PubMed climbed from 64th to fourth place among the domains ChatGPT cites most often in U.S. answers between March and September 2026. WebMD, over the same stretch, fell from 18th to 79th. That’s not a small ranking shuffle. It’s Ahrefs, which tracked 200 of the most-cited domains across 2.4 million answers in March and 3.8 million in September, describing a pattern plainly: ChatGPT “seems to be prioritizing original sources over sites that summarize them.” In other words, primary sources are gaining ground on the explainers that summarize them.
Table of Contents
The Pattern Behind the Numbers
The sites that climbed in Ahrefs’ data are primary sources: they’re where information originates, not where it gets repackaged. Government health and research agencies, regulators, and hands-on product testers made up most of the gainers. The sites that fell share the opposite trait. Ahrefs specifically categorizes WebMD and Verywell Health as sites that summarize research rather than produce it, and groups Healthline and Investopedia into the same “health and finance explainer” bucket, all of which lost ground. Britannica, Investopedia, and the New York Post dropped out of the top 200 entirely.
This isn’t a story about those specific brands losing favor for arbitrary reasons. It’s a story about what kind of content earns a citation when an AI system is synthesizing an answer. A page that explains, in its own words, research that exists somewhere else is one layer removed from the primary sources ChatGPT actually needs to point to. A page that IS the research, the regulatory filing, the hands-on test, has nothing between it and the claim being made.
This Isn’t Just Noise From a Changing Question Mix
A reasonable objection to any citation-tracking study is that the questions people ask change over time, so a shift in which sites get cited might just reflect a shift in what’s being asked about, not any real change in how ChatGPT picks sources. Ahrefs checked for exactly that by isolating roughly 25,000 prompts that appeared among its highest-volume prompt set in both March and September, the same questions, asked both times. Every site on their chart moved in the same direction for that controlled subset as it did in the full dataset. That rules out the most obvious alternative explanation and makes the underlying pattern harder to dismiss as a side effect of trending topics.
Reddit and Wikipedia Are Down, But Not Out
Two of the internet’s biggest aggregators of other people’s content, Reddit and Wikipedia, both saw sharp drops in how often they’re cited directly in ChatGPT’s U.S. answers: Reddit went from being cited in 19.1% of answers in March to just 2.3% in September, and Wikipedia fell from 14.1% to 2.9% over the same window. Ahrefs is careful to note both sites remain among ChatGPT’s most-cited domains overall; a drop from 19% to 2% of answers is still a meaningful presence, just a much smaller one than before.
Separately, in a worldwide sample looking at which domains ChatGPT searches most often using site-specific queries, Reddit still ranked ninth, a reminder that “cited in the final answer” and “searched as part of producing that answer” are two different measurements that don’t always move together.
Most ChatGPT Answers Still Don’t Search the Web at All
It’s worth sitting with a figure that’s easy to miss in a story about citation rankings: across 66.6 million free, logged-out responses analyzed from August and September, only 14.4% involved ChatGPT running a web search at all. The rest were answered from the model’s own training, with no live citation of any external source, which means the entire citation-ranking conversation above applies to a minority of these responses, not the majority of them.
That sample covers only free, logged-out use, which the report says triggers fewer searches than paid or reasoning models, so it shouldn’t be read as a figure for every ChatGPT user. When a search does happen, it typically triggers a small number of background queries (around 1.6 on average) that ChatGPT runs before composing its answer, and OpenAI’s own documentation describes this as something the model decides to do on its own when a question would benefit from current information, not something that happens on every query by default.
What Actually Gets Searched, When a Search Happens
Of the roughly 15 million individual background searches in Ahrefs’ sample, 6.8% were site-specific queries limited to one domain, spread across 116,000 different domains. Among the roughly 160 domains targeted most often by those site-specific searches, 75.9% were brand sites, meaning company websites like a retailer’s own domain or a manufacturer’s own product pages, not news outlets or review sites. That’s a useful detail for any business wondering whether this entire conversation is relevant to them: if a prospective customer asks ChatGPT something that triggers a site-specific search of an actual company’s own domain, the company’s own site, not a third-party review of it, is disproportionately likely to be the thing getting searched directly.
