ChatGPT is quietly swapping its sources. Primary wins, aggregators lose.
Over the last two years, the AEO shortcut was relatively straightforward. Get a couple of mentions on Reddit, find your way to the right Wikipedia page, or receive a large review on YouTube, and ChatGPT would cite you. Recently, that shortcut has become less reliable, and faster than most people noticed. New data from Ahrefs shows that sources that ChatGPT trusts are shifting, and the aggregators are losing blood.
Let’s look at Reddit first. In March of this year, ChatGPT cited Reddit about once in every 5 answers. In September, that number dropped to once in every 43 answers. Similar trends can be seen with YouTube and Wikipedia. YouTube went from once in every 370 answers, to once in every 500. Wikipedia went from once in every 7 answers, to once in every 35 answers. The sources are big, but the number of times ChatGPT cites them is decreasing.
Source: Ahrefs data (https://ahrefs.com/blog/query-fan-out-chatgpt/), coverage: Search Engine Journal (https://www.searchenginejournal.com/chatgpt-most-cited-sites-primary-sources-ahrefs/592682/)
The citations are shifting to the primary sources, including research institutions and regulatory bodies. A clear example of that shift is with PubMed. It went from the 64th to the 4th most cited domain among ChatGPT’s 200 most-cited domains. The FDA appeared at the 8th spot. Sites that summarize research like WebMD and Verywell dropped, along with many others. The sample size in this case isn’t small. The data runs across 2.4 million answers in March and 3.8 million answers in September.
This is the uncomfortable part. Nothing on Reddit and Wikipedia got worse. Your content didn’t change. The model’s tastes did. I’ve written about this before. The model layer moves under you (https://sandeepamar.in/when-the-underlying-chatgpt-model-updates-aeo-citation-patterns-shift-even-if-your-content-has-not-changed/), and your citations can swing without you touching a word. This is that playing out across the whole citation graph.
Microsoft already said the quiet part
Pay attention to what the platforms say, not what they advise. In a post about the evolving role of the index (https://blogs.bing.com/search/2026/5/Evolving-role-of-the-index-From-ranking-pages-to-supporting-answers/), Microsoft points out the unit of value shifts from documents to “groundable information, discrete, supportable facts with clear provenance.” Read that twice. Provenance. A system optimizing for provenance will step beyond the aggregator and reach for the thing the aggregator was summarizing in the first place. PubMed over WebMD is exactly that in the wild.
Google frames it as quality. Same idea. Its guide to generative AI features (https://developers.google.com/search/docs/fundamentals/ai-optimization-guide) keeps hammering home the importance of first-hand, expert-led content, and serves as a reminder to all these features run through Google’s core ranking systems. First-hand is the focal point. A recap of someone else’s study is not first-hand.
What actually wins now
Quality and trust cues dominate recent research. In the largest recent study (a 252,000-trial study across 6 models (https://arxiv.org/abs/2605.25517)), trust cues and topic match dominated determining source citations. A primary source is a trust cue.
This reinforces a point I’ve made previously. The brand’s own site is only surface number 6 in how LLMs create connections. The top five are third-party and comparison content (https://sandeepamar.in/googles-ai-optimization-guide-is-not-the-aeo-bible-its-a-press-release/). The new development? In that third-party content, the primary sources are outpacing the summarizers. Being in the right list isn’t going to help if that list is just a list.
Bottom line
ChatGPT is rapidly trading aggregators for primary sources. You can’t sit this trend out. In the last 6 months, Reddit has gone from 1 in 5 to 1 in 43. If your AEO plan relies on citing Reddit threads or Wikipedia, you have a time bomb. The best way to put it is to be a primary source or be cited by one. Time is on the side of primary research and testing. There is a high probability it is just going to skip the middleman and give the results. It’s not going to turn back.
