PubMed jumped from 64th to fourth in Ahrefs’ ranking of ChatGPT’s U.S. citations between March and September, while WebMD slid from 18th to 79th. Search Engine Journal reported the findings on October 12. The pattern suggests ChatGPT is rewarding the origin of information over the sites that repackage it.
Ahrefs is an SEO vendor with its own tracking product, so the figures describe its sample, not ChatGPT as a whole. The ranking covers the 200 domains cited most often. March rests on 2.4 million U.S. answers, and September on 3.8 million.
Climbers on the chart were mostly public-sector health and research sites, regulators, and product reviewers who test items by hand. Health and finance explainers moved the other way. Britannica, the New York Post, and Investopedia all left the top 200. Ahrefs wrote that ChatGPT “seems to be prioritizing original sources over sites that summarize them.” That is the vendor’s interpretation of a ranking, not a statement from OpenAI.
Ahrefs puts WebMD and Verywell Health in a group of health sites that digest research, and Healthline and Investopedia join them in a broader explainer category. The wider list of gainers also includes the CFPB and FTC, plus help pages published by Google and Microsoft.
The methodology carries a built-in risk. Ahrefs adds prompts to its tracking over time, so a rank shift could stem from fresh questions in the pool instead of any behavior change by ChatGPT. To test that, the company isolated roughly 25,000 prompts present in its top 150,000 by volume in both months. Each domain on the chart moved the same way on that shared set, which supports the trend but covers a small slice of the full sample.
Reddit and Wikipedia posted the steepest falls in share. Reddit appeared in 19.1 percent of U.S. answers in March, then 2.3 percent in September. Wikipedia fell from 14.1 percent to 2.9 percent. Ahrefs says both remain among the most-cited domains.
Search Engine Journal’s Matt G. Southern noted that this differs from Promptwatch data he covered in August, which tracked how Reddit’s slice of ChatGPT Search citations dropped from 3.83 percent on average to 0.52 percent. Ahrefs counts the share of answers citing a domain, a different denominator. Two vendors using two methods still point the same direction for Reddit.
The report also measured how often ChatGPT searches at all. In a sample of 66.6 million free, logged-out replies spanning August and September, 14.4 percent ran a web search. Those searching replies averaged about 1.6 fan-out queries, the background lookups ChatGPT performs before answering. Among 15 million fan-outs, 6.8 percent were site: searches. Of the site: searches aimed at the top 160 or so domains, 75.9 percent went to brand sites such as apple.com and ikea.com.
Fan-out figures from other studies run higher. Nectiv measured 7.61 per prompt across about 4,000 prompts sent to GPT-5.6 Sol through the API. The samples differ in model, prompts, and collection method, so the numbers do not compare directly. Ahrefs adds that free, logged-out use triggers fewer searches than paid or reasoning models, and that its fan-out data is a single-model snapshot that could change.
For search teams, the stake is where an explainer page sits relative to its sources. A page that summarizes a regulator or a study now competes with that source for the citation. Ahrefs recommends investing in pages only you can produce: first-party studies, product testing you run yourself, and details about what you sell.
Ahrefs’ citation ranks do not measure traffic, so none of this proves lost visits. Audit how often ChatGPT cites your pages for the topics you care about and which outside sources it names instead, then repeat the audit after each OpenAI model update.
Search Engine Journal, Matt G. Southern, October 12, 2026, reporting Ahrefs data.
