ChatGPT Citations, Part 1: What Happens Between Your Question and the Sources It Shows
We traced ChatGPT Web across four model settings: how it rewrites a question, how many searches it runs, and why being retrieved is not being cited.
Read →本站有中文版。
切换到中文The methods Sourso uses to check engines and rewrite pages come from research we run and publish. Each paper is written up here one finding at a time, in the tone of the paper, with the numbers as reported and the limitations stated.
We traced ChatGPT Web across four model settings: how it rewrites a question, how many searches it runs, and why being retrieved is not being cited.
Read →Candidate pages churn from run to run, yet the ones that keep coming back are the ones ChatGPT cites. Recurrence and snippet length carry a measurable signal.
Read →Most tools measure ChatGPT through Google results or the API. We test how much of what ChatGPT Web actually cites each proxy recovers, by page, site and entity.
Read →With only Google and API signals, a model predicts which pages, hostnames and entities ChatGPT Web cites at AUC above 0.74. The ceiling is candidate coverage.
Read →Generative engines answer election questions by citing web pages. We investigate whether GEO could shift what voters see, starting with what engines cite.
Read →Our first VoteGEO experiment: whether rewriting an election page raises a party's exposure in AI answers only when the rewrite is told which party to favor.
Read →The second VoteGEO experiment compares eleven rewriting strategies: reorganizing a page or arguing a position moves exposure more than evidence or style.
Read →The third VoteGEO experiment reruns the rewrites under 59 voter personas: a stated partisan identity shifts baseline exposure more than one rewritten page.
Read →