How Many YouTube Videos Drive AI Citations? What 100 Million Citations Actually Show

0:00 YouTube Citations What 100 Million Show
Hey, I’m Jeremy Ashburn with PushLeads and helping small businesses get more leads without struggle or frustration. And we’re talking about a topic today about uploading videos to You– Now, the actual official, um, data that I’m seeing here is about twenty-nine percent YouTube, um, of the AI equation for showing up in Google for AI overview is posting on YouTube. So if you are a local business and you’re not posting on YouTube, a lot of us aren’t. I wasn’t until about literally three or four weeks ago, uh, because my videos would get zero views. But guess what? It doesn’t matter about zero views. It matters that you’re posting consistently as a blog on your website, as YouTube, um, on YouTube, etcetera. So we’re gonna look here in this little video about how many video actually need to get more visibility, um, and grab some of that AI, um, slice of the, of the… So let’s look at the number that we already have. We’ve got about three hundred six videos, um, on, um, on PushLeads. You can go to, you know, you’re probably on YouTube right now, or if you’re on LinkedIn or somewhere else, you can go to YouTube and see it. Um, and only two of these videos have only been, ever been cited in, um, an AI answer. And so, um, look, you know, we have a lot of content, and it didn’t make a difference. However, though, in the last sixteen days, we’ve been doing, uh, content that’s laser-targeted. So we did a whole bunch of research and figured out what are people searching for in AI, and then we started making content answering the common questions that people search for on AI. And one of the kinda companies I go out– I work with is restoration businesses. So water damage, fire damage, mold damage, you name it. So I did a lot of, you know, content around restoration marketing, restoration SEO, etcetera. I, I have a restoration company I’ve taken from four fifty to six hundred and fifty K in revenue, and so, um, so super interesting.

1:51 Channel Volume Does Not Drive AI Citations
Um, but in the last sixteen days, we basically, you know, posted forty-three shorts, and we also posted a long, some long videos. And so we’re gonna look at the data here. So let’s take a look here. So the first question is, does channel volume dot drive… So the first question here is, does channel volume drive AI citations? The short answer is no. There’s no published study that, um, establishes a view count threshold, and so essentially we have a correlation data of zero. The evidence base is from Otterly AI, March of twenty twenty-six, um, where they looked at a hundred million AI citation instances over thirty days across ChatGPT, Google’s AI reviews, AI Mode, Perplexity, Copilot, and Gemini. So many AIs, so little time . But what replaces volume is a per video answer fit. So what that basically means in plain English is making videos answering questions that your customers are asking or what are people actually searching for. And you have to have an extractable transcript. You have to have chapter structure. You have to have a rich description and regency, so you’re posting frequently. And when you’re posting, the unit is the video itself. Now, for bonus points, you should also have the video embedded on a page or blog post on your site. So if you are a local business and your SEO person isn’t talking about video and they’re not posting the video embedded in the site, and you might wanna consider working with somebody who actually is doing, is jumping into the whole AI game. A lot of ACE– SEO agencies are very old school and haven’t, you know, made the switch or they’re just relying on tools. They’re actually not doing the work. And I honestly can’t stand that, you know, not doing the work. But I’m gonna get off my high horse. So let’s look at the correlation with citation frequency.

This is part of our AI search visibility guide.

