AI Citation Data From 45 Businesses: YouTube Is the Most-Cited Source in AI Answers

We tracked AI citations for 45 businesses every week for 11 weeks. YouTube turned out to be the single most-cited source in AI answers across the entire dataset — and we were winning 1.5% of the slots available to us. Here is the full data set, the measurement trap that nearly hid it, and what we changed.

The short version

Between late June and early September 2026, our system ran weekly AI visibility snapshots across 45 client and in-house domains, probing Google AI Overviews, ChatGPT, and Perplexity with real prompts and recording every source each engine cited. That produced roughly 1,200 prompt-and-answer records per week.

Five findings worth your time:

• Consistent content work does move AI citations. Our own domain went from 48 AI Overview citations to 102 over 11 weeks — a 113% increase — while citation share rose from 38.7% to 44.4%.
• YouTube is the most-cited domain in AI answers, by a wide margin. At the most recent snapshot it accounted for 400 citations across the fleet. Reddit was second at 229. Facebook was third at 125.
• Almost nobody is claiming those video slots. In a three-week window, 324 prompts returned an answer citing a YouTube video. Five of those videos were ours. That is a 1.5% capture rate on the highest-volume citation source in AI search.
• Video citations, once won, are unusually durable. One of our videos held a Perplexity citation for three consecutive weekly snapshots. A YouTube Short — not a long-form video — earned a Google AI Overview citation.
• Most AI visibility tools will tell a local business it has zero citations when it does not. We had a client reading zero on a commercial data provider while our own live prompt probes showed its domain cited 18 times in a single week. More on this below, because it is the most consequential thing in this article.

What counts as an “AI citation,” exactly?

The terminology in this space is a mess, and the mess causes real reporting errors. Three distinct things get called the same name:

• Citation. The AI answer links to a specific URL on your domain as a source. This is the one that sends traffic and the one that is unambiguously measurable.
• Mention. The AI answer names your brand in the text without linking to you. No traffic, but it means the model treats you as a known entity in the category — which is arguably the more valuable long-term asset.
• Visibility share. Your citations as a percentage of all citations across a defined prompt set. Useful for tracking direction, meaningless without knowing the prompt set.

These move independently. Our own domain, for example, is cited far more often than it is named — AI engines happily scrape our pages as source material while rarely presenting us as an authority by name. Those are two different problems requiring two different fixes, and a dashboard that blends them into one “AI visibility score” will hide both.

How we measure it

We run two instruments in parallel, deliberately.

Instrument one: a commercial mentions index. This queries a provider that has indexed hundreds of millions of AI responses and reports how often a domain appears as a cited source or a named brand. It is cheap, fast, and gives a comparable number across a whole client roster. A full-fleet weekly pull across 39 domains costs us under four dollars.

Instrument two: live prompt probes. This is our own harness. For each client we maintain a prompt set — roughly 8 to 34 prompts per engine — written to mirror how a real buyer would ask. We fire them at Google AI Overviews, ChatGPT, and Perplexity, capture the full answer text and every cited URL, then evaluate whether the brand was named, whether its domain was cited, and which competitors took the slots instead.

Running both is more work. It is also the only reason we caught the error described later in this article.

11 weeks of data: what actually moved

Here is our own domain’s trajectory, one row per weekly snapshot, measuring Google AI Overview citations where our domain was the cited source:

Two things are happening at once, and it is worth separating them. Citations grew 113%. But AI search volume for the prompts where we appear also grew 113%, from 6,300 to 13,390. Part of our gain is us getting better. Part of it is the pond getting bigger — AI Overviews are appearing on more queries in our category than they did in June. Anyone showing you an AI visibility chart that goes up and to the right without decomposing those two effects is selling you the tide as if it were the boat.

Citation share is the metric that controls for this, because it measures our citations against everyone else’s on the same prompt set. Share moved from 38.7% to 44.4% — a real gain, but a far more modest one than the raw citation count implies. That is the honest number.

The fleet picture

Across all tracked domains over the same window:

Five domains crossed from zero to a non-zero citation count during the window. One regressed. And a meaningful group — including several local service businesses — sat flat at zero for the entire 11 weeks according to the commercial index.

That last group is where things got interesting, and where we found we had been reading our own data wrong.

