only in English.[2]
What it means for AI visibility reporting
For Australian brands paying for AI share-of-voice reports, the study is a caution about method rather than a verdict on any product: it tested no commercial tracking tool.[2] A report built on one run of each prompt, typed exactly as a customer would, against Google or the API, is measuring something that only partly overlaps with what ChatGPT shows.
The authors argue generative engine optimisation should treat being retrieved and being cited as separate measures, allow for ChatGPT rewriting queries, and track sites and entities as well as exact URLs.[1] Search Engine Journal adds that every setting in the paper predates GPT-5.6 and GPT-6, so any method should be retested as models change.[2]
The paper is a preprint posted to arXiv on 11 October and has not been peer reviewed.[1]
SOURCES & CITATIONS
James Dore is founder and editor of Prompt the Market and Strategy Director at Bushnote, a Sydney strategy, SEO and AI search agency. He has a background in strategy, media and financial services, and led consumer product marketing at MONEYME. He writes on AI search, answer engines, technology policy and media economics.
