ChatGPT's most-cited sites shift toward original sources

Ahrefs published How ChatGPT Does Query Fan-Out (15 Million Queries Studied) on 12 October 2026. Search Engine Journal’s Matt G. Southern wrote up its citation findings the same day in ChatGPT’s Most-Cited Sites Tilt Toward Primary Sources, Data Shows. We read both. This is Ahrefs’ own data, published by a vendor that sells the tracker behind it.

What did Ahrefs find?#

Ahrefs compared the 200 domains ChatGPT cited most in US answers in March and September 2026. The chart covers 2.4 million answers in March and 3.8 million in September, drawn from Ahrefs’ Brand Radar database of ChatGPT citations.

In that chart PubMed climbed from 64th to 4th and the FDA entered at 8th, while health sites that summarize research, WebMD and Verywell Health among them, fell. Ahrefs writes that ChatGPT “seems to be prioritizing original sources over sites that summarize them.”

Who rose and who dropped out?#

Ahrefs sorts the movers into groups. The citations Reddit, Wikipedia and YouTube lost, it says, went to research bodies, regulators, brands and hands-on testers, while explainer and aggregator sites fell.

• Government health and research: PubMed, FDA, CDC, MedlinePlus, DailyMed, NIH, NHS.
• Regulators: SEC, EPA, FTC, CFPB, NHTSA, CPSC.
• Hands-on testers: RTINGS, CarComplaints, RepairPal.
• Health and finance explainers: WebMD, Healthline, Verywell Health, Investopedia, Livestrong.

Southern adds the specifics: WebMD went from 18th to 79th, and Britannica, Investopedia and the New York Post fell out of the top 200. He reports Reddit cited in 19.1% of US answers in March and 2.3% in September, and Wikipedia going from 14.1% to 2.9%. Ahrefs says both are still among ChatGPT’s most-cited domains.

Is it only a change in prompts?#

Ahrefs notes that Brand Radar adds prompts over time, so a rank can move because the questions changed. To check, it compared about 25,000 prompts that were among its 150,000 highest-volume prompts in both March and September. It says every site in the chart moved in the same direction on those identical prompts.

Ahrefs also says the prompts in any set affect which sites show up. The post does not publish the prompt list, and it does not say how a domain’s rank is calculated beyond how often it is cited.

How often does ChatGPT search?#

The same post analyzed 66.6 million free, logged-out GPT-5.6 Luna responses from August and September, across all countries. Only 14.4% involved a web search, and searching responses ran about 1.6 fan-out queries on average, according to Ahrefs.

Ahrefs calls these figures a snapshot of one model, and says logged-out free sessions search less than paid or reasoning models. It says other studies found far more fan-outs per prompt, but those used different models, prompts and collection methods, so the numbers are not directly comparable.

What can publishers take from it?#

Ahrefs’ reading is that if content mainly summarizes what someone else publishes first, ChatGPT may cite the original source instead. It says the pages holding up best are ones only you can publish: original research, hands-on testing and your own product facts.

The practical step is to cite and link the original source in your own copy, and to put your effort into material nobody else holds. Southern points out the ranks do not measure traffic, so the check is how often ChatGPT cites your own pages for your core topics, and which sources it names instead. For related data, see how OpenAI licensing deals relate to ChatGPT citations and our AI search coverage.