The 50,000-brand citation study: who answer engines actually trust
Every few months someone claims to know which brands "win" AI search. This year the guessing stopped: three independent research groups published citation datasets big enough to take seriously. Semrush and Growth Memo tracked 50,000 brands across 1,094 topics in ChatGPT for six months. Peec AI analyzed 30 million cited sources across five engines. And HubSpot and Wix Studio between them classified over a million citations by content format. Read together, they answer the question in this headline with unusual precision. The answer is not what most marketing plans assume.
The scoreboard nobody owns
The Semrush study is the largest brand-level look at ChatGPT visibility published to date: 1,094 US topic categories, each tracked monthly from January through June 2026 through five buyer prompts, producing over 600,000 citations. Its central finding is that topic ownership barely exists. A brand "owns" a topic when it appears in at least four of the five related prompts with a five-point lead over the runner-up. Only 15.2% of categories cleared that bar.
It gets stranger. In the high-demand half of topics, the half that accounts for 98% of AI search volume in the sample, ownership drops to 11.3%. The topics people actually ask about are the ones nobody has claimed. And once a brand does take a topic with a real margin, it keeps it: owners with a five-point lead held first place in 90% of month-over-month comparisons. The land grab is real, and it is still early.
The trust list is short, and it is not your website
Peec AI's dataset answers a different question: when engines cite, where do the citations go? Across 30 million cited sources in the US, the top ten domains are Reddit, YouTube, LinkedIn, Wikipedia, Forbes, G2, Yelp, Facebook, Medium and TechRadar. Not a single brand-owned site among them. The engines disagree with each other, though, and the differences are big enough to change strategy:
The concentration numbers behind that table are stark. Contently's April 2026 tally put Wikipedia at 47.9% of ChatGPT's top-ten citation share and Reddit at 46.7% of Perplexity's. And where a Google results page spreads attention across roughly ten organic slots, a Q2 2026 analysis of 5,000 queries found AI answers typically cite just three to six domains. The funnel into an AI answer is far narrower than the one into a results page, and much of it runs through platforms you do not control.
Authority is not the predictor you think it is
The most uncomfortable finding in the Semrush data is what failed to predict ownership. Comparing each topic owner against its runner-up on classic SEO strength, the owner was better on branded search volume 55.7% of the time, on Authority Score 52.5% of the time, and on organic traffic just 48.4% of the time. Two of those three are statistically indistinguishable from a coin flip.
The Q2 citation analysis reached the same conclusion from the content side: pages built around quotable data points, clear definitions and extractable tables were cited more often than higher-authority pages without those properties. The researchers called it linkability beating authority. Authority still functions as a floor. It just stopped working as a ladder.
"Domain authority predicted the topic winner about as well as a coin flip. Whatever decides who owns a topic in AI search is not captured by broad SEO strength."
There is a sharper edge to this than lost visibility. Webflow data reported by CMSWire found that when AI does mention a brand, it misrepresents it roughly a third of the time. Absence is not the only risk. Being described wrongly, at scale, by a system users trust, is the other half of the problem.
The three habits the most-cited brands share
They publish liftable evidence. The Princeton-led GEO benchmark, the study that named generative engine optimization, measured what actually moves citation likelihood: adding expert quotations lifted visibility in generative answers by 41% on its position-adjusted metric, and adding statistics by around 37%. HubSpot's 2026 dataset independently found the same structural signals correlated with citations: statistics, visible last-updated dates, author bios and FAQ blocks. Engines cite passages that carry their own proof.
They match format to intent. Across the HubSpot and Wix Studio datasets, listicles, articles, product pages and category pages earn the most citations overall, and comparison content tops ChatGPT specifically with a 95% citation rate, the highest of any format on any engine. Third-party discussion threads punch far above their weight on Perplexity, where discussions take 17.35% of citations, more than double the cross-engine average.
They cover the whole topic, not the hero keyword. Ownership in the Semrush data required showing up across definition, comparison, alternatives, use-case and buying prompts at once. Brands that won one prompt and vanished from the next four did not register as owners at all. Winning a topic means answering all five ways a buyer asks about it.
What to do this week
Pick the one topic your revenue depends on most. Write down the five questions a buyer would ask inside it, then check what each engine actually cites for those questions. If the answer is Reddit threads and review sites, that is your distribution plan, not a nuisance. Then rebuild your own key page so every claim carries a number, a source and a date. The scoreboard says 85% of topics are still open. That is not a threat. For one more quarter, it is an invitation.