YouTube was the most cited domain we saw in Google AI Overviews, at 10.4%
If you search for whether you need video to show up in AI answers, you'll get a consistent answer: yes, start a YouTube channel, Google gives its own platform a leg up. The first part of that holds in our data. What follows from it doesn't.
Between 22 August and 30 September 2026 we collected answers for questions we track, across two engines that return citations: Google AI Overviews and our ChatGPT engine, which queries OpenAI's API with web search available. Three brands in adjacent B2B software categories, two accounts. It's one narrow slice, so read every number below as ours and not as a law.
- Google AI Overviews: 612 of 891 completed answers carried citations, 5,139 citations in all. youtube.com had 533 of them, 10.4%, the single most cited domain. linkedin.com was second at 195, reddit.com third at 151.
- A YouTube link turned up in 333 of the 612 cited answers. Slightly more than half.
- ChatGPT engine: 206 of 958 answers carried citations, 1,484 citations. youtube.com had 10, which is 0.7%, in 8 answers.
The gap between the two engines is about fifteen times, on citation share
10.4% against 0.7%. BrightEdge published the same direction from a much bigger corpus, putting Google at roughly 30 times ChatGPT on absolute query volume. We're measuring a different thing, share of the citations an engine actually emitted, and we get about 15 times. Both readings say the same thing to a buyer: this is Google behaviour, and treating it as general advice about AI search is a mistake.
We're also honest that 10 citations tells you almost nothing about ChatGPT beyond the fact that YouTube was rare there in our sample. Don't build a plan on that number.
59% of the distinct YouTube links skipped to a moment inside the video
This is the part nobody seems to have counted. Our 533 citations resolved to 272 distinct URLs and 228 distinct videos. 161 of those 272 URLs carried a t= parameter, a deep link to one second inside the video, and those accounted for 303 of the 533 citations.
The median was 75 seconds. 76 of the 161 pointed somewhere in the first minute. The deepest was 13 minutes. In the raw payloads the links sit inline in the answer text, next to the particular sentence they support, which is what a timestamp is for.
Google documents the mechanism. Search tries to detect a video's segments automatically, and where you supply your own they get priority, either through structured data or the YouTube description. YouTube's own help page gives the spec for the description: start the first timestamp at 00:00, list at least three in ascending order, keep each chapter to 10 seconds or more. Automatic chapters are on by default for new uploads and you can turn them off.
The unit being cited, then, is a segment of a video. That changes what a useful video looks like here. A rambling 20 minute walkthrough with one good answer at 14:30 does worse than a five minute video that answers the question at 0:40 and says so in the description.
11 of the 533 citations were on the channel of the brand we were measuring
2.1%. We matched the cited video's channel name against the brand whose question produced the answer, so this undercounts a channel named something other than the company. Even allowing for that, roughly 98% of the video Google cited in these answers belonged to somebody else.
The 228 videos came from 193 different channels, and 171 of those channels were cited for exactly one video. The busiest single channel held 36 citations, under 7% of the total. There's no big video to get into. There are a hundred small ones, mostly made by people who don't work for any of the vendors: independent recruiters, trainers, consultants reviewing tools they use.
Two more things worth knowing before you plan a shoot. Only 42 of the 228 videos had a title shaped like a roundup or comparison, and they carried 22% of the citations, so the standard advice to make review and versus content points at about a fifth of what got cited. And recency mattered: the median cited video was 272 days old, and 64% of the citations came from videos published inside the last year.
What we'd actually do
In order, cheapest first.
- Find out which videos are already being cited for your questions before you make one. That list is short, specific, and usually a surprise.
- For the videos you can influence, the ones where a partner, customer or friendly reviewer already featured you, ask for a chapter marker on the segment that mentions you. That's a two minute edit to a description.
- If you do publish your own, put the answerable claim inside the first 90 seconds and mark it in the description. 47% of the timestamps we saw landed in the first minute.
- Don't do any of this for ChatGPT. In our sample it barely looked at YouTube.
Where this could be wrong
Three brands and two accounts is a small sample, drawn from adjacent B2B software categories. A consumer or healthcare category could look completely different, and published per-industry breakdowns suggest they do. The widely quoted figure of nearly 30% of AI Overview citations going to YouTube comes from a far bigger corpus than ours. We got 10.4%, and we think that difference is the point: the headline number doesn't transfer to your category.
The AI Overview references reach us through DataForSEO's ai_overview element and we store the reference URL as given, with no rewriting. We haven't independently re-rendered those searches in a browser to confirm Google's own markup carries the same timestamp, so the honest claim is that the timestamps are in the payload we're served.
Only two of the engines we run produce citation rows today, so Perplexity, Claude and Gemini are absent from every count here. The roundup test is a keyword match on video titles, not a viewing of 228 videos. And a share is not a ranking factor: we measured what got cited, which doesn't prove that adding chapters causes a citation.