YouTube is an Answer Engine
LLMs aren't watching your videos. They're reading them. Once you accept that, video stops being a brand asset and becomes a search asset.
I started out in content, back when the job was words on a page and the only question that really mattered was whether Google liked them. Video sat somewhere else entirely. It belonged to the social team, it was measured in views and watch time, and the closest it came to my work was when I dropped an embed into a blog post to nudge time on page. For a long time that division of labour made sense, because search engines indexed text and video was something people watched rather than something machines read.
But that, my friends, has changed, and the reason is fairly unglamorous. The systems that now sit between a brand and its buyers, meaning ChatGPT, Perplexity, Google AI Overviews and the rest, do read video. And they read it well. They read the transcript, the description, the chapter titles and the metadata around it.
“LLMs aren't watching your videos. They're reading them.”
Jessica Finch
Once you accept that, video stops being a brand asset and starts being a search asset, and it needs to be planned by the people who understand how questions get answered rather than by the people who understand how films get made.
For years marketers treated YouTube as a distribution channel and very little else. You made the film, you put it somewhere people could watch it, and the platform was the shelf rather than the strategy. What has shifted is not the platform's audience but its readability, and the reason YouTube has become so useful to language models comes down almost entirely to structure. Most social platforms give a crawler an unruly mess of short posts, replies and images with very little to anchor them. YouTube gives it something closer to a document. Every video carries a transcript, a description, metadata, and in many cases chapters and timestamps that break the content into labelled sections. That is clean, structured text of exactly the kind these systems are built to ingest, attribute and quote.
It is an article with a face on the front.
What the research actually shows
The most substantial work I have seen on this is Otterly.AI's YouTube Citation Study 2026, which analysed more than 100 million AI citation instances across ChatGPT, Google AI Overviews, Google AI Mode, Perplexity, Microsoft Copilot and Gemini over a thirty-day window. It is worth reading in full, but a handful of findings should change how you plan.
Start with the reality check. Social and video platforms account for only around 5.54% of all AI citations, and brand-owned domains still make up the majority at 52.2%, so nobody should be reallocating a budget on the strength of this. Within that slice, though, YouTube took 31.8% of citations and Reddit 46.4%, meaning the two platforms between them represent 78.2% of all social citations in AI search.
The finding that matters most is that popularity barely registers. The correlation between citation frequency and view count was minus 0.03, likes minus 0.02 and channel subscribers minus 0.03, all of which is noise. In practice, 40.83% of cited videos had fewer than a thousand views and 35% of cited channels had under ten thousand subscribers. Length, on the other hand, matters enormously. Long-form video accounted for 94% of citations against 5.7% for Shorts, with nearly all the Shorts activity confined to Google's own surfaces, and the cited videos clustered between ten and twenty minutes.
Structure compounds it. Of the videos carrying timestamps, 78% were cited more than once, usually across two to five different chapters, because Google's systems appear to treat a chapter marker much the way they treat a subheading on a web page. One well-structured video can become five separate citable answers. Descriptions carry more weight than you would expect too: description length was one of only two metadata features with any meaningful relationship to repeat citation, at r = 0.31, and the average description among cited videos ran to around 334 words. Finally, the opportunity is not evenly distributed. Perplexity accounted for 38.7% of YouTube citations and Google AI Overviews for 36.6%, while Gemini and Copilot sat at 0.2% and 0.5%. If your buyers live inside Copilot, video is not where I would start.
A note on a headline you may have seen. Some coverage claims YouTube has overtaken Reddit. That comes from a different dataset with a different methodology, and I would be careful about repeating it.
“AI engines don't reward popularity. They reward clarity.”
Thomas Peham, Otterly
The practical consequence is that YouTube rewards the same discipline that good bottom-of-funnel content always required, which is answering a specific question properly and in full. A ten-minute video that works through one real question, with the answer stated clearly near the start and the reasoning laid out afterwards, gives a language model something concrete to extract. A polished brand film about transformation and partnership gives it nothing at all, however many people watch it.
