Where AI Citations Come From (Mostly Not Your Site)
Across 25 million links cited by AI engines, 84% pointed at domains the brand doesn't own — but on the evaluation-stage prompts that decide shortlists, the ratio flips. Here's what that changes about where you spend your AEO effort.
Updated · Published
Frequently asked questions
What share of AI citations come from a brand's own website?
A minority. Muck Rack's May 2026 edition of "What Is AI Reading?" analysed more than 25 million links cited by ChatGPT, Claude, and Gemini across 17 industries and attributed 84% of citations to earned media, meaning sources outside the brand's own domain. The figure has stayed between 82% and 89% across three editions since July 2025.
Does that mean on-site AEO is pointless?
No. It means on-site work has a ceiling, not that it has no value. Your pages still have to be crawlable, parseable, and accurate or you cannot be cited even on the queries where you would win, and Google's own guidance is that its AI features run on core Search ranking. The finding argues for rebalancing effort toward off-site surfaces, not abandoning your own site.
Which websites do AI engines cite most?
Across published studies the recurring names are Reddit, Wikipedia, LinkedIn, YouTube, and large editorial outlets. Semrush tracked 100 million-plus citations over 13 weeks in 2025 and found those domains consistently at the top. But the mix differs sharply by engine, and your own category's list will differ from any published one.
Why does Yext say 86% of AI citations come from brand-managed sources?
Because it measured a different query universe with a different taxonomy. Yext's October 2025 study analysed 6.8 million citations from 1.6 million location-based queries in retail, financial services, healthcare, and food service, and counted both first-party websites (44%) and third-party listings the brand can edit (42%) as brand-managed. Muck Rack's 84% earned figure comes from industry-topic prompts and classes anything off the brand's own domain as earned. The two studies largely classify the same kinds of citations in opposite directions, and neither describes your prompts.
Do different AI engines cite different kinds of sources?
Substantially. In Profound's analysis of roughly 680 million citations, Wikipedia accounted for 7.8% of ChatGPT's citations but only 0.6% of Google AI Overviews' and Perplexity's, while Reddit ran at 6.6% on Perplexity against 1.8% on ChatGPT. Engines also cite at different rates: Muck Rack recorded citations in 96% of ChatGPT responses, 82% of Gemini's, and 55% of Claude's.
Does the type of question change which sources get cited?
Yes. Muck Rack found industry trend queries cited journalism in 46% of responses, more than double the rate for how-to and comparative queries, which lean instead on reference material and brand-owned pages. Your prompt mix therefore determines which off-site surfaces are worth the investment.
Do evaluation-stage prompts cite different sources than general prompts?
Substantially. Ten Speed's September 2026 analysis categorized just over 7,000 citations from 220 B2B evaluation-stage prompts — comparisons, alternatives, use cases, buying-intent questions — run through Peec AI across ChatGPT, Perplexity, Claude, and Gemini. Product pages were the single largest cited format at 24.1%, community sources like Reddit and YouTube totalled 4.2%, and formats a marketing team can publish summed to 88.3% of all citations. The caveat: the study does not check whether a cited comparison page or listicle sits on the recommended brand's domain or a competitor's.
How do I find which third-party sites shape my AI visibility?
Export the cited domains from your tracked prompts on every engine, including the runs where your brand was not mentioned, then group them by type and rank them by citation count. The top off-site domains for your prompts are your target list. Open the cited pages and check whether they represent you accurately.
The standard AEO checklist is almost entirely about your own website: headings, schema, llms.txt, answer blocks, freshness dates. The citation data says that is a minority of the game. Muck Rack's May 2026 edition of What Is AI Reading? analysed more than 25 million links cited by ChatGPT, Claude, and Gemini across 17 industries, and attributed 84% of those citations to sources outside the cited brand's own domain. Paid and advertorial content accounted for 0.3%.
Key takeaways
84% of AI citations pointed at domains the brand does not own. Across three editions of the study going back to July 2025 the figure has held between 82% and 89%, which makes it a property of how these systems source information rather than an artefact of one model update.
Read the taxonomy before the headline. Muck Rack's "earned media" bucket includes journalism, academic work, government and NGO sources, encyclopedic sites, social platforms, and other companies' corporate content. It is closer to "not yours and not paid for" than to "press coverage."
Journalism alone was 27% of citations, and stayed in a 25–27% band across all three editions.
The mix is engine-specific. Wikipedia was 7.8% of ChatGPT's citations but 0.6% of Google AI Overviews' and Perplexity's in Profound's 680-million-citation analysis; Reddit ran at 6.6% on Perplexity against 1.8% on ChatGPT.
Query type shifts it again. Trend questions cited journalism in 46% of responses, more than twice the rate of how-to and comparative questions.