Who actually wins AI citations

Because our prompt harness stores every cited URL, we can aggregate across the whole fleet and see exactly which domains AI engines reach for. Here is the top of that list at the most recent snapshot:

Look at the composition. Of the top ten most-cited domains in AI answers across 45 businesses, exactly one is a reference work and none are traditional publishers. The list is video, forums, social platforms, and directories.

This should reframe how you think about the work. If your AI visibility strategy consists entirely of publishing blog posts on your own domain, you are competing for a minority of the available citation slots while ignoring the majority. Your website matters. It is not where most of the citations live.

The YouTube finding

YouTube is not just first on that list. It is first by 75% over second place, and it holds that position across every engine we probe.

So we asked the obvious follow-up: of the prompts where an AI engine cited a YouTube video, how many cited our video?

In a three-week window, 324 prompts across the fleet returned an answer that cited at least one YouTube video. Five of those citations were videos we own. Every one of the five was on our own domain’s prompt set — not a single client video has ever been cited in an AI answer in the entire dataset.

A 1.5% capture rate on the single highest-volume citation source in AI search. That is not a tuning problem. That is a strategy that never accounted for the channel.

The external research points the same direction. Ahrefs analyzed 75,000 brands and found that branded web mentions correlate with AI Overview visibility at 0.664 while backlinks sit at 0.218 — roughly three times the predictive strength. Their follow-up analysis extending to ChatGPT and Google AI Mode found YouTube mentions to be the single strongest individual signal at approximately 0.737. Their study also found that about 26% of brands have no AI Overview mentions at all.

Correlation is not causation, and Ahrefs says so plainly in their own write-up. Large brands have both more mentions and more AI visibility; the mentions may be a symptom of being a real brand rather than the cause of AI visibility. But when third-party correlational research and our own first-party citation counts independently put YouTube at the top of the list, the practical conclusion is the same either way.

Video citations are durable, and Shorts count

The five video citations we did earn taught us more than the 319 we missed.

Two videos were responsible for all of them. One held a Perplexity citation for a question about local search visibility across three consecutive weekly snapshots — it was still holding the slot at the most recent pull. The other was a YouTube Short answering a cost question, and it earned a Google AI Overview citation across three snapshots before dropping out.

Three things follow from that small sample:

• Video citations persist. Unlike a blog post that gets displaced as fresher content arrives, a cited video held its slot week after week. The citation behaved more like an asset than an impression.
• Shorts are citation-eligible. A vertical Short, produced in a fraction of the time of a long-form video, earned a citation in Google AI Overviews. The format that is cheapest to produce is not excluded.
• Question-shaped videos win. Both cited videos answer a specific question a buyer would actually type or say. Neither is a brand overview, a testimonial, or a service page in video form.

The measurement trap that nearly hid all of this

This is the section to read twice, especially if you are paying for an AI visibility report.

One of our restoration clients had been reading zero across the board on the commercial mentions index. Zero citations, zero mentions, zero citation share, for 11 straight weekly snapshots. We had written it up as the cleanest possible baseline — a business completely absent from AI answers, ideal for measuring the lift from a distribution campaign.

Then we cross-checked against our own live prompt probes for the same weeks. Its domain was cited 18 times: 11 in Perplexity, 5 in Google AI Overviews, 2 in ChatGPT.

Not zero. Cited in all three engines, in the same week the commercial tool reported total absence.

The explanation is straightforward once you see it. A commercial mentions index reports on the AI responses it has indexed. Its corpus is built from prompts with meaningful search volume. “24 hour water damage company Rogers Arkansas” does not have meaningful national search volume — it has a handful of desperate homeowners a month. So it is not in the index, and a business that gets cited exclusively on hyper-local prompts reads as invisible.

The practical consequences:

• If you are a local service business, a commercial AI visibility index will systematically understate you, often to zero. Your citations happen on prompts too small for the index to have sampled.
• The tool is not broken and it is not lying. It is answering a different question than the one you asked. It reports on a national corpus; you wanted to know about your city.
• Only live prompt probes answer the local question. If nobody is firing the actual prompts your actual buyers use, nobody knows your actual AI visibility.
• Beware the strategy built on a false zero. We were one approval away from spending money on a campaign to fix a problem that partly did not exist, and we would have measured its success against a baseline that was wrong.