What Primary Sources Mean for a Small Business’s Content
Ahrefs’ suggestion, as Search Engine Journal reports it, is to focus on pages that only you can create: original research, hands-on testing, and your own product information. For a web hosting company, an e-commerce store, or any small business publishing content, that’s a specific, actionable filter to run every piece of planned content through before writing it. A generic explainer about a topic that twenty other sites have already explained, in roughly the same way, is exactly the content category Ahrefs’ data shows losing ground.
A page built around something only your business actually has, your own uptime numbers, your own pricing and plan specifics, a real test you ran, a specific problem you solved for an actual customer, is the category gaining it.
This doesn’t mean every post needs to be original research in a formal sense. It means the parts of a post that cite genuinely unique information, your own data, your own documented experience, your own specific product details, are doing more work toward AI-answer-engine visibility than the parts that restate general industry knowledge already available everywhere else. One practical habit: before publishing, ask whether a page could be swapped out for a dozen competitors’ versions of the same topic without anyone noticing the difference. If yes, the page needs something concrete and specific that only your business can provide. If it already has that, it’s positioned the way Ahrefs’ data suggests actually matters right now.
A Quick Self-Audit for Your Own Content
Translating that guidance to focus on pages only you can create into something concrete doesn’t require new tools, just an honest pass through your existing content with a specific question in mind for each page: if a competitor copied this page’s structure and wrote their own version of it, would the result actually be different in any way that matters? A few prompts that help answer that:
- Does this page state a number, a result, or a fact that comes from your own operation (your own uptime data, your own support response times, your own documented test of a product or process), rather than restating a figure that’s already published somewhere else?
- Could this page’s core claim be verified by a reader going directly to your own evidence, a screenshot, a dataset, a dated record, rather than only by trusting your summary of someone else’s research?
- Does the page explain a specific situation you’ve actually handled (a particular migration, a particular troubleshooting case, a particular customer scenario), rather than describing the topic in general terms that apply to any business in the category?
A page that answers “no” to all three isn’t necessarily bad content, but it’s exactly the kind of page Ahrefs’ data suggests is losing ground as a citation source, and it’s worth asking whether a planned rewrite could add even one of those three elements before publishing.
A Caveat Worth Keeping in Mind
Ahrefs itself flags an important limit on this data: citation rank isn’t the same as traffic. Being cited more often by ChatGPT doesn’t automatically translate into more visitors, and the company’s own fan-out sample comes from a snapshot of a single model over two months, not a permanent, fixed ranking system. Models get updated, citation behavior can shift again, and a different AI system might weight sources differently than ChatGPT currently does.
The sensible response to data like this isn’t to treat September 2026’s snapshot as a permanent rulebook, but to periodically check which primary sources and explainers a given AI platform cites for topics relevant to your own business, since the report’s authors note these patterns can and do change with model updates.
Key Takeaways on Primary Sources
- ChatGPT’s most-cited domains in U.S. answers shifted toward primary sources between March and September 2026, according to Ahrefs data reported by Search Engine Journal.
- PubMed rose from 64th to fourth, while WebMD fell from 18th to 79th.
- Ahrefs checked a fixed set of about 25,000 prompts and found every site on its chart moved the same way, so the shift isn’t just a change in the questions people ask.
- Only 14.4% of the free, logged-out responses in Ahrefs’ sample involved a web search, so citation rankings describe a minority of those responses.
- For a small business, the useful response is to publish pages only you can create, such as your own data, tests and product details: the kind of primary sources AI systems cite.
- Citation rank is not traffic, and the pattern may change when OpenAI updates its model.
Related reading: our post on why ChatGPT now cites Reddit far less covers the Reddit side of this data.