Video strategy for AI search: search against discovery, and what the citation data actually shows

The short answer

The useful half of the original page was the split, and it survives. Discovery video and search video are different products with different titles, lengths, pacing and success measures. Making one and hoping it does both jobs is the most common waste in a video budget.

The unusable half was the certainty about YouTube’s citation share. We went looking for independent evidence and found the opposite of a settled answer.

• LLM Pulse, updated 28 September 2026 on a seven-day rolling window across five engines, ranks youtube.com first at 9.8% of citations.
• Semrush, from 230,000 prompts and more than 100 million citations across three engines over thirteen weeks, put Reddit, Wikipedia, LinkedIn, Forbes and Medium in the top five, with YouTube appearing in Google AI Mode’s top five and not in the overall list.
• Figures circulating elsewhere put YouTube’s share on individual engines at under 3%.

That is a spread of more than five times on the same question. Nobody should be building a video budget on a number that unstable, including from us.

So this page keeps the framework and replaces the certainty with a method. Make the two kinds of video deliberately, then measure your own citation rate rather than borrowing anyone’s.

Key takeaways

• Search video and discovery video are different products. Different titles, lengths, structure and metrics. The split is the part of this page that holds up.
• YouTube’s citation share is genuinely contested, from 9.8% to under 3% depending on who measured, which engines, and when.
• Engine mix matters more than the average. Gemini and Google AI Mode route through YouTube far more than ChatGPT does, so your engine mix decides whether video is worth funding.
• We removed our own proprietary claim. A statistic from a dataset we do not publish is not evidence, and our own rules say so.
• The only reliable number is yours. Run a frozen prompt set and count how often a video of yours is cited, which Pepper’s GEO platform reports alongside page citations.

A note on where this comes from. We run organic for more than 250 enterprises and track more than 10 million prompts across every major engine. We are deliberately not deriving a percentage from that, which is the correction this page is making. The citation figures below are other people’s, dated and attributed. The production advice is our craft judgement, labelled as such.

Book a growth audit if you want to know whether engines cite your video today, before you commission more of it.

What is a video strategy for AI search, and what does the citation data say?

Here is the most recent domain-level ranking we could open, from LLM Pulse, updated 28 September 2026 on a seven-day rolling window across ChatGPT, Perplexity, Gemini, Google AI Mode and Google AI Overviews.

Now the problem. Semrush studied the same question over thirteen weeks to October 2025, across 230,000 prompts and more than 100 million citations on ChatGPT, Google AI Mode and Perplexity. Its top domains were Reddit, Wikipedia, LinkedIn, Forbes and Medium YouTube appeared in Google AI Mode’s top five and not in the overall ranking.

Three things explain part of the gap, and none of them resolves it.

• Different engines. LLM Pulse covers five surfaces including two Google properties that route heavily through YouTube. Semrush covers three.
• Different periods. Semrush’s window closed in October 2025. LLM Pulse’s is a rolling week in September 2026. Eleven months is a long time in this category.
• Different counting. LLM Pulse counts a domain once per answer at most and excludes answers with no citations. Semrush does not describe the same rule.

Both are vendor studies. LLM Pulse sells AI search measurement software and ranks its own product first on its own comparison of AEO tools, which we noted when we reviewed that category. Semrush sells an AI visibility toolkit. Neither is disqualified by that and both should be read with it in mind.

What we take from it, and it is deliberately modest: video is cited often enough to be worth making, the share is contested, and the variation by engine is larger than the variation between studies. That is a reason to measure your own, not a reason to fund a channel.

The split that still holds

This was the good idea on the previous version and it needs no statistics to stand up.

Discovery video

Made for a recommendation feed. The viewer did not go looking for you, so the first three seconds decide whether they stay.

• Title: a hook. Curiosity, tension or a promise, not a literal query.
• Length: longer, typically eight to fourteen minutes, because watch time is the currency.
• Structure: fast opening, varied pacing, a presenter on camera.
• Thumbnail: high contrast, expressive, readable at small size.
• Measured on: click-through rate, retention curve, subscribers.