We now treat the live probes as the source of truth for any business whose demand is geographically bounded, and use the commercial index for national-scale comparison only. Anything reporting a local client’s AI visibility off the commercial index alone is showing a number we no longer trust.

The structural gap: we were measuring what we were not acting on

Here is the part that was genuinely embarrassing to find.

Our prompt harness records, every week, exactly which prompts get answered without citing our client. At the latest snapshot that was 773 prompts across the fleet. Broken down by the intent we classify each prompt with:

That top row is the important one. 259 prompts where a business’s own core subject is being explained to a buyer, right now, by somebody else.

Meanwhile, on the other side of the same system, our content pipeline decides what to publish each day by pulling from a queue of strategy plays generated from keyword and ranking data. Those two things had no connection. The pipeline was producing content and hoping some of it earned citations, while a precise, weekly-refreshed list of 773 answers we were losing sat in a database view that nothing read.

We had built the eyes and never connected them to the hands.

What we changed

The fix is unglamorous and took an afternoon:

1. A candidate view. Takes the latest snapshot’s gap prompts, filters to active publishing clients, scores each one on how winnable it looks — core-topic and buyer-intent prompts rank highest, head terms lowest — and boosts anything showing the same gap in two or more engines, since a gap visible across engines is a real content absence rather than one model’s quirk.
2. Deduplication that actually works. Exact matching is not enough. “What’s the average cost of a metal roof in North Carolina?” and “What is the average cost of a metal roof in North Carolina?” are the same content brief. We use fuzzy text similarity to collapse near-duplicates, and we check against every play generated in the previous 45 days so the pipeline cannot be handed the same brief twice.
3. Injection into the existing queue. Surviving prompts become content plays in the same queue the pipeline already reads, carrying the prompt itself as the target and the gap evidence as the rationale — which engines showed the gap, on what date, at what intent. Nothing downstream had to change.
4. A daily schedule that runs 30 minutes ahead of the step that promotes plays to publication, so gap-driven briefs compete on equal footing with keyword-driven ones.

First live run: 51 briefs across 26 businesses, capped at two per business per run so no single client’s queue gets flooded. Every one of them is now a question we have watched an AI engine answer without our client, in the last seven days.

A few examples of what that list actually contains:

• A roofing company that does not appear when AI is asked what a metal roof costs in the two states it operates in — a question with obvious commercial intent, showing the gap in two engines.
• A restoration company absent from “who to call for fire damage restoration” in its own service region, across all three engines.
• A dumpster rental company missing from a straightforward sizing question its buyers ask constantly.

None of those are exotic. They are the most basic questions in each business’s category, and something else was answering them.

The levers, ranked by measured opportunity

Based on where the citation slots actually are in our data, here is the priority order we now work in:

1. Video on question-shaped topics. 400 citations at the last snapshot, 324 open slots in three weeks, 1.5% captured. Largest measured gap between opportunity and effort by a wide margin, and the citations persist once won.
2. Content aimed at known gaps rather than keyword volume. 259 core-topic prompts where the business is absent from its own subject. Cheapest to act on if you already have a content pipeline, because it is a targeting change rather than a new capability.
3. Forums and community platforms. Reddit at 229 citations and Quora at 40 mean roughly 78 gap slots sit on platforms where the answer is a genuine participant, not a page. This raises a real question about who does the participating — the agency or the business owner — that we think is a decision to make deliberately rather than default into.
4. Directories, reviews, and Google Business Profile. Angi, Expertise, BBB and Yelp together are the workhorses of local AI answers, and Google’s own properties account for 153 citations. Listing consistency and review velocity remain the cheapest AI citations available to a local business.
5. Social platforms. Facebook at 125 citations was the most surprising number in the dataset and the one we are least equipped to explain. We are treating it as a measured observation pending a proper look.

