A popular story about AI search says the engines run on Reddit and YouTube now, and that the way to get cited is to be all over both. We wanted to see where ChatGPT, Perplexity, Gemini, Claude, and Google AI Mode actually get the information in their answers, so we classified the sources each one cited by the kind of site it is. The open web still does most of the work, and the Reddit-and-YouTube version of the story is only true for particular engines.
How we measured this
We looked at about 151,000 sources that the five engines actually cited in their answers, drawn from roughly 23,000 answers to the same set of prompts run across all of them. The prompts are the kind buyers ask when they compare and choose in competitive categories, not a random crawl of the web, so treat the numbers as a benchmark sample and the pattern as directional. We grouped each cited domain into a type: general web and publishers, forums like Reddit and Quora, video that is almost entirely YouTube, reference and academic like Wikipedia and arXiv, SEO and marketing publications, review sites, LinkedIn, and blogs. Every percentage below is an engine’s share of its own cited sources.
The open web does the heavy lifting
On all five engines, general web and publisher pages are the overwhelming majority of citations, between 86 and 97 percent. None of them is mostly forums, mostly video, or mostly anything else. We expected the forum share to be larger, given how loudly Reddit gets discussed as an AI source, and it just is not, on any of the five.
Video shows up most on Google AI Mode
The sharpest difference between engines is video. Google AI Mode cited video, again almost all of it YouTube, in about 8 percent of its citations. That is more than any other engine and roughly ten times ChatGPT, which sat under 1 percent. Perplexity came second near 6 percent. Gemini, which is Google’s own model, used far less than AI Mode at around 2 percent. Putting a topic on YouTube for AI is really a Google AI Mode move, and we dug into that channel on its own in video is becoming AI visibility too.
Whether AI cites Reddit depends on the AI
Forums split the engines the same way. ChatGPT leaned on them most, around 5 percent of its citations, with Perplexity next near 3.5 percent. Claude barely touched them, a fraction of a percent. Reddit is a real signal for ChatGPT and Perplexity and almost invisible on Claude, which is worth knowing before you spend a quarter seeding threads for an audience that mostly reads Claude. On why Reddit shapes answers even when it sends little traffic, see AI search may use Reddit, but that does not mean Reddit gets the click.
ChatGPT reaches for references, Claude reads the mainstream
Two smaller patterns stood out. ChatGPT cited reference and academic sources, Wikipedia and arXiv and the like, more than any other engine, close to 4 percent, in line with its habit of reaching for something that looks authoritative. Claude was the opposite kind of outlier. Almost 97 percent of what it cited was general web and publisher pages, with the least forums and the least video of the five. It reads the mainstream web and mostly skips the community and video layers the others dip into.
| Engine | Forums | Video | Reference |
|---|---|---|---|
| ChatGPT | 4.9% | 0.8% | 3.7% |
| Perplexity | 3.5% | 6.1% | 1.7% |
| Gemini | 1.6% | 2.3% | 1.9% |
| Claude | 0.1% | 0.0% | 1.2% |
| Google AI Mode | 2.7% | 8.2% | 0.8% |
Why “Reddit is the top AI source” can be true and still misleading
You will see claims that Reddit is the single most-cited site in AI answers, and on some engines a single site like Reddit or Wikipedia genuinely does sit near the top of the individual-domain list. Both things are true at once. A source type stays small even when one site inside it ranks high, because the long tail of ordinary publisher domains, added together, dwarfs any single platform. Ranking individual domains and grouping source types answer different questions, and this post groups types.
What this does not show
The sample skews toward competitive commercial categories, so a corpus of medical or academic questions would almost certainly push the reference share up. The shares are pooled across the sample, so one very active category can pull a figure; the ordering, with Google AI Mode highest on video and Claude lowest on forums, is the robust part rather than the exact decimals. And a type based on the domain misses a real page that happens to sit on a general-web host, so “general web” is a catch-all rather than a clean category.
What to do with it
The split gives you a per-engine map instead of one plan. Video pays off mostly when Google AI Mode matters to you. Being present and well-regarded in the right forums helps on ChatGPT and Perplexity and does little on Claude. Across all five, the largest lever is unchanged: be a credible page on the open web the engines already trust, because that is where the clear majority of citations come from no matter which engine you look at. This is the next question after our earlier finding that the five engines barely cite the same sources: they do not read the same web, and now we can see they do not read the same kinds of sites either.