A counter-headline exists and is also true: Yext measured 86% of citations from brand-managed sources — on location-based queries, counting third-party listings as brand-managed. The clash with the 84% is a clash of query mix and taxonomy, not of data.
A third pole arrived in September 2026: across 220 B2B evaluation-stage prompts, Ten Speed found product pages the largest cited format at 24.1% and Reddit, YouTube, and forums at 4.2% combined. The shortlist-deciding prompts lean on vendor-publishable pages — with the caveat that "publishable by a marketing team" does not mean "published by yours."
None of this makes on-site work worthless. It puts a ceiling on it, and the ceiling is lower than most AEO checklists imply.
The number is easy to misread in a flattering direction, so start with the definition. Muck Rack groups cited sources into earned, owned, and paid, and its earned category covers journalism, academic research, government and NGO material, encyclopedic sites like Wikipedia, social platforms, and third-party corporate content.
That last item is the one to notice. Another company's blog post about your category counts as earned media from your point of view. So does a Wikipedia article, a Reddit thread, a university page, and a government dataset. The honest reading of 84% is not "press coverage drives AI answers." It is 84% of citations pointed somewhere you cannot edit.
Muck Rack is a PR software company, and its framing leans toward the journalism slice of that bucket, which is the slice its product addresses. Worth holding in mind. It is also worth noting that the broader reading survives the discount: whether the off-site share is 84% or somewhat lower under a stricter taxonomy, it is not a rounding error, and it has been stable for a year across three separate measurement runs.
Category
Share of citations
What it includes
Earned
84%
Journalism, academic, government and NGO, encyclopedic, social platforms, third-party corporate
Journalism (subset of earned)
27%
Editorial coverage from news outlets
Paid and advertorial
0.3%
Sponsored placements, advertorials
The 0.3% deserves a moment on its own. Whatever else is true about AI search, you cannot buy your way into the citation set. That is a meaningful difference from every paid channel marketers are used to reaching for when an organic number goes the wrong way.
One vendor study with a self-serving frame would not be enough. Two more, with different methods and different commercial incentives, land in the same place.
Semrush tracked the top 25 most-cited domains weekly across ChatGPT Search, Google AI Mode, and Perplexity for 13 weeks, covering 230,000-plus prompts and over 100 million citations. The names at the top were Reddit, Wikipedia, LinkedIn, YouTube, and Google's own properties. Not a single brand's product site among them.
Profound's analysis of roughly 680 million citations found commercial .com domains taking 80.4% of all citations and .org 11.3%, with the leading individual sources again being platforms and publishers rather than vendors.
Our own 42-day citation study shows the same thing from the other end. Google AI Mode cited about 24 domains a day and ChatGPT about 17, and for a single broad prompt — "alternatives to Nike for high-performance running" — Google drew on 445 distinct domains over six weeks. Any one brand's own site is one entry in a pool that size, competing against hundreds of pages written about the category rather than by a participant in it.
A statistic making the rounds in AEO feeds this August appears to flip the conclusion outright. Yext analysed 6.8 million citations from 1.6 million queries across ChatGPT, Gemini, and Perplexity and headlined the result 86% of AI citations come from brand-managed sources. Set beside Muck Rack it reads like a contradiction: 84% not yours against 86% yours, from corpora of comparable size, on overlapping engines.
Both numbers are defensible. They differ on the two dials that decide every citation study: what was asked, and how the sources get bucketed.
The query universe. Muck Rack's corpus is industry-topic prompts across 17 industries. Yext's is location-based queries — branded and unbranded, objective and subjective, run across thousands of locations in retail, financial services, healthcare, and food service, using July–August 2025 data. That is the "pharmacy open near me" end of AI search, where the correct answer is one business's address, hours, and phone number. Local queries resolve to the business's own footprint almost by construction, because nobody else publishes your opening hours.
The taxonomy. Yext's 86% is 44% first-party websites plus 42% listings — business profiles on platforms like Google, Apple Maps, Yelp, MapQuest, and TripAdvisor. A listing sits on a domain the brand does not own but holds data the brand can edit. Classify by domain ownership, as Muck Rack does, and most of that 42% lands in the earned bucket; classify by manageability, as Yext does, and it lands in owned. The same MapQuest citation is evidence for both headlines. Strip the listings out and Yext's own-domain share is 44% — higher than a general corpus, as you would expect for local queries, but no longer a refutation of anything.
The vendor symmetry from earlier holds perfectly here. Muck Rack sells PR software and its framing foregrounds the journalism slice; Yext sells listings management and its framing foregrounds the listings slice. Each study finds the game centred on the surface its vendor sells. Neither is fabricating; both are choosing the frame, and both publish enough method to see the choice.
