YouTube is becoming an increasingly important source of information for AI search.

In our analysis of AI-generated answers across Google AI Overviews, Google AI Mode, Gemini, ChatGPT, and Perplexity, YouTube emerged as the single most frequently cited domain among the sources we tracked.

More than 5% of all citations in the dataset pointed to YouTube.

That raised a simple question:

What makes one YouTube video more likely to be cited by an AI system than another?

To find out, we analyzed more than 1,000 YouTube URLs that appeared in AI-generated answers, looked at citation frequency across multiple AI platforms, and examined dozens of video transcripts in greater depth.

The results point to a very different definition of a “good” YouTube video.

Traditional YouTube optimization tends to focus on views, subscribers, engagement, thumbnails, retention, and production quality.

AI search appears to care about something else:

How useful is the information in the video for answering a specific question?

And several patterns stood out.

YouTube accounted for more than 5% of AI citations

Across the AI responses we monitored, YouTube represented more than 5% of the total citations we observed.

The citations were distributed across five major AI search environments:

Google AI Overviews
Google AI Mode
Gemini
ChatGPT
Perplexity

Google’s AI search experiences accounted for approximately two-thirds of the YouTube citations in the dataset.

The approximate distribution was:

AI platform	Share of YouTube citations
Google AI Overviews	38%
Google AI Mode	29%
Gemini	18%
ChatGPT	10%
Perplexity	5%

The exact distribution varies by query set and topic, but the overall pattern is clear: YouTube is not merely appearing occasionally in AI answers. It is a meaningful source of information for AI search systems.

This creates an interesting opportunity for anyone working on Generative Engine Optimization (GEO).

If AI systems use YouTube as a source, then optimizing the information contained in YouTube videos becomes part of the broader AI visibility equation.

The most-cited videos were not necessarily the most popular videos

One of the most interesting findings was what we did not find.

The videos appearing most frequently in AI answers were not obviously determined by conventional YouTube popularity signals.

We found cited videos from channels ranging from relatively small audiences to channels with millions of subscribers.

There was also no obvious relationship between video length and citation frequency.

The shortest video among the highest-cited videos was only 52 seconds long.

The longest was approximately 30 minutes.

Both appeared among the most frequently cited videos.

This suggests that AI citation is not simply a proxy for YouTube popularity.

A video does not necessarily need to be long, highly produced, or published by a major channel to become useful to an AI system.

85% of the highest-cited videos used a comparative structure

The strongest content-structure signal was surprisingly consistent.

Among the highest-cited videos we analyzed, 85% used an explicit comparison, ranked list, or testing format.

Only 15% were straightforward informational or how-to videos without a comparative structure.

The formats broke down approximately as follows:

Video structure	Share	Average weekly citations
Comparison / “X vs Y”	40%	24.3
Ranked list / “Best of N”	30%	21.8
Tested review / “I tested N”	15%	23.1
How-to / tutorial	10%	14.6
Informational / explainer	5%	11.2

The difference is notable.

Videos built around comparisons, rankings, and evaluations generated substantially more citations than straightforward explanatory content.

Why?

Because these formats map naturally to the questions people ask AI systems.

Consider the difference between:

“How does category X work?”

and:

“What are the best options for category X?”

The second question requires an AI system to identify options, compare them, evaluate them, and make a recommendation.

A video that already performs that analysis provides a highly useful source.

“I Tested” videos had the highest average citation rate

The “I tested” format was particularly interesting.

Only around 12% of titled videos used an “I Tested” or “I Tried” format.

Yet those videos had the highest average citation rate among the title formats we identified, at approximately 23.7 citations per week.

That was higher than:

“How to…” videos: 21.4
“[A] vs [B]” videos: 22.1
“Best…” videos: 19.8
Brand-centric review titles: 13.2

The implication is not necessarily that putting “I Tested” in a title causes more citations.

Instead, the format may signal something more important: first-hand evaluation.

AI systems are increasingly being asked questions that require judgment:

Which option is best?
What should I choose?
Is this worth it?
What are the alternatives?
What are the differences?
Which one should I use?