3:38 Correlation Factors Description Recency Hashtags
So Otterly AI looked at a hundred million instances in March of twenty twenty-six, and what they found here is that description length, uh, makes a difference, and that was one of the strongest leathers, uh, and that was one of the strongest levers, uh, regency. If you’re publishing fresh, that accounts for point, um, um, a third of the equation. If you have hashtags, that’s like, you know, point, that’s like point two of the equation. The duration and the likes doesn’t make a difference. The view channel count, the subscribers of video views, none of that makes a difference. So everything that looks like clout, which is sorta like subscribers, views, likes, library size, sits at or below zero. So what does that mean? That basically bottom line is if you post videos answering questions and you give value, ’cause AIs love when you give value to the web. Now, not really salesy kinda content. Not like, "I’m the best SEO company in the world." AIs don’t care about content like that. They care about actual data, right? So AIs are like Spock , you know, where it’s like, "Give me the data." That was a really bad Spock, um, imitation. But anyway, so you need to have post the video. You need to have a description. You need to have, be posting frequently, and you need to have hashtags. That’s sort of like what we’re seeing here. Um, and the crazy thing is AI is kinda leveling the, the playing field here. So small channels are getting cited. Um, the distribution of AI-cited videos in their channels, about forty percent of the cited videos had under a thousand views. And this is where I keep saying it doesn’t matter that you post a video that has zero views or has one or five. That’s, that’s one of the reasons why I stopped posting on YouTube. I’m just like, "What’s the point here, um, if you don’t post shorts?" Even I– when I was posting shorts, I wasn’t getting a lot. And it doesn’t matter. It’s that Google wants to see that you’re active and that YouTube is one of the channels that AIs can easily see that you’re posting on. And it’s the easiest, one of the easiest channels for AIs to capture. You don’t have to be logged into YouTube just to get, you know, the data, whereas Facebook has multiple layers and blah, blah, blah. So thirty-six percent of the cited videos were, had, were under like fifteen likes, and then thirty-five percent of the cited channels had under ten thousand subscribers. So the median channel, the median cited channel had about forty-one videos, and our

5:58 Video Rich Result Checklist Schema Captions
channel has three hundred and twenty-six. So being eight times larger than the median channel brought us two citations, but not because of volume, because of our targeting. So what actually earns a citation? So bottom line here is that title If you have a title, a transcript, and a good description along with a video where, of course, you’re talking, right? It can’t just be a slideshow video with no audio, um, then that is the– that’s the winner. Winner, winner, chicken dinner, right? So, um, if the titles mention the query, only two of our titles did this, it was the strongest signal we observed. It was a– we have– when you have a clean, corrected transcript, so up– creator, creator uploaded captions run ninety-nine percent accurate, versus if you just let Google sort of do the transcript, then it’s like eighty-five to ninety-five percent accurate. So Google and AIs like accuracy. And then chapters in the real description. Seventy-fi– seventy-eight percent of the cited videos were cited more than once, and they had– and, and they had chapters in the descriptions. So chapters are kind of behaving like an H2 in a blog post. So super critical. Um, so here’s the share of YouTube citations by engine. Perplexity is, um, is thirty-eight percent, um, AI Overviews is thirty-six percent. Google’s AI mode is nineteen percent. ChatGPT is like four percent. Gemini, strangely enough, being a Google product, only is– I think co- Copilot is half a percent. Um, so basically, YouTube is twenty-nine point five percent of all AI overview according to Bright Led– Bright Edge study with thirty million sources, and YouTube is cited two hundred times more than any other platform. So bottom line here is that if you are an SEO agency, you need to be getting your clients to be posting on YouTube, or you need to come up with a strategy of getting content, which I’ve already solved that problem, you know, so that’s what happens with Builder. Um, but anyway, so, uh, this is pretty wild though. It’s all about YouTube, um, and of course, if you wanna show up for ChatGPT, you need to be posting on Reddit ’cause they sort of highlight more of the Reddit. So here’s the next question. How many videos do you need for regular, uh, to show up in regular Google searches? So the volume here is irrelevant too. A properly structured video can earn a, a rich result. A thousand bad, badly structures embeds lear- earns… A thousand bad… A thousand badly structured embeds learns– earns nothing. A thousand badly structured embeds… A thousand badly structured embeds learns nothing. A thousand badly structured embeds… A thousand badly structured embeds earn nothing. And the twenty twenty-two golden rule, the twenty twenty-throu… and then the twenty twenty-three rule still governs. Google shows an AI thumbnu… Google shows an AI thumbnail only when the video is the main content of the page. Blog posts with supporting video no longer qualify, and it’s all per video work. The schema, the real watch page, the caption, the chapters, the thumbnail, all the video thumbnet, vide- the video sitemap, all of that has to be done, and that’s just all part of your checklist. So here’s your rich video… So here’s your video rich result checklist. Give the video its own watch page. So the page need to be… So when you’re publishing content, the page needs to be on the site. It needs to be an inner hidden page or a blog post. Add the video object schema to the page, um, add the video object schema to the page. The JSON-LD works there, and this is what the made– and this is what makes the page eligible for a thumbnail and duration badge. Add a clip or seek to action markup. This produces key moments in the SERP, and upload the corrected captions, not the auto ASR. We know that doesn’t work as well. And this feeds Google and every other AI engine. And you need to have a unique thumbnail per video. Google says specifically don’t reuse thumbnails. They don’t like that. And then submit it, and then submit a video sitemap via Search Console. You could submit up to fifty thousand URLs. So if you’re doing this over and over again or you make a system that does this, you’re going to be successful.