What survives the framing fight is genuinely useful. For location-shaped queries, engines lean hard on structured business data and barely on forums: Yext put Reddit and similar sites at 2% of citations once location and intent were applied — against Reddit's 6.6% of Perplexity's citations in Profound's general corpus — and its CDO summarised it as "forums are less influential than expected." The per-engine split moves too: in Yext's data Gemini favoured websites at 52.1% of citations while ChatGPT leaned on listings at 48.7%. If your buyers ask questions with a location in them, listings hygiene is a citation lever the 84% story undersells, and it is the cheap kind: the audit is opening your own profiles. If your prompts have no location in them, Yext's numbers do not describe your game at all.
The reconciliation fits in one sentence: how much of your AI citation mix you control is not a property of AI search, it is a property of your prompt set. Which is one more argument for measuring your own mix — the audit below — instead of adopting either vendor's headline.
"Go off-site" is not a strategy until you know which off-site. The engines diverge more than they converge.
ChatGPT
Google AI Overviews
Perplexity
Claude
Share of responses with citations
96%
—
—
55%
Average citations per response
~5
—
—
~13
Wikipedia share of citations
7.8%
0.6%
0.6%
—
Reddit share of citations
1.8%
2.2%
6.6%
—
YouTube in top sources
No
1.9%
2.0%
—
Character of the source mix
Encyclopedic and factual
Balanced across social and professional sources
Community discussion
Cites least often, most sources when it does
Citation rates are from Muck Rack (Gemini sat at 82% of responses and about 8 citations each); domain shares are from Profound. Blanks are where the studies do not report a comparable figure, which is itself a reminder that no single dataset covers every engine cleanly.
Two practical consequences. First, a category whose buyers use Perplexity should treat community presence as core infrastructure, while one whose buyers use ChatGPT should care disproportionately about how reference sources describe the category. Second, these shares move: inside a single 13-week window Semrush watched Reddit's presence in ChatGPT responses fall from roughly 60% to about 10%, and Wikipedia's from around 55% to under 20%. Build a strategy on one engine's current source preferences and you have built on sand.
Within a single engine, what gets cited depends on what was asked. Muck Rack found industry trend queries cited journalism in 46% of responses, more than double the rate for how-to and comparative queries, which lean on reference material and brand-owned pages instead.
This is the finding that makes the whole thing actionable, because it means your own prompt mix decides where off-site effort pays. A brand whose tracked prompts are mostly "how do I do X" questions has more on-site leverage than the 84% headline suggests. A brand whose prompts are mostly "what is happening in this market" questions has almost none, and needs editorial coverage. Most brands have both and have never separated them.
So separate them. The prompt set you already track can be split by question type in an afternoon, and the split tells you which half of the work to fund.
For a year, the missing dataset in this argument was the one marketers care about most: not "what does AI cite across an industry" but "what does AI cite on the comparison and alternatives prompts that decide a shortlist." In September 2026, Ten Speed — a B2B content agency — published an analysis of exactly that slice: 220 hand-written evaluation-stage prompts (20% comparisons, 20% alternatives, 20% use cases, 19.5% feature and integration questions, the rest buying-intent and brand-direct) run through Peec AI's monitoring of ChatGPT, Perplexity, Claude, and Gemini, with just over 7,000 resulting citations categorized by content format.
The mix looks nothing like the general corpora:
Format
Share of citations
Product pages
24.1%
Articles (blog, news, PR)
17.4%
Comparison pages
13.3%
Listicles
13.2%
How-to guides
8.9%
Homepages
7.8%
Review-site profiles (G2, Capterra)
7.2%
Community discussion
3.7%
Alternatives pages
2.4%
Category pages
1.2%
Video
0.5%
Product pages — the format general studies barely register — were the single largest category. Reddit, YouTube, and forums combined for 4.2%, with YouTube at roughly 1%. Everything a marketing team can publish sums to 88.3%. Comparison-format prompts over-earned: 20% of the prompt set drove roughly 27% of citation volume, about 1.33× their share.
Set 88.3% next to Muck Rack's 84%-not-yours and Yext's 86%-brand-managed and you have three headline numbers apparently at war. It is the same reconciliation a third time: general industry prompts source from the open web, location prompts source from listings, and evaluation-stage prompts source from the category's own vendors. The controllable share of your citation mix is a property of your prompt set — now demonstrated at all three points of the triangle.