There is a second benefit that has nothing to do with citations, and something a little more obvious: customers have questions, so answer them. When a prospect watches you explain the thing you sell, in your own voice, they form a view on whether you know what you are talking about long before they reach your pricing page. That is particularly true of the awkward questions, the ones about cost, about limitations, about who the product is genuinely not right for. Answering those on camera costs you the occasional poor-fit enquiry and earns you a good deal of credibility with everyone else. The happy accident of this moment is that the videos which build the most trust with people are largely the same videos that get cited by machines, because both are looking for a direct, specific and honest answer.
That logic extends to the money you already spend on creators, and this is the part I find most interesting. LLM crawlers filter out standard advertising units, so a pre-roll ad has essentially no influence on what ChatGPT says about your product. What these systems draw on instead is the organic narrative inside the transcript, which is to say the words a creator actually speaks. Because a brand partnership inside a creator's video does not necessarily surface in the video's metadata, there is every chance that sponsored content mentioning your brand is being read and cited as ordinary organic information. A partnership that used to be justified on reach and brand lift is now also building the source material that AI systems draw on when someone asks about your category.
How to optimise YouTube for brands
If I were setting this up from scratch, this is the order I would work in.
- Answer the damn question. Sounds simple and it is, so do it. Start from the ten questions your sales team fields most often, checked against the follow-up prompts ChatGPT and Perplexity suggest around your category. One question per video, and make the question the title. Otterly found the average title among cited videos ran to nineteen words, so there is no prize for being brief or clever.
- Structure your script. State the answer plainly in the first thirty seconds, then show your working in clearly separated sections, each of which stands on its own as an answer to a sub-question. Use specific numbers, named frameworks and real examples rather than general advice. Close by restating the key point in a sentence or two.
- Optimise your thumbnail. This one is for humans rather than machines, and I would not pretend otherwise. There is no evidence that a thumbnail influences whether an AI cites you. It influences whether a person ever watches the video, which is the other half of the reason you are doing this at all. Faces, a short piece of text that matches the question, and consistency across the channel so people recognise you.
- Chapter the video properly. YouTube's requirements are that the first timestamp starts at 00:00, that there are at least three timestamps in ascending order, and that each chapter runs a minimum of ten seconds. Name each one as a question or a task rather than “Intro” or “Part Two”.
- Write the description as a standalone summary. A short paragraph on what the video covers, the tools, products and concepts named in it, the chapter list, and a link to the related page on your own site.
A word on AI-generated video
I would avoid it, and I want to give the reasons rather than just the opinion.
The first is that YouTube has moved against it directly. On 15 July 2025 the platform renamed its “repetitious content” policy to “inauthentic content” and updated the guidelines to identify material that is mass-produced or templated. YouTube's position is that such content was never eligible for monetisation and that this was a clarification rather than a new rule, but the clarification made the target explicit, and channels have been demonetised under it since.
The second reason is more fundamental. What makes a video worth citing is that it contains something a language model could not have produced on its own, whether that is a figure from your own client work, a judgement formed by having got it wrong before, or a demonstration of your product doing an actual job. A generated video is a model writing text for another model to read, and there is nothing in the middle that was not already in the training data. It is difficult to be the best available answer when you are a summary of the existing ones. The third reason is that people notice. Trust is the reason to be on YouTube at all, and synthetic presenters spend that trust rather than build it.
than a hundred generated ones.
YouTube really is a no-brainer for brands, and the best thing is that performance and brand now work as one. Here's to reaping the benefits.
Lisa Steingold is an organic search and AI visibility specialist. When she's not working on brands, she's riding her motorbike.
Which questions are your buyers asking, and do you come back as the answer?
Free, live, no sign-up.
Run your Answerz Share checkSources
- Otterly.AI, YouTube AI Citation Study 2026. More than 100 million AI citation instances across ChatGPT, Google AI Overviews, Google AI Mode, Perplexity, Microsoft Copilot and Gemini, over a thirty-day window.
- Jessica Finch, on YouTube's role in AI visibility, as cited by the author.
- YouTube, “inauthentic content” policy update, 15 July 2025.
