Study: ChatGPT's own search explains citations better than page content
University of Toronto researchers tracked 90 days of ChatGPT, Gemini and Perplexity answers to commercial questions — 158,099 answers, 468,037 citations — and found that page content explains only 2.8% of which pages ChatGPT cites within a given website, while whether ChatGPT’s own rewritten search queries retrieved the page explains 32%, a gap the study calls statistically decisive (95% CI 24–42%). Content still matters, the authors found, but mainly through matching the wording of those rewritten queries rather than through traditional signals like FAQ blocks, structured data or page speed, which showed effects only when comparing across different websites, not between pages of the same site.
A separate set of controlled experiments found the models’ own prior knowledge also skews results: with every version of a product holding identical facts, revealing its brand name changed ChatGPT’s top-three ranking rate 2.1 times as much as merely repeating the prompt, and 3.3 times as much for Gemini.
Why it matters: It supplies the mechanism behind a finding the wire has already logged — that content signals barely predict citations once domain is controlled for — by showing the missing variable is whether the engine's own search retrieves the page at all, which argues for optimizing to match an engine's rewritten queries rather than traditional on-page signals.
The record: ChatGPTGeminiPerplexity
Via arXiv ↗
Posted to the wire October 13, 2026. Edited by Joe Balewski.