Summary. Between March and September 2026, the sites ChatGPT cites most in US answers moved away from the ones that explain a subject and toward the ones that publish it first. That is the pattern in an Ahrefs study published on 12 October 2026, drawn from its Brand Radar data: about 2.4 million answers in March and 3.8 million in September. PubMed climbed from 64th to 4th. The FDA arrived at 8th. Government health, research, regulator and product-testing sites rose as a group; health and finance explainers fell. The two big community sources shrank sharply. Reddit's share of answers went from 19.1% to 2.3% and Wikipedia's from 14.1% to 2.9%, though both are still among the most-cited domains. Search Engine Journal reported the study the same day, and its reading supplies the steepest falls: WebMD from 18th to 79th, and Britannica, Investopedia and the New York Post out of the top 200. The measured change spans March to September 2026.
The obvious objection is that people simply asked ChatGPT different things in September. Ahrefs checked. On about 25,000 prompts that sat in both months' top 150,000, the movement ran the same way.
Credibility: HIGH. We read the Ahrefs post itself. It states its method, its sample sizes and the prompt-mix control. Ahrefs sells Brand Radar, the tracker that produced the data, so it has a commercial stake in citation studies. Ahrefs names three limits of its own: one model, US answers only, and citation rank is not traffic.
Facts
- PubMed rose from 64th to 4th among ChatGPT's most-cited domains between March and September 2026, and the FDA entered at 8th. WebMD fell from 18th to 79th, and Britannica, Investopedia and the New York Post dropped out of the top 200 (the WebMD and top-200 figures are Search Engine Journal's reading of the study).
- Reddit's share of US ChatGPT answers fell from 19.1% in March to 2.3% in September, and Wikipedia's from 14.1% to 2.9%. Both remain among the most-cited domains.
- To check that prompt mix was not driving the change, Ahrefs compared about 25,000 prompts present in both months' top 150,000 and found the same direction of movement.
- A separate fan-out analysis of 66.6 million free, logged-out GPT-5.6 Luna responses from August and September found 14.4% triggered a web search, searching responses averaged about 1.6 fan-out queries, and 6.8% of roughly 15 million fan-outs used the site: operator; of those aimed at the ~160 most-searched domains, 75.9% went to brand sites.
Sources
Ahrefs: How ChatGPT Does Query Fan-Out (15 Million Queries Studied), Ahrefs blog, 2026-10-12 — https://ahrefs.com/blog/query-fan-out-chatgpt/
Search Engine Journal: ChatGPT's Most-Cited Sites Tilt Toward Primary Sources, Data Shows, Matt G. Southern, 2026-10-12 — https://www.searchenginejournal.com/chatgpt-most-cited-sites-primary-sources-ahrefs/592682/
Stay ahead of AI search
The changes that move your rankings and AI citations — and the exact move to make — before they cost you. Free, three times a week.
For marketing leaders, a citation report built on a March baseline now describes a source mix ChatGPT has since moved away from. The prompt control is what makes this worth acting on. Without it, the story would be that people asked different questions. With it, the change sits in where the answers come from. The pattern is simple to state. PubMed and the FDA publish the research and the rulings; WebMD and Investopedia explain them. In this data, the publisher gained and the explainer lost. A page on your site that restates what a regulator or a study already says is competing with the original, and in September the original won more often. A page that carries your own test results, specifications or documentation is the original.
The fan-out analysis points the same way from a separate dataset. When a free, logged-out GPT-5.6 Luna response used the site: operator on one of the roughly 160 most-searched domains, 75.9% of those searches went to a brand's own site. That is a small slice, 6.8% of fan-outs, so read it as a hint about where ChatGPT looks when it wants an answer from the source, nothing larger.
Two cautions keep all of this a direction. It is one model in one market, and a high citation rank does not guarantee a visit. And Reddit and Wikipedia lost share without leaving the most-cited list, so dropping them from a plan on this evidence would overcorrect. Keep the AI-visibility tracker in the reporting pack, start it from a fresh baseline, and put the next round of content hours into the pages only you can write.
Actions this week
1. Ask your SEO team to re-run your core prompt set in ChatGPT now and compare the cited domains against your March to September tracker history before anyone reports a trend line.
2. List your ten highest-traffic informational pages and mark which carry original data, testing or official documentation and which only summarise other sources. Ask the content team to add a sourced original figure to each page in the second group.
3. If Reddit or Wikipedia was a named line in your citation plan, write down whether you keep it at its current weight, and why.
Ahrefs: How ChatGPT Does Query Fan-Out (15 Million Queries Studied), Ahrefs blog, 2026-10-12 — https://ahrefs.com/blog/query-fan-out-chatgpt/
Search Engine Journal: ChatGPT's Most-Cited Sites Tilt Toward Primary Sources, Data Shows, Matt G. Southern, 2026-10-12 — https://www.searchenginejournal.com/chatgpt-most-cited-sites-primary-sources-ahrefs/592682/