Where AI answers get their sources
The short version
- Semrush found 62% of AI citations are "ghost citations" — the source gets linked, but the brand is never named in the answer text.
- A separate 25.1% of appearances were the reverse: named in the answer with no link at all.
- The "AI runs on Reddit" story is engine-specific. One peer-reviewed measurement found user-generated content made up 14.2% of AI Overview references against 41.4% of the organic first page.
- Which domains get cited depends heavily on which engine you ask. There is no single source hierarchy to optimise for.
The instinctive model of an AI answer is that it reads a few pages, summarises them, and lists what it used. Reality is messier, and the mess is where the useful findings are.
Most citations do not say your name
This is the finding that reframes the rest. Semrush, working with Kevin Indig, analysed 3,981 domain appearances across 115 prompts, 14 countries and four engines — ChatGPT, Google AI Overviews, Gemini and Google AI Mode.1
Read that again: in only 13.2% of appearances did a brand get both a link and a namecheck. The most common outcome by a wide margin was being used as raw material — your page informing the answer while the reader learns nothing about who you are.
This has a direct consequence for reporting. If your dashboard counts brand mentions, it is blind to roughly six in ten of your appearances. If it counts citations, it misses the quarter where you were named with no link. Counting either alone understates you, and in different directions.
Which domains actually get cited
Two large studies, measuring different engines, and they tell a more complicated story than the headlines.
The peer-reviewed one comes from Xu, Iqbal and Montgomery at Washington University in St. Louis, who analysed 61,212 references across 7,583 Google AI Overviews collected in March and April 2026.2
| Hostname | Share of AI Overview references |
|---|---|
| youtube.com | 5.49% |
| en.wikipedia.org | 4.39% |
| facebook.com | 3.68% |
| instagram.com | 3.65% |
| usatoday.com | 2.80% |
Notice how flat that distribution is. The most-cited domain on the internet’s largest AI answer surface accounts for about one reference in twenty. There is no small club to break into; the long tail is genuinely long.
Peec AI’s analysis of 30 million sources across five engines found the mix shifts depending on which engine you ask: Search Engine Land’s coverage described ChatGPT as favouring Wikipedia, Reddit and editorial sites such as Forbes, while Perplexity leaned toward Reddit, LinkedIn and G2.3 That study is where most people encountered the “AI cites Reddit and YouTube most” framing.
The Reddit story, corrected
That framing is not wrong, but it is being generalised past what the data supports.
Xu and colleagues measured how much user-generated content appears in AI Overviews compared with the organic results shown on the same page. UGC made up 14.2% of AI Overview references against 41.4% of first-page organic results.2
The popular claim
"AI answers run on Reddit. Get your brand into forum threads and you win AI search."
What the measurements support
Some engines lean heavily on community content — Perplexity most of all. Google's AI Overviews use proportionally less user-generated content than the organic results beside them, and cite domains rated more credible than those results.
The same study found AI Overview–cited domains scored higher on a composite credibility measure than the first-page results displayed alongside them: 0.732 against 0.645.2 Whatever Google is doing in that surface, it is filtering toward established sources, not away from them.
So “go and get cited on Reddit” is reasonable advice for Perplexity visibility and poor advice if your buyers use AI Overviews. Which is the general shape of nearly every tactic in this field: correct somewhere, wrong elsewhere, and useless as a universal rule.
How many sources an answer uses
Pew Research Center’s browser-panel study logged 68,879 Google searches from 900 US adults in March 2025 and looked at the AI summaries returned. Summaries were typically 67 words long, and 88% cited three or more sources.4
Sixty-seven words, three or more sources. That is the competitive shape of the surface: a very short answer assembled from several places, where each contributor gets a sliver of a sentence. It explains why single, self-contained, quotable claims do disproportionately well — there is not room for anything longer.
Numbers we deliberately left out
Several widely circulated citation rankings — including some naming precise per-brand citation shares — come from publishers who disclose in their own methodology that the figures are modelled estimates rather than observed engine output. We also found a claim that content updated within 30 days earns 3.2× more citations; we could not locate the underlying study, its sample or its method, so it is not here.
What to do with this
- Track citations and mentions as separate series Given only 13.2% of appearances are both, one number cannot represent your presence. Two can.
- Find out which engine your buyers use before choosing tactics The Reddit play and the Wikipedia play and the editorial-PR play are each right for a different surface. Pick after you know, not before.
- Audit the sources cited around you, not just your own Your competitor's citations are a list of the pages shaping your category's answers. Most of them will not be your competitor's own site.
- Give up on a source shortlist With the top domain under 6% of references, there is no set of ten places to be. Breadth of credible third-party presence beats a placement strategy.
Questions we get asked
Is a ghost citation worth anything?
Directly, very little — the reader does not learn your name and rarely clicks. Indirectly it tells you the model considers your page a usable source on that question, which is the precondition for being named later. Treat it as a leading indicator, not an outcome.
Should we be publishing on Wikipedia?
You should not be editing your own Wikipedia entry; that is against their policy and it tends to end badly. The finding that ChatGPT leans on Wikipedia is an argument for being the kind of organisation that gets covered by sources Wikipedia cites, which is a much slower game.
Do these distributions hold outside the US?
The Semrush study covered 14 countries; the AI Overview measurement study is US-focused. We would not assume the domain mix transfers — local media, local forums and language-specific reference sites almost certainly shift it.
How stable are these numbers?
Not very. The studies here were run between March 2025 and June 2026 and disagree with each other in places. Anything in this field older than about a year should be read as history rather than as a current benchmark.
The short of it
Being used and being credited are different things, and the gap between them is most of your appearances. Which sources feed an answer depends on which engine produced it. And the citation pool is far wider than the “get on Reddit” advice implies.
Measure both events, per engine, and be suspicious of any tactic that claims to work everywhere.
Sources
- Semrush (Margarita Loktionova, with Kevin Indig), "The Ghost Citations Study", 9 June 2026. 3,981 domain appearances, 115 prompts, 14 countries, four engines.
- Xu, Iqbal & Montgomery (Washington University in St. Louis), "Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact", arXiv 2605.14021, 13 May 2026. 55,393 queries, 7,583 AI Overviews, 61,212 references.
- Peec AI, reported in Search Engine Land, "AI search engines cite Reddit, YouTube, and LinkedIn most: Study", 31 March 2026, covering Peec AI's analysis of 30 million sources across five engines.
- Pew Research Center, "Google users are less likely to click on links when an AI summary appears in the results", 22 July 2025. 900 US adults, 68,879 queries logged March 2025.