10:07 Shorts vs Long Form Citation Rates
And then what about shorts versus longs? Um, short-term videos versus long-term videos. Shorts are only accounting for about five point seven percent of citations. They’re really too short to yield a usable co, um, transcript. It’s really only cited inside Google services, and all forty-three percent of our shorts are near useless. They didn’t show up. Oh man, it brought people to the channel, but it’s useless. Um, so it’s still fine for the reach and the brand. The long form content, the longs, is ninety-four percent of citations. It has enough spoken con, um, content to extract from. If you notice, this video is about ten or twelve minutes. I’m really targeting for videos between twelve to fifteen minutes. Um, you have to have chapters, which, which drive citation. The largest cited cluster is like, you know, on average ten to twenty minutes. So twelve to fifteen minutes is kinda your go-to. And then fifty percent of cited videos were under eight minutes. What about embedding on the site?

11:04 Embedding Video on Site GSC Data
Does this help? Yes, for probably engagement. So Wistia’s f- f- so Wistia’s five hundred thousand session data on shorter time– on longer time is page credible. So dwell time is not really a confirmed ranking signal. But yes, it helps for click-through rate. Me- the– and the mechanism is clear. Bright Edge measured roughly thirty percent click-through rate from video object schema thumbnails across twenty-five thousand pages, and they found that a thumbnail draws the eye. And of course, the famous stat is junk. Fifty-three percent more likely to hit page one is going back to a twenty- two thousand and nine claim with no accessible meth- methodology. So our own data with thirty-nine pages, our average Google Search Console impressions, pre versus post, went from two hundred and twenty-seven to five hundred and fifty-seven, and then our average position went from twenty-seven to twenty-four. So a hundred and forty-five percent of keywords per page rose. They actually rose from thirteen to seventeen. So this is sort of us measuring. It’s not really enough data for us to say, "This is it." So there was no suc, su… So there was no su… There was no ch- control group, so we can’t separate the embed f- effect from normal content aging. And of course, pages accumulate impressions and over And of course, pages accumulate impressions and keywords over months, whether or not you add video. So better pages tend to be the ones that have the videos added to them. So anyway, it’s early on data, but that’s kinda what we’re seeing. So the honest, um, caveat here is that we measured a change, not the cause. Without a control group, content aging, and selection, selection bias explain an unknown share of every lift.

