We ran an experiment. We took a sample of real buyer and category questions, put each one to five AI engines, and logged every source each engine cited. That left us with 127,198 citations pointing at 11,647 different websites, across ChatGPT, Claude, Gemini, Perplexity, and Google AI Mode. The first thing the data showed: the five engines barely read the same web. Seven in ten sources were cited by only one of them. Just 2.7% were cited by all five.
What we measured
For this experiment we used a sample of about 16,400 AI answers collected between March 29 and June 27, 2026. Each answer came from putting a real buyer or category question to one of the five engines and recording the sources it cited. We then pulled the source list out of every answer and counted citations at the domain level. We looked only at the sources the engines cited, which are public websites they link to, not at the brands the questions were about.
Keep one thing in mind. The questions skew commercial: people checking how AI talks about their product or category, not general-knowledge trivia. So read this as how AI cites sources for buying and brand questions, the case that matters if you sell something. It is not a model of the whole web.
The five engines barely overlap
Of the 11,647 domains cited at least once, 69.6% were cited by exactly one engine. Only 2.7%, which is 309 domains, were cited by all five. Add in the domains cited by four engines and you still reach only about 7% that most engines agree on.
“Get cited by AI” quietly assumes there is one AI to optimize for. There is not. A page Perplexity loves can be invisible to ChatGPT. Optimizing for “AI search” as a single target means optimizing for an average that no engine actually is.
Gemini cites three times more than ChatGPT
The engines do not even cite the same amount. Gemini listed 11.0 sources per answer on average; ChatGPT listed 3.7. Perplexity (8.6), Google AI Mode (7.8), and Claude (6.8) sat in between.
More sources per answer means more chances for any given page to make the cut. On Gemini, a single answer can name ten sites or more; on ChatGPT, it is often three or four. A page that is the eighth-best match has a real shot on Gemini and almost none on ChatGPT.
Each engine has a citation personality
The clearest split was over user-generated platforms. Google AI Mode sent 11.2% of its citations to YouTube and 4.0% to Reddit. Perplexity also leaned on YouTube, at 8.8%. At the other extreme, Claude sent 0.02% of its citations to YouTube and 0.01% to Reddit. It almost never cites either.
That surprised us. Claude is not citing fewer sources because it is lazy; it lists 6.8 per answer. It just pulls them from documentation, vendor pages, and editorial sites rather than forums and video. If your buyers ask Claude, a great YouTube presence does close to nothing. If they ask Google AI Mode, it may be the whole game.
Reddit and Wikipedia are not running this
Widely shared studies, like 5WPR’s index of 680 million citations, put Reddit and Wikipedia near the top of the AI source list. In our commercial sample, neither was close. Reddit was 1.8% of all citations and Wikipedia was under 0.6%. YouTube, at 4.9%, was the only big consumer platform that showed up strongly.
Where did the other 90% go? To vendor docs, product pages, and a very long tail of category sites. Of the 11,647 domains, the top 10 accounted for 20.6% of citations and the top 100 for 42%. Nearly 43% of cited domains were cited exactly once.
| Source type | Share of all citations |
|---|---|
| Vendor, product, and long-tail sites | 90.6% |
| YouTube | 4.9% |
| 1.8% | |
| GitHub | 0.7% |
| Wikipedia | 0.6% |
What we cannot claim
This is one experiment, and an honest one names its edges. We measured citations, not clicks or sales; a cited domain is not proof anyone visited it. The sample is commercial-intent questions, so the source mix would look different for health, news, or general trivia. Engine behavior also moves: these numbers describe March to June 2026, and the platforms change what they cite often. And we counted at the domain level, so a brand with ten cited pages and one with a single cited page both count once.
What to do with this
A few things follow directly from the data.
- Stop optimizing for “AI.” Pick the engines your buyers actually use and check them separately. With 2.7% overlap, winning one barely carries to the others.
- Match the source type to the engine. YouTube and Reddit move Google AI Mode and Perplexity; documentation and vendor pages move Claude and ChatGPT.
- For commercial questions, do not chase Reddit and Wikipedia. They are a rounding error here. Clear product pages, docs, and category coverage do more.
- Mind the long tail. Almost half of cited domains were cited once, so a specific page on your exact topic can get pulled in where one broad post never would.
What we keep coming back to
The takeaway is the one we keep coming back to. “Get cited by AI” is the wrong goal. “Get cited by the engine your buyers actually use” is the real one, and on these numbers, that is five separate jobs. The average across engines hides the thing that matters, which is the same reason a single AI answer is not a visibility report: you have to read each engine on its own.
