PubMed went from the 64th most cited domain in US ChatGPT answers to the 4th. WebMD went from 18th to 79th. That is six months of data, and it describes a business model losing its position.
A recent piece on Search Engine Journal walked through new Ahrefs data on which domains ChatGPT cites. The conclusion Ahrefs drew is blunt: ChatGPT appears to prioritize original sources over sites that summarize them. Here is what I think that means for anyone who publishes content for a living.
What the Ahrefs numbers actually show
Ahrefs analyzed the 200 most cited domains across 2.4 million US ChatGPT answers in March and 3.8 million in September. To rule out the obvious objection, that people simply asked different questions, it compared about 25,000 prompts that appeared in both months. Every site on the chart moved in the same direction.
The losers share a profile. Britannica, Investopedia and the New York Post dropped out of the top 200 entirely. Healthline, Verywell Health, WebMD and Investopedia were grouped as health and finance explainers, sites whose core product is a readable version of research done elsewhere.
The community sites fell too. Reddit appeared in 19.1 percent of US ChatGPT answers in March and 2.3 percent in September. Wikipedia fell from 14.1 percent to 2.9 percent. Both are still on the list, but the era of "just get mentioned on Reddit" as an AI visibility strategy looks a lot shorter than the people selling it predicted.
The middleman layer is the one being cut
For twenty years, the most reliable content business online was translation. Take a dense primary source, a clinical study, a tax regulation, a product spec, and rewrite it so a normal person can read it. Rank for the question. Monetize the visit.
A language model does that translation itself, on demand, for free. When it can read the primary source directly, the summarizer adds nothing the model cannot produce in the answer. So the model goes upstream, and the summarizer loses the citation that used to justify its existence.
I wrote earlier this year that your website is the source, not the destination. The Ahrefs data is the clearest evidence yet. Being the destination for a summary is a shrinking position. Being the place where a fact first appears is a growing one.
Two numbers most people will skip
The second half of the study matters as much as the rankings. Of 66.6 million free, logged-out responses from August and September, only 14.4 percent involved a web search at all. When a response did search, it averaged about 1.6 fan-out queries.
Read that carefully. Most answers come from what the model already knows. Citations only happen in a minority of conversations, and in those, the model runs very few searches. Competition for those slots is brutal, and it is getting more concentrated, not less.
The other number is about brands. Of 15 million fan-out queries, 6.8 percent were site: searches spread across 116,000 domains. For the roughly 160 most searched domains, 75.9 percent of those site: searches went to brand websites such as apple.com and ikea.com.
When the model wants a fact about a company, it increasingly asks the company. That is good news if your site actually contains the fact, and bad news if your pricing, specs and policies live in a PDF nobody can parse.
What counts as a primary source for a business
Most companies hear "primary source" and think of academic research. That is too narrow. For a business, a primary source is anything that exists nowhere else until you publish it.
Your own pricing, stated plainly on a page a machine can read. Your product specifications and compatibility details.
Data from your own customer base, aggregated and anonymized. Test results from running your product against a real problem. A clear explanation of your process, with the numbers attached. Ahrefs itself points to original research, hands-on testing and proprietary product information as the content hardest for others to replicate.
Notice what is not on that list. The "ultimate guide" that restates what ten other sites already say. The blog post explaining a Google update that Google already explained.
The comparison table built from competitors' marketing pages. That content was always weak. Now it is also invisible to the systems where buying research increasingly starts.
How I would rebuild a content plan this quarter
Start with an audit, not a brainstorm. List your top 50 pages by traffic and mark each one honestly: does this page contain a fact that originated with us, or is it a rewrite? In most companies I see, the ratio is worse than 1 to 4.
Then shift the budget. Fewer explainers, more originals. One proprietary data point published every month beats twenty generic articles, both for AI citations and for the humans who still read. Put the facts in plain HTML, with dates, units and sources, so a model fetching the page can quote it without guessing.
Finally, measure the right thing. Traffic from AI answers is a lagging indicator. The leading indicator is whether your brand appears in the answer and which source the model used to put it there. That is exactly the question GEOflux.ai (geoflux.ai) is built to answer: not only whether you are mentioned, but which pages and which third parties are shaping the mention.
The summarizers had a good twenty years. The next ten belong to whoever owns the original.