12:43 AI Citation Half-Life How Fast They Churn
So it’s AI citations are extremely volatile. We saw this ourself. We saw a video that showed up in AI overview in one day and was gone the next, replaced by a competitor’s video, and that’s the norm, not the bug. Uh, multiple independent studies converge on a citation’s half-life, measured in days and weeks, not months. So the change is what we’re reporting. We’re tracking citation share across a pump… a prompt. We’re tracking citation sca… We’re tracking citation s- we’re tracking citation share across a plump… across a prompt slunk… We’re tracking citation share across a prompt cluster over time, and we’re not reporting anything as a win yet. And of course, we have the citation half-life across studies, independent measurements of how, you know, how citations actually churn. Here’s a couple independent measurements of how fast AI citations churn. Sample one, basically we saw from tracker with looking at ten thousand brands and seven models, we saw a thirty-day half-life. For another sample, we saw seventy-three percent of URLs cited once, never cited again. Digital Authority looked at eleven hundred URLs, and ten point six were still cited after twenty-eight days. Get Mentioned looked at five thirty thousand… five hundred thirty hundred thousand citations, and sixty-nine percent of those sources changed within two days. And then Max AEO is tracking single URLs, and the median, and the median was six consecutive days cited. And then domain-level visibility outlasts URL visibility because the engines are swapping one page for another. So build the site, don’t build the assets. A couple more questions. How do engines read a video? They’re primarily looking at the transcript. They’re looking at the description, chapters, metadata. Google’s models also processing our, our… Google’s models are also processing the audio and structure. So say the target– So say the target question verbatim. So say this– So say the target question verbatim early and answer it inside the first forty to sixty questions. Should you republish transcripts on a page? Yes. Google permits a video plus its own transcription on the same page. The risk is using that over and over again across multiple pages. Google doesn’t like that. What about the ultimate… What about the length of the video? Roughly eight to fifteen minutes. Duration only– Duration correlates at only point… at only zero point two… Correl– Duration correlates verily… Duration barely correlates. So l-length is not the leather, so length is not the lever. Having enough substance to report real chapters is, and the largest chapter was, you know, at ten to twenty minutes. Um, and

15:31 Local Reality Check Maps vs AI Overviews
the local reality check is that Google AI overviews are an informational thing. Whitespark. Whitespark’s, Whitespark measured five hundred and forty manual queries. Ninety-two, ninety-two percent of… Ninety-two percent of informational local queries were triggering an AI overview and ninety-seven percent of hybrid intent ones. But the near me searches like plumber near me, roofer near me, water damage near me, they still belong to the map. So that’s on direct near me searches, and that’s showing up in the map ninety-three percent of the time. So transactional demand hasn’t… So transactional demand has not moved yet. But what happens when AI answers show up in the map results? It’s probably gonna happen at one point. So video wins the top of the funnel. It wins how much does water damage restoration cost? That was one of the ones that we actually were able to get cited on. Cited on the Google business, Google Business posting reviews, and the map pack still win the call. So you gotta be posting regularly, um, across your site. So what should you do?

16:33 Action Plan Retrofit Optimize Track
So what should you do next? Stop counting and start matching questions. One, create one video that ma– create one video that’s eight to fifteen minutes long that answers each question. One question per video, and you wanna question the… and you wanna title the question, and you wanna title that as the search. How much does water damage cost? Water damage costs X amount, blah, blah, blah. And you wanna retrofit… We’re gonna retrofit the su– the two hundred and sixty-six long-form videos we own with chapters, corrected captions, descr– three hundred plus word descriptions, and that’s essentially zero new production. It’ll produce the highest ROI. We’re gonna fix the embed technicals, the real iframe in the DOM, video object schema, video sitemap, and we’re gonna, we’re gonna click t– and we’re gonna kill click to inject facades. Build the sh– build the plump… We’re gonna build the prompt cluster tracker. We’re gonna be… We’re, we’re gonna track fifty to one hundred prompts daily across four engines. Report, report share to share over time, and we’re gonna run a match pair experiment. We’re gonna be publishing sixty to eighty paired pages, and some are embedded on site, and some are embedded on the site, some are not, measuring that over ninety to a hundred and eighty days, and we’re gonna look at differences. And then we’re gonna project the fun– we’re gonna protect the fundamentals, the Google My Business posting, the reviews on the map pack. Video is addictive, but not the replacement. So the real answer to how many is roughly one well-bid… one well-built video answering one question worth owning. For our, for our current data, we’re saying that that works. I’m Jeremy Ashburn with PushLeads, helping small businesses get more leads without struggle and frustration. Talk to you soon. Thanks.