Content that contains actual testing, evaluation, and recommendations can provide useful evidence for answering those questions.

The title patterns were surprisingly predictable

Among the videos for which titles were available, several formats dominated.

Approximately:

31% used a “How to…” format
24% used a “Best…” format
18% used an “X vs Y” format
12% used an “I Tested…” format
15% used a declarative or brand-centric title

These formats correspond closely to the types of prompts that generated citations.

The most common prompt patterns included questions such as:

“What is the best…”
“How do I…”
“A vs B…”
“Is X worth it?”
“What are the best options for…”
“What is currently best?”

This points toward an important principle for AI-oriented YouTube content:

Don’t optimize only for keywords. Optimize for questions.

A video designed around a question-and-answer structure can potentially match a much larger set of AI queries than a video focused narrowly on a single keyword.

The best videos appeared across many different questions

Citation frequency was only one dimension of performance.

Another was prompt breadth: the number of distinct questions for which a video was cited.

The highest-performing video in our dataset was cited for more than 35 distinct prompt phrasings within the monitoring period.

The questions were not simply duplicates.

They represented meaningfully different ways of asking about the same broader subject.

Among the top 10 videos, the median was more than 20 distinct prompt phrasings.

Videos further down the rankings typically appeared for approximately 9–14 different prompt phrasings.

This suggests that the strongest AI-visible videos may not answer just one question.

They cover an entire topical cluster.

That can include:

The main question
Comparisons
Alternatives
Selection criteria
Use cases
Advantages and disadvantages
Who should use a particular option
What to look for
Current recommendations

The more of this decision-making context a video contains, the more opportunities an AI system has to use it as a source.

Semantic matching appears to be extremely important

We also compared video transcripts against the prompts associated with their citations.

In approximately 60% of the transcripts we analyzed, the passage that appeared most relevant to the video’s citations was highly semantically similar to the highest-frequency prompt cluster associated with that video.

The estimated semantic similarity was above 0.85 in these cases.

In practical terms, the video was often talking about essentially the same question that the user was asking.

This may sound obvious, but it has an important implication.

A video can be “about” a topic without necessarily providing the information an AI system needs.

For AI visibility, the content needs to contain the answerable information.

If people ask:

“Which option is best for X?”

a video that spends five minutes introducing the category before eventually discussing the options may be less useful than one that immediately explains the options and provides a recommendation.

The first 90 seconds matter

One pattern appeared repeatedly when we examined transcripts.

Approximately 73% of the transcripts contained an explicit first-person credibility or research claim within the first 90 seconds.

Examples included statements establishing:

Personal experience
Testing performed by the creator
Professional experience
Research conducted
The number of options evaluated
Relevant expertise

Another pattern was even more qualitative.

High-citation videos tended to provide a direct answer or recommendation relatively early, before moving into supporting detail.

Lower-citation videos were more likely to spend significant time on introductions, channel information, background, or other material before reaching the substantive answer.

This suggests a potentially important rule for AI-oriented video creation:

Give the answer before the preamble.

If an AI system is using a transcript to determine whether a video can answer a question, the useful information should be easy to find.

Named entities showed the largest quantitative difference

One of the strongest signals in our transcript analysis was the density of named entities.

The highest-performing group averaged approximately:

5.4 named products, tools, or other entities per transcript.

The lowest-performing group averaged:

0.8 named entities per transcript.

That’s a difference of approximately 6.75×.

This makes sense when you consider the types of questions that generate AI citations.

AI systems frequently answer questions involving comparisons between specific options.

A useful source therefore needs to contain the specific entities being discussed.

Generic statements about a category are less useful when the user’s question is:

“What is the difference between A and B?”

A transcript that explicitly discusses A, B, their use cases, differences, limitations, and pricing provides far more extractable information.

Current information can create additional citation opportunities

We also found a signal around current-year references.

Approximately 27% of the highest-cited videos included the current calendar year either in the title or within the first 30 seconds of the transcript.

This was particularly relevant for queries asking about the current state of a category.

