Ahrefs studied 66.6 million free, logged-out ChatGPT answers from August and September 2026, plus roughly 15 million of the follow-up searches ChatGPT runs behind the scenes. The full study, published by Ahrefs on October 12 and updated October 22, upends a handful of assumptions marketers have been repeating about ChatGPT citations since early 2026. If you have been building a GEO plan around guesses rather than data, this is the update to read.
Below are five common claims about ChatGPT citations and search behavior, checked against the Ahrefs numbers. Some hold up. Others do not survive contact with 66.6 million real responses.
What Ahrefs Actually Measured
Ahrefs pulled free, logged-out ChatGPT responses running on GPT-5.6 Luna, specifically because logged-out sessions strip away chat history and personalization. That gives a cleaner read on default behavior, the kind most first-time users and casual searchers get.
The fan-out numbers (how ChatGPT breaks a prompt into sub-queries before answering) come from that same 66.6 million-response set. The citation numbers are a separate dataset, pulled from Ahrefs’ Brand Radar tool, covering 2.4 to 3.8 million US answers per month between March and September 2026. Ahrefs compared about 25,000 prompts that showed up in both the March and September samples, specifically to rule out the idea that shifting prompt mixes were driving the citation changes.
That matters for how you read every stat below. Fan-out behavior and citation ranking are two different measurements, on two different samples, and Ahrefs is careful to say so. Treat them as two related but separate signals, not one combined metric.
Myth 1: ChatGPT Searches the Web for Almost Every Answer
This is the biggest surprise in the whole report. Only 14.4% of the logged-out responses Ahrefs studied contained a web search at all. The rest, more than 85%, were answered straight from the model’s own training, with no live lookup.
When ChatGPT did search, it ran an average of 1.6 fan-out queries per response, meaning it usually split one prompt into one or two sub-searches rather than a long chain of them. That is a much lighter search footprint than most GEO commentary assumes.
Ahrefs flagged one wrinkle worth knowing: an earlier version of its own post described this figure as the share of responses that searched the web, then corrected it on October 22 to the share of responses containing fan-out queries specifically. The two aren’t identical, and the update is a useful reminder to read methodology notes on any AI search study, including this one.
What this means for you: do not assume every reader question about your product or industry triggers a live crawl of your site. For a lot of free-tier ChatGPT traffic, the model is answering from memory. Fresh content still matters, but it is competing for a smaller slice of live searches than the hype suggests, and that smaller slice is exactly where ChatGPT citations get decided.
Myth 2: Site-Specific Searches Mostly Target Big Retailers
Partly true, but the shape is different than most people assume. Of the roughly 160 most-searched domains using the site: operator, 75.9% were direct brand sites, like apple.com, ikea.com, nike.com, adidas.com, and samsung.com. Retailers such as Walmart, Costco, and Amazon made up 11.9%. Reddit and Facebook combined for just 1.9%.
Apple’s own domain led the pack with 9,442 site: mentions in the dataset. Reddit still showed up, ranked ninth with 2,586 mentions, but it was nowhere near the top of the list the way some SEO forum chatter implies.
Government and regulator domains also cracked the top 150, including gov.br, gov.uk, service-public.fr, the NHTSA, the FDA, and the SEC. That is a meaningful signal for any brand in a regulated space: official government pages are already earning ChatGPT citations through this kind of site-targeted search, not sitting on the edge case list.
Myth 3: Reddit and Wikipedia Dominate ChatGPT Citations
This was true earlier in 2026. It is not true anymore, according to the same Brand Radar data. Between March and September, Reddit’s citation share dropped 88%, going from roughly one in five ChatGPT answers down to about one in 43. Wikipedia fell 80% over the same stretch, from roughly one in seven answers to about one in 35.
YouTube slipped too, down 27%, from about one in 370 answers to one in 500. Meanwhile, PubMed jumped from 64th to 4th among the most-cited domains, and the FDA entered the rankings at 8th. WebMD and similar health explainer sites fell in rank over the same window.
Ahrefs’ own read on this: ChatGPT “seems to be prioritizing original sources over sites that summarize them.” My take is that this lines up with what a lot of us have noticed anecdotally in GEO work this year, that aggregator and forum content is losing ground to primary research, regulators, and brand-owned pages. The Ahrefs numbers are the first large-scale confirmation I have seen of that shift.
We have covered a related angle on this before. If you want more detail on why two AI engines can answer the same question with completely different source lists, see our piece on why ChatGPT and Perplexity cite different sources for the same query.
Myth 4: Official and Government Sources Rarely Get Cited
Also false, and this is where the ChatGPT citations data gets genuinely useful for content planning. In a separate 20,000-prompt US test, 84.4% of factual questions produced at least one ChatGPT citation to an official source, and 63.4% of citations overall went to official websites.
Comparison-style questions (“best X for Y”) cited official sources less often, at 68.3%, with 46.4% of citations landing on official sites. That gap makes sense. A comparison question invites third-party opinion by nature, while a factual question about, say, a drug interaction or a tax deadline, pulls toward the source that actually owns that information.
Publisher partners that have licensing deals with OpenAI also moved. Their share of citations among ChatGPT’s top 200 cited domains rose from 7.5% in March to 15.8% in July, then settled at 12.5% by late September. Only one of those 17 partner publishers and platforms, Reddit, showed up in the top 150 site:-searched domains, which tells you licensing deals and organic site-search visibility are two different kinds of ChatGPT citations, not the same thing.
