A Citation Lab Finds ChatGPT Cites Fewer Pages and Uses Them More
The geo-citation-lab snapshot has 602 prompts. ChatGPT averages 6.88 citations and 0.2713 influence; Perplexity averages 16.35 and 0.0646.
Direct answer
The paper on citation selection and citation absorption says the geo-citation-lab snapshot has 602 prompts and 21,143 valid search-layer citations. ChatGPT averages fewer citations per prompt than Perplexity, and a higher mean influence on fetched pages.
The paper "From Citation Selection to Citation Absorption" splits generative-engine visibility into two measurements. Citation selection is the platform triggering search and choosing sources. Citation absorption is a cited page contributing language, evidence, structure, or factual support to the answer. The public geo-citation-lab snapshot it uses has 602 controlled prompts across ChatGPT, Google AI Overview/Gemini, and Perplexity, 21,143 valid search-layer citations, 23,745 citation-level feature records, 18,151 successfully fetched pages, and 72 extracted features. A later section says the fetch success rate is 76.44%, and that the cleaned search-layer rows number 21,181.
The reported means separate breadth from depth. Mean citations per prompt are ChatGPT 6.88, Google 12.06, and Perplexity 16.35. Mean fetched-page influence is ChatGPT 0.2713, Google 0.0584, and Perplexity 0.0646. A table of triggered prompts says Google ran on 602 prompts with a 99.67% trigger rate, a median of 12 citations, and a maximum of 37. Perplexity triggered on all 602, with a median of 17 and a maximum of 27. ChatGPT's search-layer summary uses 587 observed prompts in the cleaned report.
Formatting is not the same as absorption
Q&A pages show mean influence 0.0947 versus 0.1005 for non-Q&A pages, a relative difference of -5.74%. News is selected often, while news_media pages average 0.0726 influence and encyclopedia pages average 0.2144 in the domain-type table the paper cites. On multi-constraint tasks, ChatGPT averages 3.4 citations, Google 12.6, and Perplexity 17.7. The authors call the compression reading an interpretation, not a causal claim. Official, news, and vertical sources account for 79.12% to 87.52% of citations across the platforms in that selection table.
Source: arXiv 2604.25707.
FAQ
- Does Q&A formatting raise influence?
- The paper says Q&A pages show mean influence 0.0947 versus 0.1005 for non-Q&A pages, a -5.74% relative difference.
- How many pages were fetched?
- The abstract says 18,151 pages were fetched successfully, and a later section reports a 76.44% fetch success rate.