Ahrefs, an SEO tools provider, has published changes in the rankings of domains cited in ChatGPT responses for the U.S. Comparing March and September, the medical literature database PubMed rose from 64th to 4th, while health information site WebMD fell from 18th to 79th. Sites with source material, such as research data and hands-on product tests, appear to have moved up the rankings.
Research databases surge as explainers fall back
The analysis covered the top 200 most frequently cited domains in ChatGPT responses for the U.S. It was based on about 2.4 million responses in March and about 3.8 million in September.
The domains that rose in the rankings were mainly PubMed, government health and research agencies, regulators, and testing sites that try and evaluate products firsthand. The FDA and RTINGS are cited as examples.
By contrast, WebMD and Verywell Health, which break down research for readers, as well as health and finance explainer sites such as Healthline and Investopedia, fell in the rankings. Britannica, Investopedia, and the New York Post dropped out of the top 200.
Ahrefs says ChatGPT appears to favor original sources over sites that summarize information.
An era when the “owners of information” are favored over “summarizers”?
Health and finance explainers have traditionally been strong performers in search engines. They have interpreted research papers and public documents, translating them into plain language. But when AI generates an answer, it takes on that “translation” work itself. If so, it seems natural that AI would want to refer to the original source rather than the person who translated it.
Another notable trend is the rise of product-testing sites. Unlike articles that simply rearrange specification tables, measurements taken firsthand and real-world impressions are unique to those sites. The criteria for being cited by AI may be coming down to whether a site has original content.
Ahrefs also recommends that site operators focus on pages that only they can create. Examples include original research, hands-on product testing, and information about their own products.
Citation shares for Reddit and Wikipedia plunge
The share of U.S. responses citing Reddit fell from 19.1% in March to 2.3% in September. Wikipedia also declined, from 14.1% to 2.9%. Even so, Ahrefs says both remain among the domains ChatGPT cites frequently.
Another analytics tool, Promptwatch, reported a similar decline for Reddit in August. It found that reddit.com’s share of ChatGPT Search citations fell from an average of 3.83% in late July and early August to 0.52% in mid-August. However, Ahrefs measures the “share of responses that cited Reddit,” so its methodology differs from Promptwatch’s.
In a separate global sample, reddit.com ranked ninth among the domains ChatGPT most often specified in site: searches—searches limited to a particular site. This suggests it still has a presence in search even after its citation share fell.
Citation rankings and traffic are different things
One important point is that these rankings reflect citation frequency, not the number of visitors to a site. A lower ranking does not mean that the site’s traffic fell by the same amount. Conversely, appearing near the top does not necessarily mean people are clicking through and reading the site.
Also, Ahrefs’ tracking system adds questions over time, so ranking changes may result from changes in the questions being asked. To account for this, the company compared about 25,000 questions that ranked among the top 150,000 high-volume questions in both March and September. It said that every site shown on the chart moved in the same direction.
The reality is that most answers involve no search
Another study looked at how often ChatGPT searches the web. Of 66.6 million responses to free, logged-out users in August and September, 14.4% involved a web search. The model covered was GPT-5.6 Luna, and the data came from all regions.
Among responses that involved a search, an average of about 1.6 searches ran in the background before the answer was generated. These searches are called fan-outs. Of 15 million fan-outs, 6.8% were site: searches limited to particular sites, covering 116,000 domains. Among the roughly 160 most frequently searched domains, 75.9% of searches were directed at official company websites such as Apple and IKEA.
Another research firm, Nectiv, processed about 4,000 questions through GPT-5.6 Sol via API and found an average of 7.61 fan-outs per question. However, the figures cannot be compared directly with 1.6 because the models, questions, and collection methods differ. Ahrefs also says that free, logged-out users search less often than users of paid and reasoning models, which is one factor behind the difference.
What site operators should verify
Ahrefs’ findings are a snapshot of a single model collected in August and September. The company also notes that the trend could change with future model updates.
For site operators, a practical step is to regularly check which pages ChatGPT cites on their core topics—and which sources it chooses instead. If explanatory articles are a primary source of traffic, it is worth revisiting that review whenever OpenAI updates its models.
Comments