Myth 5: Fan-Out Query Wording Stays Pretty Stable Month to Month
Not according to the data. Ahrefs tracked the actual words people’s prompts triggered inside fan-out queries, and the shift from August to September was fast. The word “official” appeared in 13.1% of English fan-out queries in September, up from 10.1% in August. “Price” rose from 6.04% to 8.88%, a jump of roughly 47% in a single month.
“Reviews” grew by about a third over the same period, and “specifications” entered the top 50 fan-out terms for the first time. On the other end, “meaning” and “definition” fell by roughly a fifth, though both started from a small base.
That is a fast-moving signal. If ChatGPT users are increasingly fan-searching for “official,” “price,” and “specifications,” that tells you which page types (spec sheets, pricing pages, verified product pages) are becoming more valuable targets for ChatGPT citations right now, this month, not next year.
How to Check Your Own ChatGPT Citations
The Ahrefs numbers are a population-level view. Your own pages might already be picking up ChatGPT citations, or losing them, in ways that match or break from the overall trend. A quick manual check takes about ten minutes.
- Open a logged-out ChatGPT session (no memory, no chat history) and ask five to ten real factual questions a customer would type about your product category.
- Note which responses show a visible “Searched the web” or source-citation marker. Based on the Ahrefs data, expect that marker on well under a fifth of them.
- When citations do appear, record which domains show up: your own site, a competitor, a government page, a forum thread, or a licensed publisher.
- Repeat the same prompts with the word “official” or “price” added, since Ahrefs found both terms rising fast inside fan-out queries this fall.
- Run the same prompt set again in two to three weeks. ChatGPT citations shift with model updates, so a single snapshot tells you less than a short trend line.
This will not replace a dedicated AI-visibility platform, but it is a free, fast way to sanity-check whether the Ahrefs trends (fewer searches overall, more weight on official and brand-owned pages) show up in your own category. If the pattern holds, your highest-value move is making sure your pricing, specification, and “official” pages are complete, current, and easy for a crawler to parse, since those are exactly the page types gaining ground in ChatGPT citations right now.
Search Engine Journal’s own coverage of the study, written by Matt G. Southern, adds a useful caution here too: citation rank is not the same as traffic. A domain can climb in ChatGPT citations without sending you a single extra visitor, since most chat interfaces do not surface clickable links the way a traditional search results page does.
What the Limits of This Study Are
Ahrefs is upfront about the gaps, and you should weigh them before changing a content strategy. The fan-out data reflects one model, GPT-5.6 Luna, over two months. OpenAI changes retrieval behavior often, and the authors say results could shift with the next model update.
Logged-out, free-tier sessions search less than paid or reasoning-mode sessions, so this is a floor on search frequency, not a ceiling. The study also notes that its fan-out and site: volumes differ from other published research, including a Nectiv report that found 7.61 fan-outs per prompt and a 64% site: share, roughly ten times what Ahrefs measured. Different prompt sets and different models explain most of that gap, but it is a reminder that no single study is the final word on ChatGPT citations or ChatGPT search behavior generally.
One more caveat worth naming plainly: Ahrefs sells Brand Radar, the tool behind the citation half of this data. That does not make the numbers wrong, but it is a reason to read the conclusions as a vendor’s interpretation of its own data, not a neutral third-party audit.
What This Means for Your ChatGPT Citations Strategy
Pull these threads together and a practical picture forms. Most free-tier ChatGPT answers never touch a live search, so the content you publish competes for a visible but smaller slice of total ChatGPT traffic. When ChatGPT does search, it increasingly favors brand-owned pages, government and regulator sources, and original research over forums and general-purpose encyclopedias.
In my experience advising on GEO work, the practical move is to treat “official” as a target word, not a tone. If Ahrefs is right that fan-out searches for “official” grew by 3 full percentage points in a single month, then pages that read as the official source, specs pages, pricing pages, and regulator-adjacent documentation, are the ones most likely to pick up ChatGPT citations in the near term.
For a broader framework on building content that earns this kind of visibility, our guide to content strategy for AI search walks through the planning steps in more depth. If you want to track whether any of this is actually showing up in your own AI-driven traffic, our comparison of Ahrefs and Semrush AI visibility tools covers how to monitor it.
Frequently Asked Questions
Does ChatGPT search the web for every question?
No. Ahrefs found that only 14.4% of logged-out, free-tier ChatGPT responses contained a web search at all. The majority of answers come straight from the model’s training, with no live lookup.
Has Reddit really lost ground in ChatGPT citations?
Based on Ahrefs’ Brand Radar data, yes. Reddit’s citation share dropped 88% between March and September 2026, from roughly one in five answers to about one in 43.
What is a fan-out query in ChatGPT?
A fan-out query is one of the sub-searches ChatGPT runs behind the scenes to answer a single user prompt. Ahrefs found an average of 1.6 fan-out queries per response that included any search at all.
Should I stop targeting Reddit and Wikipedia-adjacent content entirely?
Not entirely. Both still appear in ChatGPT citations, just far less often than earlier in 2026. The bigger opportunity, per the data, is building content that reads as an original or official source rather than a summary of one.
Why This Matters
This study replaces guesswork with a large, dated sample, even with its limits. The headline for most sites: ChatGPT searches less than assumed, and when ChatGPT citations do appear, they are pulling away from forums and summary sites toward brands, regulators, and original research.
Next steps worth taking this week: audit your highest-traffic pages for whether they read as an “official” source on their topic, check your spec and pricing pages are complete and crawlable, and revisit the Ahrefs study again after the next ChatGPT model update, since the authors themselves expect these numbers to move.
Leave a response