What to do with this if you run a business

Concretely, in order:

1. Write down 10 to 20 prompts a real buyer would use. Not keywords. Full questions, including the local and the desperate ones. “Who do I call when my basement floods in Asheville” is a prompt. “water damage restoration” is a keyword.
2. Ask the engines yourself. Run them through Google AI Overviews, ChatGPT, and Perplexity by hand. Record who gets cited. This takes an afternoon and is more accurate for a local business than most paid reports.
3. Do not trust a zero. If a tool says you have no AI visibility, verify by hand before building a strategy on it, especially if your customers are within driving distance.
4. Make a video for your three highest-intent questions. A Short counts. Answer the question directly in the first ten seconds. Based on our data this is the highest-yield unit of work available right now.
5. Fix your listings and reviews before writing more blog posts. Directories and Google properties are cited constantly in local answers and take a fraction of the effort of a content program.
6. Track share, not count. AI answer volume is growing in nearly every category. Raw citation growth will flatter you. Share against a fixed prompt set will not.

What we still do not know

In the interest of not overselling an 11-week dataset:

• We cannot yet connect citations to revenue. We can prove citations moved. We cannot prove they produced customers, and any agency claiming a clean line from AI citation to closed job right now is guessing.
• ChatGPT is moving the wrong way for us — 82 mentions in August, 66 in September on a monthly pull. We do not have a diagnosis. Monthly sampling gives slow feedback, which is its own problem.
• Facebook’s 125 citations are unexplained. We have not established what kind of Facebook content gets cited or whether it is reproducible.
• Correlation caveats apply throughout. Our clients receiving more content also tend to be the clients receiving more of everything. We are reporting what moved alongside the work, not proof the work caused it.
• Eleven weeks is a short series. Some of what looks like signal will be seasonality or model-update noise. We will know more in a quarter.

Frequently asked questions

What is an AI citation?

An AI citation is when an AI answer engine — Google AI Overviews, ChatGPT, Perplexity, Gemini — links to a specific page on your website as a source for its answer. It differs from an AI mention, where the engine names your brand in the text without linking to you.

How do you track AI citations for a local business?

By running the actual prompts your local buyers use directly against each engine on a fixed weekly schedule and recording every cited source. Commercial AI visibility indexes report on a national corpus of high-volume prompts and will frequently report zero for a local business that is in fact being cited on local prompts. We found one client reading zero on a commercial index while live probes showed 18 citations in the same week.

Which websites get cited most often in AI answers?

In our dataset of 45 businesses, YouTube was the most-cited domain with 400 citations at the latest snapshot, followed by Reddit at 229, Google properties at 153, and Facebook at 125. Directories including Angi, Expertise, BBB and Yelp accounted for most of the remainder in local queries. Traditional publishers were largely absent from the top of the list.

Does YouTube help with AI citations?

Our data says yes, and it is the largest untapped channel we measured. YouTube was the single most-cited domain across every engine we probe, and 324 prompts in a three-week window cited a YouTube video. Independent research from Ahrefs across 75,000 brands found YouTube mentions to be the strongest individual correlate of AI visibility at approximately 0.737, ahead of branded web mentions at 0.664 and well ahead of backlinks at 0.218.

Do YouTube Shorts get cited in AI answers?

Yes. One of the two videos cited in our dataset was a vertical Short, and it earned a Google AI Overview citation across three consecutive weekly snapshots. Short-form video is not excluded from AI citation.

How long do AI citations last?

Video citations in our data proved unusually durable — one video held its Perplexity citation across three consecutive weekly snapshots and was still holding at the most recent pull. We do not yet have enough data to state a reliable average lifespan for page citations.

Do backlinks help with AI citations?

Less than most SEO programs assume. The Ahrefs 75,000-brand study measured backlinks at 0.218 correlation with AI Overview visibility versus 0.664 for branded web mentions — roughly a third of the predictive strength. Backlinks still matter for traditional rankings. They should not be the headline metric in an AI visibility program.

How often should AI visibility be measured?

Weekly, against a fixed prompt set. Monthly sampling is too slow to catch a decline while you can still act on it — our own ChatGPT numbers dropped 20% between monthly pulls and we had no intermediate data to explain when or why.

The one-sentence takeaway

The citation slots in AI answers are mostly not on your website — they are on video, forums, and directories — and the fastest way to find out where you are losing them is to stop reading aggregate visibility scores and start asking the engines the questions your buyers actually ask.

All first-party figures in this article come from weekly AI visibility snapshots run across 45 client and in-house domains between June 23 and September 7, 2026, probing Google AI Overviews, ChatGPT, and Perplexity. Client domains are described by category rather than named. Third-party figures are attributed and linked.