Two caveats before this becomes a victory lap for on-site work. First, the vendor symmetry holds yet again: a content agency measured the citation landscape and found that content wins, just as the PR vendor found journalism and the listings vendor found listings. Second — and this is the limitation the study itself does not dwell on — "brand-controllable" describes the format, not the owner. Ten Speed never checks whether a cited comparison page or listicle sits on the recommended brand's domain. A comparison page is controllable by whoever wrote it, and in most categories the page AI cites for "alternatives to X" was written by one of X's competitors. The sample is also one agency's B2B SaaS client base, a single snapshot, from one monitoring tool, with no per-engine breakdown.
What the study does establish, caveats intact: on the prompts closest to a buying decision, the citation pool is made of pages someone's marketing team wrote. Whether yours is in it is the part you control.
If you want to get featured in the third-party listicles and comparison pages already influencing LLM answers, you can use AI citation outreach tools like ListBrew. It finds pages that feature your competitors but not you, shows which of them are cited by Google AI Overviews, ChatGPT, Gemini, and Perplexity, and surfaces editor contacts so you can pitch for inclusion.
Being straight about the limits matters more than the headline here, because "your website doesn't matter" is a conclusion the data does not support and one that would cost you real visibility.
The big corpora are citation counts across all queries in an industry, not brand-specific commercial queries. Ten Speed's evaluation-stage data closes part of that gap and confirms the expectation — vendor pages do far better on high-intent prompts than across a general corpus — but it cannot tell you whose vendor pages win in your category, and its sample is one agency's client base on one day. That experiment is still the one you run yourself, on your own prompts, and it is described in the next section.
The categories are also softer than the percentages imply. A bucket that contains both a Reuters investigation and a competitor's blog post is not measuring one thing, and the published summaries do not give a clean own-domain share to set against the 84%. The closest published figure — Yext's 44% for first-party websites — comes from an all-local corpus in four sectors, which answers a different question than a general prompt set does.
Most importantly, on-site work is a precondition, not a competitor to off-site work. If AI crawlers cannot reach your pages, or your content is unparseable, or your facts are stale, you forfeit the citations you would otherwise win — including the how-to and comparative queries where brand-owned pages do get cited. Google's own guidance on AI features is explicit that its generative features run on core Search ranking and that there is no separate playbook. Structured data and clear answer structure remain worth doing. They are just not the whole job, and the checklist industry has been selling them as if they were.
Digiday's July 2026 survey of agency practitioners lands on the same balance. Nate King of RPA put the overlap with conventional SEO at "80 to 90% of the same tactics" while stressing that the remaining "10 to 20% is super critical" — and VML's Heather Physioc was blunter about the vendors promising more: "Anyone promising anything… is a liar."
The general studies tell you the shape of the problem. Only your own data tells you where to act, and the audit is a couple of hours of work.
Export every cited domain from your tracked prompts, on every engine, including the runs where your brand was never mentioned. Those runs are the informative ones: they show which sources are carrying an answer you are absent from.
Group the domains by type. Your own site, editorial outlets, forums and communities, review platforms, reference sites, video, competitors' own domains. A ranked list of 300 domains is data; seven buckets with percentages is a decision.
Compute your own domain's share of all citations. That number is the ceiling on what on-site optimisation can reach for these prompts. If it is 6%, you now know what the other 94% of your effort should be aimed at.
Rank the third-party domains by citation count. The top ten or twenty are your real target list. It will be category-specific and it will not match Reddit-Wikipedia-LinkedIn, because your buyers ask narrower questions than a general corpus samples.
Open the cited pages and read them. Are you listed? Described accurately? Priced correctly? Present at all? A review profile three versions out of date and a comparison thread that omits you are concrete, fixable gaps, and fixing them is faster than earning a new citation from scratch.
Re-run it quarterly. Source preferences shift fast enough that a target list built a year ago is pointing at the wrong surfaces.
Then work the list. The channels that move off-site citations are the unglamorous ones: accurate review-platform profiles, genuine participation in the communities where your category is discussed, relationships with the journalists who actually cover your market, and documentation good enough that other people cite it. None of it is a hack, which is precisely why it holds. As Digiday's practitioners noted, the manipulation-shaped tactics — keyword stuffing, link schemes, llms.txt as a magic file — are the ones showing the least effect.
On-site AEO became the default advice because it is the part of the problem a brand fully controls, and controllable work is easier to sell, scope, and report on. The citation data is a reminder that the controllable part is not the large part.
The reframe worth carrying: your website is how you become citable, and the rest of the web is what decides whether you get cited. Both halves need budget, and for most brands the second half currently has none.
Aeonav is an open-source, self-hosted AI visibility platform that runs your prompt sets across ChatGPT, Claude, Gemini, Perplexity, and Google's AI surfaces, recording every citation, mention, and competitor named alongside you. Because it stores every cited domain and the data is yours to query directly, the source-mix audit above is a grouped query over your own results rather than a research project.