Search video

Made for somebody who asked a question. They already want the answer, so the job is to give it quickly and make it quotable.

• Title: the literal question, phrased the way a buyer would type or say it.
• Length: shorter, typically two to six minutes, because nobody came for entertainment.
• Structure: the answer in the first thirty seconds, then numbered chapters.
• Support: burned-in captions, a full transcript rendered on the page that embeds it.
• Measured on: citations and whether the transcript gets quoted, not retention.

The transcript is the part teams skip and it is the part that matters most for AI search. An engine quotes text. A video with no readable transcript on the embedding page is, to a crawler, a page about a video rather than an answer. That is the same retrievability problem covered in what a crawler actually receives, and it applies to video more severely than to anything else.

Where this falls short: the lengths and structures above are Pepper’s craft judgement from the channels we run. They are not measured optima. We have not tested them against a control and we are not going to present them as though we had.

How to split the channel

The previous version prescribed a fixed split. We would not, because the right answer depends on which goal you are funding and on your engine mix.

• If the goal is audience, discovery leads. Search video still earns its place, because it compounds quietly, but the channel lives on the feed.
• If the goal is citations, search video leads and discovery becomes the thing that makes the channel credible enough to be worth citing.
• If your buyers use ChatGPT more than Gemini, weight this down overall. The engine that routes most heavily through YouTube is not the one most of your buyers are using, on the data above.

The honest version, and the one Pepper briefs: pick the goal, let the split follow, and revisit when you have your own citation numbers rather than ours.

Our methodology: how we judge a video claim

Weights sum to 100. Source availability and stated method carry 60 between them. That is Pepper’s judgement rather than a fact, and it is the judgement this page rests on. We publish our weightings the same way we publish [how we rank GEO agencies](https://www.pepper.inc/blog/geo-agency-ranking-methodology/), so you can argue with them precisely.

What would change these weights. If an independent, non-vendor study of citation shares appears, the third criterion drops sharply. None exists today that we could find.

Which video earns its budget

Two questions decide whether a video format deserves funding, and neither is production value.

• A chaptered answer video with a rendered transcript scores highest on citation and lowest on audience. It is the format nobody enjoys making and the one engines can actually use.
• A long-form discovery piece scores the reverse. It builds the channel and is rarely quoted.
• Shorts and vertical video score low on both for AI search specifically. They build reach, they rarely carry a quotable transcript, and we would not fund them as a citation play.
• A recorded webinar with a transcript sits in the middle and is the cheapest thing most teams already have sitting unused.

Video for AI search at a glance

What this costs

• Adding transcripts to what you already have: cheap and unglamorous. This is the highest-return item on the page and most teams have a library of untranscribed webinars.
• Making search video: cheaper than discovery video, because it needs no hook, no heavy edit and no thumbnail testing. A clear presenter and good chapters.
• Making discovery video: expensive, and priced as brand work. Judge it on audience, not citations.
• What you should not pay for: a video package sold on a citation share figure. Ask which study, which engines, and which period. The spread on this page is your reason to ask.

For what the broader programme costs, our generative engine optimization cost benchmarks collects 30 published prices.

Where Pepper fits

We are an agentic organic growth engine, and on this page the honest position is that we removed our own claim rather than defending it.

• An organic growth partner. A growth team is attached to your account, a senior strategist plus always-on agents, accountable for the number rather than for a content calendar.
• Our GEO platform. Genuinely self-serve and included rather than billed as a separate licence. You log in and set up a workspace. Define your brand profile, competitors and personas, then connect GA4 and Search Console, and manage your own themes, prompts and analytics across six engines including ChatGPT, Perplexity, Gemini, Claude and Google AI Overviews.
• The Agent Atlas. Your team builds, versions and runs its own agents, and keeps them.

