OctoLink GEO

Page Properties Explain 2.8 Percent of ChatGPT Citations

Author Editor

A Toronto study of 158,099 answers says ChatGPT retrieval explains 32 percent of citation variation, against 2.8 percent for page properties.

GEO Citations Retrieval

Direct answer

Chen and Koudas report that, within a website, page properties explain 2.8 percent of ChatGPT citation variation, while the pages its own search retrieves explain 32 percent.

Mahe Chen and Nick Koudas, both at the University of Toronto, study what shapes citations in ChatGPT, Gemini, and Perplexity. Their HTML preprint reports two 90-day collections of answers to commercial questions: 158,099 answers and 468,037 citations. The larger set is US home internet, with 651 questions, 143,249 answers, and 396,188 citations, collected daily from 24 June to 21 September 2026. The engines named in the methods are ChatGPT with GPT-5.4 nano and web search, Gemini 3.1 Flash Lite with Google Search grounding, and Perplexity Sonar.

Comparing pages of the same website, the abstract says content matters in one way: pages whose title and paragraphs match the questions people ask are cited more. FAQ blocks, structured data, summary boxes, and page speed appear effective only across websites. Page properties explain 2.8 percent of the citation variation between pages of a website for ChatGPT. Replaying the questions through the engines' APIs shows that every page ChatGPT cites is one its own search retrieved, and that retrieval explains 32 percent of the same variation.

What the engine table shows

Table 1 covers the 651 queries. ChatGPT has 58,590 answers, 116,530 citations, 2.0 pages per answer, and no citation on 30.8 percent of answers. Gemini has 26,069 answers, 263,300 citations, 10.1 pages per answer, and no citation on 15.6 percent. Perplexity has 58,590 answers, 16,358 citations, 0.3 pages per answer, and no citation on 95.8 percent. Citations to brand websites are 64.3 percent for ChatGPT, 16.8 percent for Gemini, and 6.8 percent for Perplexity.

The authors say the engines rewrite each question into new sub-queries on every answer and cite mostly from the top of their own ranking, which is not the ranking of Google or Bing. They also say 63.2 percent of 39,059 cited URLs returned usable HTML. The outcome they model is the log of one plus the citation count, on 12,038 pages cited at least twice.

FAQ

Which on-page practices survived the within-website tests?
The paper says pages whose title and paragraphs match the questions people ask are cited more. FAQ blocks, structured data, summary boxes, and page speed looked effective only across websites.
How many pages does each engine cite?
In the main table, ChatGPT cites 2.0 pages per answer, Gemini cites 10.1, and Perplexity leaves 95.8 percent of answers without a citation.