Examples of the underlying query intent included:

What is currently best?
What has changed?
What should I use right now?
What are the best options this year?

This does not mean every video should simply add a year to its title.

The more useful takeaway is that freshness needs to be communicated explicitly when freshness matters to the query.

A video can be technically recent without making its current relevance obvious.

Production quality was not a strong signal

Perhaps the most counterintuitive finding was that production quality did not appear to determine citation frequency.

Among the highly cited videos were basic screen recordings with minimal editing and no on-camera presenter.

At the same time, professionally produced videos in comparable subject areas received little or no citation activity during the observation period.

This is important because it changes the economics of AI-oriented video creation.

If the objective is primarily AI visibility, producing a cinematic video may be unnecessary.

The priority may instead be:

Answer the right question.
Cover the relevant entities.
Provide useful comparisons.
Demonstrate expertise or first-hand experience.
Match the language of real user questions.
Make the answer easy to extract from the transcript.

In other words:

Information quality may matter more than production quality.

Video length did not show a meaningful relationship with citations

The data also argues against a simple “longer is better” strategy.

The highest-cited videos ranged from less than one minute to approximately 30 minutes.

There was no meaningful relationship between length and citation frequency across the observed videos.

This is another reason to focus on information density rather than duration.

A 60-second video that directly answers a question may be more useful to an AI system than a 20-minute video that takes several minutes to reach the answer.

The ideal length is therefore likely to depend on the information required to answer the question.

Subscriber count did not explain citation performance

Channel size was another weak signal.

The highly cited videos came from channels with dramatically different audience sizes.

We did not observe clustering that would suggest that subscriber count alone determines whether a video gets cited.

This is potentially good news for smaller creators and companies.

If AI citation is driven primarily by information relevance, a new or relatively small channel may still have an opportunity to become a source for AI-generated answers.

That is very different from traditional content distribution, where competing with established channels for attention can be extremely difficult.

Publication age wasn’t a prerequisite either

We also observed both evergreen and recently published content among highly cited videos.

Some highly cited videos had been published years earlier.

Others were comparatively new.

This suggests that AI citation does not require a video to be freshly published.

However, current-year language and explicit references to changing information appeared to provide an additional signal for queries where freshness was important.

The distinction is important:

Freshness can help, but freshness alone is not enough.

A useful evergreen video can continue to be cited if it contains information that remains relevant to the questions being asked.

Citation count and cross-platform reach are different things

Another interesting result emerged when comparing total citation frequency with the number of AI platforms citing a video.

The two metrics did not always move together.

For example, one video ranked around the middle of the top-cited group but appeared across 4 of the 5 AI platforms we monitored.

Another video ranked near the very top by total citations but appeared across only 2 platforms.

This means there are at least two different dimensions of AI visibility:

Citation frequency: How often is a source cited?

Platform penetration: Across how many AI systems does it appear?

A video could therefore have relatively high visibility within one AI environment while having limited cross-platform reach.

For brands measuring AI visibility, this distinction is worth tracking.

Different AI platforms showed different citation behavior

The platforms did not behave identically.

Google AI Overviews and Google AI Mode together accounted for approximately 67% of the YouTube citations we observed.

They also showed the broadest retrieval behavior, with a relatively diverse set of videos appearing across queries.

ChatGPT represented approximately 10% of the citations.

The number of unique videos cited was lower, but the videos it selected tended to appear repeatedly across related queries.

Gemini accounted for approximately 18% and showed some geographic and linguistic differences in the videos it surfaced.

Perplexity represented approximately 5% of citations but showed substantial overlap with videos appearing in Google’s AI results.

This reinforces an important point about GEO:

There is no single “AI ranking.”

Different AI systems can retrieve and cite different sources.

Optimizing for AI visibility therefore means understanding the source patterns across multiple systems.

What the data did not show

Some of the most useful findings were negative findings.

We did not see meaningful evidence that citation frequency was driven primarily by:

Production quality
Video length
Subscriber count
Channel size
Publication age
Traditional popularity signals

That doesn’t mean these factors never matter.

