An October 2026 arXiv preprint compared the citations ChatGPT shows in its web app against what OpenAI's Responses API and Google return, and found the API is a usable site-level proxy but a poor page-level one — about a quarter of ChatGPT's exact cited URLs are recoverable.
2026-10-19 · by Moe Ameen
A preprint posted to arXiv on October 11, 2026 — "Understanding ChatGPT Citations: Evidence from Production Web Search Across Model Generations," by Yunxuan Jiang, Shiqi Zhang, Anqi Xing, Xiangyu Chang, and Xiaokui Xiao (Xi'an Jiaotong University, SGIT AI Lab at State Grid Corporation of China, and the National University of Singapore) — set out to answer a practical question people ask when they try to track whether AI answers quote them: can you stand in for ChatGPT's live web app using OpenAI's API or a Google search? The short answer is partly. Neither reproduces the exact pages ChatGPT cites, but both match the *sites* those pages live on much more closely than the specific URLs.
The researchers recorded what ChatGPT's web app actually cited across four settings spanning GPT-5.2 through GPT-5.5 (Instant and Thinking modes), then compared those citations against two external "proxies": the OpenAI Responses API with web search enabled (using gpt-5.3-chat-latest and gpt-5.5) and up to five pages of Google organic results per query. They ran 80 questions, split evenly English and Chinese across recommendation, fact-lookup, synthesis, and analysis tasks, and repeated each question 20 times on the two Instant models. Combined, Google plus the API recovered 28.3% of the exact URLs GPT-5.3 Instant cited and 26.2% of GPT-5.5 Instant's — but 50.3% and 48.7% of the hostnames. Google alone was weak on pages (9.4% and 6.4%); the API's own source lists matched better (around 22–23% of cited pages) and tracked ChatGPT's choices more closely than Google did.
A second theme is that the pool ChatGPT's search draws from is not stable to begin with, which likely carries over to citations too. Across the 20 repeat runs of a single question, roughly 40.8% (GPT-5.3) to 44.5% (GPT-5.5) of the distinct pages ChatGPT's search retrieved as candidates appeared in only one run; the paper separately notes that candidates retrieved more often across runs are also more likely to be cited, so a one-shot check is likely to give an incomplete picture of what the model cites, not just what it searches. The study also found ChatGPT rewrites the user's query before searching (in the 5.3 and 5.5 runs with visible searches, it never used the question's exact wording) and often searches in English even for Chinese prompts.
Two things to keep straight. This is an arXiv preprint, not peer-reviewed, and its data was collected March–July 2026 on models that predate GPT-5.6 and GPT-6, so the specific numbers may not hold going forward. And despite "accuracy" being the word people reach for, the study measures *overlap and reproducibility* — how closely a proxy's sources match ChatGPT's, and how much ChatGPT's own citations vary run to run. It does not test whether the cited pages correctly support the claims in the answer. The authors' takeaway is procedural: monitor at both the page and site level, rerun prompts several times, account for query rewriting, and check which data source any commercial AI-visibility tool is actually using under the hood.
The useful reframing from this study is that being cited is a site-and-surface game, not a single-URL lottery — proxies and ChatGPT agree on the domain far more than the page, and any one page's odds swing wildly run to run. The lever a creator actually controls is how many strong, citable surfaces their domain puts in front of the engines, and how consistently. That is a production problem, and it is what [Kompozy](/) exists to solve. Kompozy generates the long-form and structured formats answer engines tend to pull from — Blog Articles and Email Newsletters for depth, [Carousel Posts](/glossary/hyperframes), Quote Graphics, and Infographic Photos for the quotable, scannable claims — and holds all of it to one voice through the [Persona Brief](/glossary/persona-brief), so a single topic becomes a cluster of consistent assets across your owned and social surfaces instead of one lonely post hoping to get quoted.
Then it ships them. [Autopilot](/glossary/autopilot) schedules and publishes the set across the eight social platforms plus blog and email behind a per-post review, which widens the domain footprint the study says engines match on. Use the preprint's own advice as a workflow: write for the question *behind* the keyword (ChatGPT rewrites it anyway), build the topic as a cluster rather than a one-off, and measure at the site level over many runs. If you want the deeper playbook, pair this with our [generative engine optimization guide](/guides/generative-engine-optimization-and-ai-answer-citations) and [AI search citation optimization](/guides/ai-search-citation-optimization) — Kompozy is the engine that turns that strategy into the volume of citable content it requires. (Kompozy's own copy generation runs on Claude and OpenAI; it is a production-and-publishing engine, not a citation tracker.)
Not exactly. The October 2026 study found OpenAI's Responses API (with web search) plus Google recover only about a quarter of the exact URLs ChatGPT's web app cites, though they match roughly half of the hostnames. The API is a reasonable site-level proxy for ChatGPT citations but a weak page-level one — and it tracked ChatGPT's choices more closely than Google did.
No — and this is the common misread. It measures overlap (how closely API/Google sources match what ChatGPT cites) and reproducibility (how much ChatGPT's own citations vary run to run), not whether the cited pages correctly support the claims in the answer. "Accuracy" of the citations' content was out of scope.
Measure at the site level and run each prompt many times. Because ~40–45% of the pages ChatGPT's search even surfaces as candidates for a question show up in only one of 20 runs, a single check is noise, and API/Google trackers will understate page-level citations while matching roughly half as well on sites. The authors also suggest checking which data source any commercial AI-visibility tool uses under the hood.
Treat it as a content-volume-and-quality problem, not a single-URL bet: build each topic as a cluster of strong, citable surfaces across your blog and social, written for the question behind the keyword. A content engine like Kompozy generates the blogs, newsletters, carousels, and graphics and publishes them across the eight social platforms plus blog and email, widening the domain footprint engines match on.