GEO experts Zhang Kai and Yao Jingang have published a research paper on arXiv proposing a two-stage measurement framework for GEO (Generative Engine Optimization) — "citation selection" and "citation absorption." They ran 602 controlled prompts across ChatGPT, Google AI Overview/Gemini, and Perplexity, collecting 21,143 citations and 23,745 feature records.
The core finding: citation quantity and citation depth diverge sharply. Perplexity averaged 16.35 sources per prompt, Google 12.06, and ChatGPT just 6.88. Yet ChatGPT scored the highest on "influence" — a metric measuring how deeply a cited page's content is absorbed into the generated answer. Simply counting citations doesn't capture how much of your content an AI search engine actually uses. This kind of measurement is exactly what Google DeepMind product lead Logan Kilpatrick recently advised companies to do — build their own benchmarks.
The paper also digs into what makes a page more likely to be deeply absorbed. Comparing high-influence and low-influence pages, the gap in word count was 11.44x, and the gap in heading count was 12.5x. Pages with code scored 76.88% higher influence than those without; pages with numbers or statistics scored 61.55% higher; pages with definition markers scored 57.33% higher; and pages with comparative content scored 55.28% higher. But Q&A format actually dragged influence down by 5.74% — evidence density matters more than surface formatting. By domain type, encyclopedia pages had the highest influence, while news pages, despite being frequently cited, had low influence.
The paper's data and code are open-source on GitHub.