83% of Google AI Overviews for How-To Questions Cite a YouTube Video (Study of 1,777 AI Overviews)

Finding 01YouTube is in most Google AI Overviews for how-to questions

Google showed an AI Overview for 99% of the 1,800 searches. Of those Google AI Overviews, 83% cited at least one YouTube video (95% confidence interval: 80% to 86%), and 4% cited no source at all.

By topic, the share of Google AI Overviews that cited YouTube was lowest for personal finance (65.1%) and Google Sheets (67.1%), and highest for home repair (95.7%), Excel (95.2%) and gardening (94.6%).

82.3% of the questions were how-to questions, 10.5% started with what or which, 4.3% asked how long, how much or how often, and 2.8% were something else, such as why or when questions.

The share of Google AI Overviews that cited YouTube was 86% for how-to questions, 67% for what or which questions and 67% for how long, how much or how often questions. For the 50 AI Overviews on other questions it was 90%, a small group.

Finding 02When Google links to a second, it quotes what is said there

36% of the 4,902 YouTube citations in Google AI Overviews linked to an exact second. Where the video had English captions, the words Google showed under the video were spoken within five seconds of that second in 97.5% of those citations. A second reader checked 60 cited seconds against the transcript without seeing the matcher's answers, with the matcher's misses oversampled. The matcher's precision was 97.5% and its weighted recall 99.6%, and it placed the start of the passage within three seconds of the reader in all 39 passages both found. The reader was a separate AI model (Claude), not a person.

Across all the YouTube citations, with or without a second, the words Google showed matched the captions alone in 80% of them, the captions and the description in 5%, and the description alone in 2%. Another 6.4% quoted a video's opening line, which the caption data lacks (see the limits below). Google's AI Overviews quoted what creators said far more often than what they wrote.

4% of the YouTube citations matched neither the captions nor the description. I asked an AI model to sort those by where the words came from, and a second reader agreed with it on only 32 of 60, so I do not use its labels. Every snippet they disagreed on read like the creator talking to the camera, in words my caption data did not have.

Where Google's quoted words were found near the linked second, I measured where they start relative to it. The median was -0.24 seconds, and the middle half of links fell between -1.08 and 0.28 (a negative number means the words start just before the linked second).

Links to an exact second were most common in gardening (58% of the YouTube citations) and personal finance (52%), and least common in Google Sheets (21%) and cooking (23%).

Finding 03The cited second is early

The median cited moment came 22 seconds into the video, and three in four came by second 59. Measured against the length of the video, the median cited moment sat 7% of the way through. 58% of the 558 cited moments came in the first thirty seconds.

Finding 04The cited moment, next to the rest of the same video

For each video with a cited moment, I compared that moment with the video's other windows. The caption data often misses a video's opening line, so I left out every window that starts in the first twenty-five seconds, on both sides. That leaves out the cited moments in the opening: this comparison covers the 230 videos whose cited moment came later. Each video counts once.

• The searcher's words. The cited moment held 44% of the question's words on average, against 31% for the other windows. It held more of them in 154 of the videos and fewer in 71.
• A chapter. A chapter from the video's description started inside 27% of the cited moments, against 12% of the other windows.
• An answer. An AI model read each window without knowing which one was cited, and judged whether it answered the question. In the 230 videos, it judged 89% of the cited moments to answer the question, against 82% of the other windows. A second reader labelled 100 of these moments, drawn at random, without seeing the model's labels. Weighted to all 5,087 labels, the model agreed with it 92% of the time (95% range 81% to 97%), so treat these two shares as approximate. The reader was a separate AI model (Claude), not a person. A first version of the model's instructions fell just short of the 90% this check requires, because it counted only complete answers; the figures here come from the second version, checked on a fresh sample.
• An announcement. A line that says what the video will show, such as "in this video" or "let me show you", was in 9% of the cited moments and 2% of the other windows.
• Words a second. The cited moments had more words in them: 3.22 words a second, against 2.99 for the other windows. They had more in 156 of the videos and fewer in 73.
• Numbers. A number was said in 23% of the cited moments and 23% of the other windows.

Finding 05Cited videos are smaller and newer than the videos YouTube ranks

For every question whose AI Overviews cited a video, I compared the cited videos with the videos in the top ten of YouTube's results for the same question that the AI Overviews did not cite. A video cited for two questions counts once for each.