It means that within the data we analyzed, they did not explain the differences in citation frequency nearly as well as content structure and information relevance.

The strongest signals were much closer to the content itself.

A possible formula for AI-visible YouTube content

Putting the findings together, a pattern begins to emerge.

The videos most likely to be useful to AI systems tend to have several characteristics:

1. They answer a specific question

The video is built around a question people actually ask.

2. They cover a broader topical cluster

Rather than answering one narrow query, they address comparisons, alternatives, use cases, and selection criteria.

3. They use comparative formats

Approximately 85% of the highest-cited videos used comparisons, rankings, or testing formats.

4. They include specific entities

The strongest group averaged 5.4 named entities per transcript, compared with 0.8 in the lowest-performing group.

5. They establish credibility

Approximately 73% of analyzed transcripts contained a first-person credibility or research signal within the first 90 seconds.

6. They provide the answer early

The most useful information tends to appear before lengthy introductions or background sections.

7. They use language that matches user questions

In approximately 60% of the transcripts examined, highly cited passages closely matched the semantic meaning of the prompt clusters that triggered the citations.

8. They make current information explicit

Approximately 27% of the highest-cited videos explicitly referenced the current year near the beginning of the content.

9. They don’t necessarily require high production value

A simple screen recording can potentially outperform a highly produced video if it contains more useful information.

We’re putting the hypothesis to the test

These findings led to a natural next question:

If these characteristics are associated with highly cited videos, can we deliberately create videos using them and increase the probability of being cited by AI systems?

That’s the experiment we’re running now.

We’ve taken the patterns identified in the research and incorporated them into an experimental AI video-generation system inside rocketblue.

The system can create videos within moments based on the characteristics we’ve observed in highly cited YouTube content.

The resulting videos are published to a dedicated YouTube channel called Rocket Research.

And now we’re doing the part that matters most:

We’re tracking whether the videos actually get cited.

This is an important distinction.

The research above identifies correlations and patterns.

It does not prove that those characteristics cause AI citations.

The only way to find out is to run the experiment.

The experiment is now live

We’re treating Rocket Research as a laboratory for AI search experiments.

The initial hypothesis is straightforward:

If we create YouTube content that closely matches the characteristics of videos already being selected as sources by AI systems, those videos should have a higher probability of appearing in AI-generated answers.

But there are plenty of ways this hypothesis could be wrong.

The videos could fail to get cited.

They could get cited only by Google.

They could appear in one or two queries and then disappear.

They could get citations without accumulating significant YouTube views.

Or we may discover that some of the patterns identified in the research are correlations rather than causal factors.

All of those outcomes are useful.

What we’re going to measure

We’re tracking the experiment across several dimensions:

Number of videos published
Time from publication to first AI citation
Number of AI citations
Number of distinct prompts producing citations
Citation frequency over time
Google AI Overview citations
Google AI Mode citations
Gemini citations
ChatGPT citations
Perplexity citations
Cross-platform penetration
Changes in citation frequency as videos age

The goal isn’t simply to produce AI-generated videos.

The goal is to understand what makes video content useful enough for AI systems to cite.

And we’ll be sharing the results as the experiment develops.

The bigger question

The significance of this experiment goes beyond YouTube.

AI search is changing the way information is discovered.

Traditional search largely answers:

“Which webpage should rank for this keyword?”

AI search increasingly asks:

“Which sources contain the information I need to construct the best answer?”

That changes what it means to create content.

A video doesn’t necessarily need to win YouTube search to be valuable.

A relatively small channel doesn’t necessarily need millions of views.

A video doesn’t necessarily need 20 minutes of production.

Instead, it may need to contain the right information, structured in a way that makes it useful for answering real questions.

If that is true, YouTube could become an increasingly important part of Generative Engine Optimization.

And we’re going to find out just how much of that can be engineered.

Follow the Rocket Research experiment as we publish the results.

We’ll share the data—including the results that don’t confirm our hypothesis.

What Makes a YouTube Video Get Cited by AI Search? We Analyzed 1,000+ Videos