A new study found OpenAI's API and Google together recover only about a quarter of the pages ChatGPT's app cites, and about half of the sites, so any audit that asks ChatGPT through the API has to say so.

Google's changelog still ends at the 16 October Indexing API quota entry, and the status dashboard shows no new ranking incident since the September spam update. Perplexity's October entries are unchanged from yesterday and still name no Sonar model. Anthropic's newest post is still the 15 October evaluator update. On the deprecations page, `claude-haiku-4-5-20251001` is still Active with "Not sooner than October 15, 2026" and no notice, and `claude-sonnet-4-5-20250929` still retires 30 November. Search Engine Roundtable's newest post is its 16 October recap. ChatGPT's release notes could not be read today (Releasebot did not respond).

The researchers recorded the data stream behind their own signed-in ChatGPT web sessions, which shows the searches ChatGPT ran, the pages it got back and the pages it chose as sources. They used 80 real questions from the public Search Arena dataset (40 English, 40 Chinese), ran each 20 times on GPT-5.3 Instant and GPT-5.5 Instant, and collected mostly between 29 March and 29 July 2026. Within three days of each app run they asked the same questions through Google (via SerpAPI, about 43 results per question) and through OpenAI's Responses API with web search, 20 times each.

Combined, Google and the API found 28.3% of the pages GPT-5.3 Instant cited and 26.2% for GPT-5.5 Instant, and 50.3% and 48.7% of the sites. Google alone found 9.4% and 6.4% of cited pages. At the level of named entities (pulled from the titles and snippets of cited pages, not from the answer text), the best combination recovered 50.7% and 45.1%; in the authors' words, "Even the union recovers only about half of the cited entities." ChatGPT's app never searched the exact wording of the question on GPT-5.3 or GPT-5.5, and 44.5% of the distinct pages it retrieved for a question showed up in only one of the 20 GPT-5.5 runs. Two collection sessions switched model partway through.

Limits: a preprint, not peer reviewed; 80 general questions, half in Chinese, none about local or service businesses; every model tested predates GPT-6; it measures cited pages and the entities on them, not which businesses an answer recommends; and it tested no commercial tracking tool.