ChatGPT's top ten domains take just 12 percent of its citations. It cites individual Reddit threads 99 percent of the time and YouTube channel pages 5 percent of the time. Domain authority is the wrong mental model. Seven rules from 730,000 conversations.
Ask most SEO teams how to rank in ChatGPT and the answer comes back in the vocabulary of Google circa 2015: build authority, earn links, get the domain trusted. It is a reasonable guess. It is also, according to the largest public dataset on ChatGPT's citation behavior, mostly wrong about the unit that matters.
Profound analyzed roughly 730,000 ChatGPT conversations that contained at least one web citation, from US English users between October and December 2025, and published a series of follow-ups through 2026. The consistent finding is that ChatGPT picks a piece of content that answers the question. It does not pick a brand and then find something from it. Here is what that looks like in the numbers, and seven rules that fall out of it.
Wikipedia is ChatGPT's most-cited domain, at about 5 percent of citations and appearing in 18 percent of cited conversations. Reddit is next at 3 percent. After that the shares fall off fast: Reuters and the NIH at around 1 percent each. The top ten domains combined capture 12 percent. Everything else, 88 percent, is spread across a tail so long that Profound measured the inequality with a Gini coefficient of .8, which sounds concentrated until you remember that the concentration is at the level of individual pages rather than sites.
That is the first thing to unlearn. There is no small club of domains that ChatGPT trusts and everyone else is locked out of. There is an enormous number of pages, most on sites you have never heard of, each of which happened to be the best available answer to one question.
The clearest evidence is on the platforms where you can tell the difference between the entity and the content. In Profound's sample of about 700,000 social citations from ChatGPT, 99 percent of Reddit citations pointed to individual discussion threads rather than subreddits or the site itself. On YouTube, 85 percent of citations went to specific videos, 5 percent to channel pages, 3.4 percent to playlists.
LinkedIn tells the same story over time. Between November 2025 and February 2026, LinkedIn climbed from around eleventh to fifth among ChatGPT's cited domains, more than doubling its citation frequency, and became the most-cited domain for professional queries across six engines. Look at what got cited and the shift is unmistakable: profile pages fell from 33.9 percent of LinkedIn citations to 14.5 percent, while posts rose from 20.9 to 26 percent and long-form articles from 6 to 8.9 percent. The engine stopped citing who someone is and started citing what they wrote.
ChatGPT's citations aligned with Google's top results 12 percent of the time in April 2025 and 33 percent by July, while its alignment with Bing fell from 26 percent to 8 percent over the same months. So Google's index has become the more important one to be in. But being in it is not the same as ranking first in it. Across 240 million ChatGPT citations, Google's first position drew 27.5 percent of human clicks and only 10 percent of ChatGPT citations. Tenth position drew 2.5 percent of clicks and 4 percent of citations. The click curve is a cliff. The citation curve is close to flat across the top ten.
And ChatGPT still chooses independently most of the time. Its sources overlapped with Google's only 39 percent in a separate 250-million-answer analysis. Being indexed and on page one gets you into the pool. What gets you cited is whatever happens on the page.
Every number above points the same way. The unit of citation is the thread, the video, the post, the article. A homepage that describes the company well is competing for almost nothing. A page that answers one buying question completely is competing for that question, and that is the contest that exists.
Turn one is 2.5 times more likely to produce citations than turn ten, and nearly four times more than turn twenty. As a conversation runs on, ChatGPT increasingly answers from what it already has. The question a person opens with is your window. Map those opening questions in your category and build the page that answers each one directly.
Sixty-six percent of cited turns had between one and four unique sources, and a typical conversation drew about six. You are not trying to be the only source. You are trying to be one of a handful, which means being clearly better on one specific point beats being adequate on everything.
Across 11.84 billion citations from eight engines in mid-2026, brand-operated sites took 57 percent overall and 47 percent on ChatGPT specifically, the lowest of the engines but still the largest single category. ChatGPT also leans on institutional sources more than its rivals, at 15.3 percent of citations in one 27-million-citation study. The pages you control are the most reliable route in. Third-party coverage helps; it does not replace them.
With alignment to Google at a third and rising, indexation and a top-ten ranking are the entry ticket. But tenth place earns 40 percent as many citations as first. Once a page is on page one, effort spent moving it from seventh to third buys very little in ChatGPT. Effort spent making the page more liftable buys a lot.
A LinkedIn profile, a YouTube channel page and a subreddit are all entities. ChatGPT cites what is inside them: the post, the video, the thread. If you are going to be on a platform at all, the work is the individual piece that answers a question, and it needs to stand alone because that is how it will be found.
Only about 18 percent of ChatGPT conversations trigger a web search. For the other 82 percent, the model answers from training, and no page you publish this month will be consulted. Your content competes only for the questions that make the model go looking: the recent, the comparative, the specific, the local. Building pages for questions the model already knows the answer to is publishing into the lane where nothing gets cited.
The core conversation data is US English, fourth quarter of 2025, from Profound's tracking. The LinkedIn figures run to February 2026 and the brand-site share to July 2026. ChatGPT changed its citation behavior materially in October 2025 and again in June 2026, so the shape of these findings is more durable than the decimals. And "cites pages, not brands" is my reading of the numbers; Profound reports the data and is more careful about drawing the line.
If the reading is right, the practical consequence is that a content audit should score each page on whether it answers a real opening question, whether a machine can lift the answer, and whether it is actually being cited on the engines your buyers use. That is the shape of what Prism Optimizer does, and it is also the shape of a spreadsheet a patient person could keep. Either way, the domain-level scorecard you are probably still looking at is measuring the wrong thing.
Score every page, not the domain.Put your site through the Prism and see which pages ChatGPT, Gemini, Perplexity and Claude cite, which they ignore, and why.
Request a demo