Have you ever wondered how ChatGPT sources its information? According to a recent study, it turns out ChatGPT tends to pull pages significantly more often than it actually cites them, concentrating its citations primarily among roughly 30 key domains.
Interestingly, AI citations in ChatGPT are far more concentrated compared to traditional search engines. Approximately 67% of citations within a specific topic are shared among only 30 domains.
From Kevin Indig’s latest study, I learned that comprehensive topical coverage through long-form pages and cluster-based models tend to perform better than the older “one keyword, one page” strategy.
The details. The visibility of citations isn’t spread out evenly. For product comparison topics, the top 10 domains capture about 46% of the citations, while the top 30 account for 67%.
Though AI visibility is slightly less concentrated than classic organic search, the competition for citations is still incredibly centralized. As Indig concludes, without building sufficient authority, you’re likely excluded from these valuable citation “seats.”
What changed. While ranking first on Google remains important, it’s not the sole factor anymore. Interestingly, of the pages ranked No. 1, only 43.2% were cited by ChatGPT, which is 3.5 times more than pages ranking beyond the top 20.
ChatGPT retrieves a vast number of pages but cites only a fraction of them. According to AirOps, about 85% of the pages retrieved were never cited, and a significant portion of citations arose from fan-out queries that lacked search volume entirely.
Why we care. Simply publishing the best answer for a keyword isn’t sufficient anymore. ChatGPT favors domains that offer comprehensive coverage of a topic, giving preference to pages that approach subjects from multiple angles.
The patterns. It turns out, longer pages typically receive more citations, and this trend varies by industry. Notably, pages with 5,000 to 10,000 characters see the most substantial lift. For pages over 20,000 characters, the average number of citations hits 10.18 compared to a mere 2.39 for shorter pages.
However, this pattern is not universal. For instance, in finance, shorter, densely packed pages often shine over lengthy guides. But in Education, Crypto, and Product Analytics, longer pages still hold their ground with little drop-off.
Looking more closely at on-page behavior, ChatGPT tends to cite from the top sections of a page. Particularly, the 10% to 20% section excels across all industries, while the bottom 10% of the page barely garners recognition.
About the data. For this study, Indig analyzed approximately 98,000 citation rows from about 1.2 million ChatGPT responses, using a variety of analytical methods to pinpoint which pages earned citations and their origins.
The study. More about these findings can be found in the study titled The science of how AI picks its sources.
Inspired by this post on Search Engine Land.
Leave a Reply