The measurement is the point here. Tracking whether engines cite a video of yours, on your prompt set, gives you a number about your category that no published study can. For what the combination produces, the Acceldata case study records 6X organic traffic and top-three keywords rising from 85 to more than 300, mostly through B2B SaaS programmes.

Where it falls short: we are not a video production company and we will not make these for you. We also will not give you a Pepper citation-share figure for video, because publishing a percentage from a dataset we do not release is exactly what this page is correcting. That makes us less useful than a supplier who hands you a number, and we think the number was the problem.

How to choose what video to make

I will write this in the first person, because a recommendation nobody will defend is worth nothing.

Start with your engine mix, because it changes the answer more than anything else on this page. It is the first thing Pepper checks before advising on video at all. If your buyers are mostly in Gemini and Google AI Mode, video is plausibly worth real budget. If they are mostly in ChatGPT, the case is much weaker on the same data. Most teams have never checked, and it takes an afternoon.

So my order is: check the engine mix, transcribe what you already have, then commission deliberately into one of the two modes. Here is the scorecard I would use on any proposed video, on a 100 point scale.

Five questions I would ask before commissioning anything, in this order.

1. “Which mode is this?” A stronger answer names one and accepts the trade. A weaker one says it will do both, which usually means a hook bolted onto an answer and neither working.
2. “Where will the transcript live?” On the page that embeds the video, rendered as text. If the answer is “YouTube has captions”, the answer is no.
3. “Which question does this answer, in the buyer’s words?” If nobody can produce the question, this is a discovery piece, which is fine as long as you say so.
4. “Which engines do our buyers use?” This is the check that reorders the budget, and almost nobody has run it.
5. “How will we know in 90 days?” Name the measure before production starts, and make it the right one for the mode.

Red flags, each one something a supplier actually says. A citation share for YouTube offered without a study, engines and period. “Video is the most-cited format”, which the evidence on this page does not support. A package priced on AI visibility with no transcript deliverable. A hybrid video sold as serving both modes. And any guarantee of citations, which nobody can promise.

Before commissioning, measure what you have. Write 25 real commercial questions your buyers would type. Run each of those 25 prompts three times across every engine you care about, because engines are probabilistic and one run is an anecdote. Then give any fix 90 days before you judge it. Four worked examples for a mid-market B2B software company:

• “best data observability platform for enterprise”
• “how do I monitor data quality across Snowflake and Databricks”
• “what should I look for in a data observability vendor”
• “who competes with Monte Carlo in data observability”

Count how often a video appears in the answers at all, and whose. That single count tells you whether video matters in your category better than any published share, because it is your category.

And the answer that loses Pepper the sale: if engines are not citing video in your category, you do not need a video strategy for AI search and you do not need us to advise on one. Transcribe what you have, spend the money on earned media, and revisit in two quarters.

It comes down to one principle: make the two kinds of video on purpose, and measure your own share rather than borrowing one. We borrowed one on the version of this page we are replacing, ours included in the count of people who should have known better.

What nobody should promise you

• A reliable citation share for YouTube. The studies we could open disagree by more than five times.
• That video is the most-cited format. One vendor ranking puts YouTube first. Another credible study does not have it in the top five.
• A citation from a video with no transcript. An engine quotes text. There has to be text.
• A split that works for everyone. The right allocation depends on your goal and your buyers’ engine mix.
• A Pepper figure for any of this. We have a large tracked prompt set and we are not going to publish a percentage from it without publishing the data, which is the correction this page exists to make.

Frequently asked questions

Is YouTube really the most-cited domain in AI answers?
It depends who measured. LLM Pulse ranked youtube.com first at 9.8% in September 2026 across five engines. Semrush’s thirteen-week study to October 2025 put Reddit, Wikipedia and LinkedIn ahead, with YouTube outside the overall top five.

What is the difference between search video and discovery video?
Discovery video is made for a recommendation feed and judged on retention and subscribers. Search video answers a specific question, opens with the answer, uses chapters and a rendered transcript, and is judged on whether engines cite it.

