YouTube as a Citation Source in AI Answers
What a February 2026 citation sample says about YouTube across AI engines, and why views, age, and citation share need careful interpretation.
YouTube can be a source in an AI answer, but its visibility is uneven. The engine matters. So does the prompt, the period measured, and the denominator behind the percentage. A video cited by Perplexity is evidence that Perplexity retrieved it for that answer. It is not evidence that the same video will appear in ChatGPT or Google AI Overviews.
Promptwatch's YouTube citation study covers February 1 through February 28, 2026. During that window, YouTube accounted for 2.74% of citations in Perplexity, 2.25% in Google AI Overviews, and 0.05% in ChatGPT. The gap is large enough that a combined "AI citation rate" would be a poor planning number.
The study used citations collected from the real interfaces of the monitored platforms. Its page-wide citation shares come from a broader tracking set, while the video attribute panels examine samples of cited videos and channels. Those are observational samples. They describe what was cited in the stated period, not the probability that any uploaded video will earn a citation.
Popularity was not a requirement in this sample
Among the cited videos analyzed for view count, the largest bucket was 10,000 to 100,000 views, at 36%. Nearly 80% had fewer than 100,000 views. At the other end, the report says about 7% had passed one million views.
That distribution is useful because it rules out a simplistic threshold theory. A video did not need millions of views to appear in the sample. It does not prove that lower view counts improve citation odds. To estimate such an effect, a study would also need a comparison group of relevant videos that were available but not cited, then account for topic, channel history, publication age, and retrieval eligibility.
The same issue applies to likes and subscribers. The report describes the properties of selected sources. It does not isolate which property caused selection. A narrow tutorial with modest reach may fit a prompt better than a broad video with a large audience, but relevance is only a proposed explanation unless the retrieval system or a controlled experiment confirms it.
For a content team, the measured conclusion is narrower and still useful: do not reject a YouTube opportunity merely because the channel is small. The February sample contains many cited videos and channels below the largest popularity bands.
Older videos remained available to AI answers
The age panel in the February 1 through February 28 study found that 54% of cited videos were more than two years old. Nearly 72% were more than one year old. Only nine of the 226 videos in that age sample were under one month old, roughly 4%.
Again, this is a distribution of cited videos. It does not demonstrate that age itself earns a citation. Older videos have had more time to accumulate links, captions, comments, and topical associations. They also outnumber newly published videos in many established categories. Without an eligible-video baseline, those factors cannot be separated.
Still, the finding changes what should be audited. Teams often look only at the latest uploads when checking AI visibility. The February data supports checking the back catalog as well. An older explainer may still be the page an engine retrieves, and an outdated description on that page may therefore matter more than the copy on a new upload.
Do not rewrite or delete an older video solely because one AI answer omitted it. First collect repeated observations. Then inspect whether the video remains indexed, whether its transcript still answers the monitored question, and whether another source consistently replaced it.
YouTube is not one channel across models
The engine split is consistent with the wider lesson in Promptwatch's live social citation data: social source shares differ by AI product. The live page changes over time, so it should not be merged numerically with the fixed February study. Use it as a current directional check, with a dated snapshot, rather than silently updating a February conclusion.
This distinction matters when setting goals. "Increase YouTube citations" has no stable denominator. A better target names the platform, prompt cohort, market, and period. For example, a team can monitor the percentage of eligible Perplexity responses in one product topic that cite at least one video from its channel over a rolling four-week window.
The phrase "eligible response" needs a definition too. Some teams divide by all runs. Others divide only by responses that searched the web or returned at least one citation. Either formula can be useful, but switching between them creates a false trend.
Measure the cited page, not only the channel
Start with a fixed set of questions that videos can genuinely answer. Run those prompts repeatedly on the engines your audience uses. For every YouTube citation, save the video URL, title, publication date, observed view band, and the passage or claim supported by the citation.
Then separate four outcomes:
- the engine cited any YouTube video
- the engine cited a video from your channel
- the answer named your brand without citing your channel
- the citation led to a measurable visit
Those outcomes answer different questions. A channel citation can support an answer without naming the brand. A brand mention can come from a third-party video. A citation does not guarantee a click.
Repeated runs are needed because source selection varies. The AI answer volatility methodology, updated August 1, 2026, recommends at least around three runs per prompt per platform in a rolling seven-day window, with more observations before making a claim about one prompt. Treat that as published sampling guidance, not a guarantee that three runs make every estimate reliable.
Turn the evidence into a restrained video test
The February findings support two reasonable tests. First, answer a specific recurring question rather than assuming broad reach will translate into citations. Second, make the answer easy to retrieve by using an accurate title, description, chapters, and captions. The study does not prove that any one metadata change raises citation rates, so record the change and watch a stable prompt cohort afterward.
For practical measurement, Promptwatch can track YouTube citations alongside prompts, engines, and citation trends. The Promptwatch review explains how that citation layer fits with crawler and visitor data. It is most useful when the team keeps the February benchmark separate from its own current measurements.
YouTube deserves attention where observed engine behavior supports it. The February 2026 sample gives Perplexity and AI Overviews teams a reason to investigate video, while the ChatGPT share in that same month was far smaller. That is a platform-specific measurement finding, not a promise that video production will cause visibility.