• Views. The median cited video had 31,226 views, against 156,854 for YouTube's top results. Question by question, the cited videos had fewer views on average in 408 of 506 questions. Leaving Shorts out on both sides, they had fewer views in 379 of 475 questions.
• Channel size. The median cited video came from a channel with 83,200 subscribers, against 164,000 for YouTube's top results.
• Age. The median cited video was 589 days old when I collected it, against 1,272 days. Question by question, the cited videos were newer on average in 445 of 506 questions, and in 401 of 475 with Shorts left out. 16% of the cited videos were under six months old, against 5% of YouTube's top results, and 10% were over five years old, against 34%.
• Length. 18% of the cited videos were a minute long or less, against 9% of YouTube's top results. Between one and three minutes the shares were close: 33% against 34%.
• Likes. The median cited video had 0.8% as many likes as views, against 0.9% for YouTube's top results, and question by question the cited videos had fewer likes per view on average in 306 of 506 questions.
• Titles. The median title held 67% of the question's words for the cited videos and 67% for YouTube's top results. Question by question, the cited titles held more of them in 242 of 506 questions and fewer in 198.
• Chapters and captions. 25% of the cited videos had chapters, against 23% of YouTube's top results, and 20% had English captions that were not auto-generated, against 22%.
• YouTube's own ranking. Only 38% of the videos cited for a question appeared anywhere in the first twenty results of the YouTube search for it.
• Channels. The 4,902 YouTube citations went to 918 channels, and half of them went to 161 of those channels.

Finding 06Shorts get cited, and the citations hold

22% of the YouTube citations in Google AI Overviews linked to a YouTube Short. I do not compare that with YouTube's search results, which listed almost no Shorts on desktop: 0.5% of the top YouTube results I compared were Shorts.

I ran every Google search three times, an hour apart. 78% of the question and video pairs cited in Google AI Overviews were cited in all three runs, and for the median question the set of videos Google AI Overviews cited did not change from one run to the next (median Jaccard overlap 1.00).

The second held too: 75% of the question, video and second combinations Google AI Overviews linked to came back in all three runs, nearly as many as for the videos themselves.

Finding 07At the cited second, more lead-ins and fewer answers than nearby

At the cited seconds, 50% of the lines gave the answer, 45% set it up and 5% were something else. Set against the seconds around them in the same 542 videos, the cited second set up the answer 46% of the time against 13%, and gave it 49% of the time against 80%. The lines around a cited second were mostly the answer itself; the cited line was more often the one that leads into it. A second reader sorted 100 of these lines, drawn at random from each of the model's three labels, without seeing the model's labels. Weighted to all 2,213 lines, the model agreed with it 95% of the time (95% range 82% to 99%). The reader was a separate AI model (Claude), not a person.

Chapters did not mark the second. 32% of the cited seconds were in videos with chapters. In those videos, a chapter other than the opening one started within two seconds of the cited second 10% of the time, against 10% for the random seconds near it. Finding 4 shows a chapter starting inside the cited moment more often than inside the rest of the video; the link itself is not placed on the chapter's timestamp.

Finding 08The AI Overview's own text does not copy the video

Finding 2 shows that the snippet under a cited video is the creator's own words. The AI Overview's answer is another matter. Leaving out the question's words, the AI Overviews shared 28% of the cited moment's words, against 26% of the words in the other stretches of the same video (230 videos; the opening left out, as in Finding 4). A run of six of the creator's words in a row appeared for 2.6% of the cited moments and 1.4% of the other stretches. The AI Overview writes its own answer; the words it shares with a video are about as likely to come from anywhere in it as from the moment it links to.

Finding 09The videos on the rest of Google's page

Google showed a video carousel on 84% of the results pages with an AI Overview, and a row of Shorts on 32% of them.

46% of the YouTube citations in Google AI Overviews went to a video that also appeared elsewhere on the same results page: in the video carousel for 33%, the row of Shorts for 10% and the ordinary results for 7%. Set against YouTube's own ranking, 21% of the YouTube citations went to a video that was both on the Google page and in YouTube's first twenty results for the question, 25% only on the Google page, 18% only in YouTube's first twenty, and 36% in neither.

Seen from the carousel's side, the AI Overview cited 31% of the YouTube videos in it: 36.6% of those in the first position, 30.8% in the second, 27.9% in the third and 27.6% in the fourth. It cited 16% of the YouTube videos in the row of Shorts and 26.6% of the YouTube videos among the ordinary results.