Does video help with AI search visibility?
Sometimes, and it depends heavily on which engines your buyers use. Gemini and Google AI Mode route through YouTube far more than ChatGPT does, so the same video strategy pays very differently across two similar companies.

Do I need a transcript for AI search?
Yes, and it should be rendered as text on the page that embeds the video. An engine quotes text. Auto-captions inside a player are not the same thing, and a video without a readable transcript gives a crawler nothing to work with.

How long should a search video be?
We aim for two to six minutes, with the answer in the first thirty seconds and numbered chapters after it. That is our craft judgement from the channels we run rather than a measured optimum, and we have not tested it against a control.

Do Shorts and vertical video get cited?
Rarely, in our experience. They build reach, and they seldom carry a quotable transcript on a page an engine can read. We would fund them for audience and reach rather than as a citation play, and judge them on that.

What split should I use between the two modes?
There is no universal split and we would distrust one. If the goal is audience, discovery leads. If the goal is citations, search video leads. Revisit once you have your own citation numbers rather than a published average.

How do I measure whether video is working for AI search?
Run a frozen set of 25 buyer questions three times each across the engines you care about, and count how often a video appears in the answers and whose it is. That is your category’s number, which no published study can give you.

Where to go next

If the constraint is that crawlers cannot read your pages at all, the three tiers of hidden content covers what a crawler actually receives, which applies to transcripts directly.

To decide where video sits in a wider plan, hero hub hygiene in 2026 covers the three layers and the routes each takes to an engine. To measure any of it, the GEO measurement stack sets out which KPIs mean anything.

Sources and further reading

• LLM Pulse, “Top Cited Domains”, data updated 28 September 2026, read at source on 30 September 2026. Seven-day rolling window across ChatGPT, Perplexity, Gemini, Google AI Mode and Google AI Overviews, described as millions of AI answers analysed weekly from thousands of real prompts. Citations resolved to registrable domain, counted once per answer at most, answers with no citations excluded. Top domains: youtube.com 9.8%, google.com 9.36%, reddit.com 6.33%, facebook.com 4.66%, instagram.com 3.18%, linkedin.com 2.66%, trustpilot.com 1.84%, wikipedia.org 1.7%, tiktok.com 1.54%. LLM Pulse sells AI search measurement software and ranks its own product first on its own comparison of AEO tools. Not hyperlinked, per our policy on vendor links.
• Semrush, “The Most-Cited Domains in AI: A 3-Month Study”, Luke Harsel, published 10 November 2025, read at source on 30 September 2026. 230,000+ prompts and more than 100 million citations across ChatGPT, Google AI Mode and Perplexity, 14 July to 12 October 2025, top 25 domains tracked weekly. Top domains included Reddit, Wikipedia, LinkedIn, Forbes and Medium. YouTube appeared in Google AI Mode’s top five and not in the overall ranking. Semrush states it cannot determine the causes behind citation shifts, only that they occurred. Its window closed in October 2025, which is why it is quoted as history rather than as a current share. Semrush sells an AI visibility toolkit. Not hyperlinked, per our policy on vendor links.

A note on the statistics that are not here. Two claims from the previous version were removed. “YouTube is the most-cited domain inside large language models”, attributed to our own 2026 dataset, is gone because we do not publish that dataset and our craft rules say a statistic derived from it is not publishable unless we do. On the independent evidence above, it is also contested. “317% higher selection rates in AI Overviews” for multi-format pages is gone because it carried no source and we could not find one. The illustrative view counts attached to an unnamed brand were also removed.

Further reading on Pepper, each checked live on 30 September 2026:

• The three tiers of hidden content, which is where transcripts live or die.
• Hero hub hygiene in 2026, where video sits in a wider plan.
• The GEO measurement stack, which KPIs mean anything.
• AI search statistics that hold up in 2026, the market numbers with their samples attached.
• How to structure content for answer engines, because a transcript still has to read as a self-contained answer.
• Prompt tracking, for building the frozen prompt set this page keeps recommending.