A page can follow every AEO checklist and still go uncited in ChatGPT. In the study's replay, ChatGPT only cited pages its own search had retrieved first, and the usual on-page features say little about which pages that search picks.

The numbers are specific. Among pages on the same website, page features explained 2.8% of the difference in ChatGPT citations. What ChatGPT's search retrieved explained 32%.

What the study measured

Mahe Chen and Nick Koudas ran two 90-day collections of daily answers from ChatGPT, Gemini and Perplexity. One covered 651 questions written for a US fiber internet provider, with 143,249 answers. The other covered 55 questions about small business-software vendors, with 14,850 answers. Together that is 158,099 answers and 468,037 citations.

They then replayed every question through the engines' APIs to see what each engine searched for, which pages it retrieved and which it cited. Perplexity could not be replayed, so the findings on how citations happen come from ChatGPT and, in places, Gemini.

The paper is an arXiv preprint posted on 10 October 2026, and I found no record of peer review. The questions were written for organisations that use the authors' monitoring platform, not sampled from real user logs.

Checklist items did not separate pages on the same site

The authors compared pages within the same website, so that a site's overall popularity could not distort the result. Only one thing held up: pages whose title and paragraphs matched the question were cited more.

FAQ blocks, structured data, summary boxes and page speed were linked to more citations only when comparing different websites. Compare two pages on one site and the link disappeared. The authors' explanation is that popular sites use these features more and get cited more for other reasons.

ChatGPT cites what its own search finds

In the replay, every page ChatGPT cited was one its own search had retrieved. It rewrites each question into new sub-queries every time it answers, and the pages those queries return rarely match Google or Bing. 97% of the pages ChatGPT retrieved were outside Google's top 50, and 99% were outside Bing's top 30.

Gemini is grounded in Google Search, yet 93% of the pages it cited were also outside Google's top 50.

Retrieval explained 32% of the within-site variation in ChatGPT citations (95% confidence interval 24-42%), against 2.8% for all the page features the authors measured. When they shuffled the retrieval data between pages of the same site, the effect fell to 0.05%.

Content still matters here, but it works at the retrieval step. The authors say it acts mainly by matching the engine's queries in meaning.

ChatGPT also cites brands it already knows

The authors asked each model, with search switched off, which brands and websites it would recommend. When a brand was on that list, ChatGPT wrote it into its own search queries 67% of the time. When it was not, 11%. And 47% of ChatGPT's citations came from websites its model named without searching.

The model can name a brand with no page at all behind it. 46% of ChatGPT's mentions of a brand that the question did not name had no cited page naming that brand. We covered a related gap in why getting cited by AI search isn't the same as being recommended.

Small firms depend on other people's pages

The second dataset is closer to a small B2B firm's position. The models barely knew those vendors: the model behind ChatGPT named them for 2-16% of questions, and Gemini's model for 60-96%.

For these vendors, ChatGPT mostly named a brand when a cited third-party page named it. A third-party page naming the vendor multiplied the odds of a mention 39.7 times. In the large-brand dataset the multiplier was 4.2.

A separate controlled experiment shows how much a name can matter. The prompt told ChatGPT's model that brand names carried no information. One product profile still reached its top three in 80% of answers under one brand name and 39% under another.

What the study does not show

  • The results are associations. Pages differ in ways the authors could not measure.
  • Pages with more referring websites got 22.7% more citations per standard deviation, but the effect was mostly Gemini (+33.7%, against +3.7% for ChatGPT). The links were measured after the citations, so the direction is not established.
  • All questions were in English and about the US market.
  • The replay used the engines' APIs and their small models (GPT-5.4 nano and Gemini 3.1 Flash Lite), which need not behave like the consumer apps. The controlled experiments also used small models and constructed inputs.
  • The authors call the chain from brand knowledge to citation a hypothesis and say testing it needs intervention studies.

What I take from it

In my reading, not the authors', page wording that matches the question still matters. It is the one on-page factor that held up within a site. FAQ blocks and schema did not separate pages on the same site. I still build both into my own work, and I now treat them as baseline work that does not get a firm cited on its own.

For a small firm, the more useful check sits outside your site. Ask ChatGPT the questions your buyers ask and read which pages it cites. Count how many of those pages name you. If none do, that is the gap to close first.

A few common follow-up questions

Does this mean schema and FAQ blocks are useless?

No. The study found no advantage for them between pages on the same site. It did not test what happens when you remove them.

Does it apply to UK firms?

The authors only tested English questions about the US market, so the study cannot say. Run your own buyer questions through ChatGPT and see.

Has the study been peer reviewed?

Not that I could find. It is a preprint (arXiv 2610.15033). The authors say they release their data and code, so other researchers can check it.