Ahrefs: ChatGPT is moving from summarizing sites to official and primary sources
An Ahrefs analysis published on October 12, 2026 finds that ChatGPT answers most prompts without going to the web at all, and that, in Ahrefs' reading, it appears to be shifting away from sites that repackage information and toward the original sources behind it. In U.S. citation data, Reddit and Wikipedia lost much of their share within months, PubMed and the FDA moved up, and most site: searches went straight to brand websites.
66.6 million answers, 15 million background queries
Louise Linehan wrote the study for Ahrefs. The team looked at 66.6 million ChatGPT responses collected in August and September 2026, together with the 15 million fan-out queries behind them. A fan-out is one of the web searches the model fires off in the background before it answers. Every response in the set came from a free, logged-out GPT-5.6 Luna session. The first finding is how rarely the model searches at all: in Ahrefs' words, "ChatGPT answers 6 in 7 prompts without searching the web". Only 14.4% of responses triggered a search, and those that did produced roughly 1.6 fan-outs on average.
The second finding concerns where those searches point. Of all fan-outs, 6.8% used the site: operator to look inside a single website, and they were spread over about 116K domains. Among the roughly 160 domains searched most often this way, 75.9% of site: searches landed on a brand's own site. apple.com led with 9,442 searches, followed by ikea.com with 7,361 and nike.com with 5,210. Ahrefs reads this as ChatGPT appearing to favor original sources over the sites that summarize them, and Search Engine Journal's coverage describes ChatGPT's most-cited sites as tilting toward primary sources. This is Ahrefs' conclusion from its own data, not a statement from OpenAI.
Reddit and Wikipedia slipped, PubMed and the FDA climbed
The citation shift comes from a separate set of U.S. citation data that Ahrefs tracks. In March, around one ChatGPT answer in five cited Reddit; by late September that was closer to one in 43. Wikipedia went from one answer in 7 to one in 35 over the same stretch. Among the 200 domains ChatGPT cites most, PubMed jumped from 64th place to 4th, and the FDA appeared at 8th. Search Engine Journal adds that WebMD slid from 18th to 79th, while Britannica, Investopedia and the New York Post no longer appear anywhere in the top 200.
The kind of question matters too. In a separate Ahrefs test of 20,000 U.S. prompts, 84.4% of factual questions, such as those on prices, product specs or company information, drew at least one citation to an official source. For comparison questions that weigh two options against each other, the figure was 68.3%. The wording of the searches is changing as well: between August and September, the word "official" rose from 10.1% to 13.1% of English fan-outs, and "2026" from 17.6% to 22.9%. The rise suggests the model is looking for current, official information more often.
When ChatGPT searches, it increasingly looks you up on your own site
If ChatGPT checks a brand's own website directly with site: and favors official sources, then a business's own product, pricing, specification and company pages become some of its most valuable assets for generative engine visibility (GEO). Ahrefs makes the same point, noting that "The pages holding up best are the ones only you can publish" and pointing to original research, hands-on testing and first-party product facts. Pages that repackage someone else's information are the ones losing citations.
Two caveats apply. First, as Search Engine Journal stresses, Ahrefs' citation ranks measure share of answers, not the traffic a site receives. Second, the data reflects the free, logged-out behavior of a single model version, and Ahrefs itself warns that the next model update could change the picture again. We covered Reddit's citation slide on August 23; this study extends that trend with a larger dataset and new findings about brand sites.
The sources say nothing about Türkiye, so this section is commentary
Neither source breaks out any data for Türkiye or for Turkish-language queries. The fan-out sample spans all countries without a Türkiye breakdown, and the citation rankings are based on U.S. data. What follows is therefore not reporting but UNALSOFT's own reading.
The open questions, one by one: what share of site: searches in Turkish queries go to Turkish brand sites, whether Reddit and Wikipedia have lost citation share in Türkiye in the same way, whether words pointing to official sources are rising in Turkish fan-outs, and which Turkish institutional sites are gaining. Our view: if the model works the same way here, a business in Türkiye that keeps the most current, clearest and most complete version of its product and service information on its own site is preparing in the direction this data points.
The UNALSOFT view
We read this data not as a traffic promise but as a change in the role your website plays. If ChatGPT looks at a brand's own pages first when it researches that brand, the product features, service scope, terms and contact details on those pages need to be single, current and consistent. Scattered PDFs, outdated campaign pages or information that lives only on third-party platforms get lost in that lookup. That is why in our agentic AI work we treat the official information layer that AI assistants will read as its own piece of work. Since citation share is not the same as traffic, the outcome still needs to be tracked against your own data.
When ChatGPT searches your site, does it find the right answer?
A short conversation is enough to turn your product, service and company pages into one current source that AI assistants can read.