Your video is read, not watched

A video earns an AI citation through its text: the description, the transcript and the chapter timestamps are what the retrieval layer can index, and the largest published study of AI video citations finds that audience metrics have essentially no relationship with how often a video gets cited.

That is a different claim from "video matters for AI visibility." Video does matter, and the citation counts are large. The point is that the thing being cited is a document that happens to have a player attached to it.

#Video citations are one domain

BrightEdge measured video citations across AI surfaces using its AI Catalyst data from May 2024 through September 2025 and reported YouTube holding a 200x advantage over its nearest rival, with Vimeo at 0.1%, TikTok at 0.1%, and Dailymotion and Twitch at zero, according to Search Engine Land's writeup. In the same dataset, 29.5% of Google AI Overviews cited YouTube, against 16.6% in AI Mode, 9.7% on Perplexity and 0.2% on ChatGPT.

Treat those engine figures as a snapshot of a period that ended a year ago. The direction has held, but the per-engine numbers in AI search move fast enough that a 2025 measurement is not a 2026 measurement.

The more recent picture comes from Otterly, which analyzed more than 100 million AI citation instances over a 30-day window across ChatGPT, Google AI Overviews, Google AI Mode, Perplexity, Microsoft Copilot and Gemini, published March 2 2026. There, 5.54% of all citations observed came from social and video platforms, and YouTube represented 31.8% of that subset, second to Reddit at 46.4%.

Read the second study's engine table carefully, because it answers a different question than BrightEdge's. Otterly reports the share of total YouTube citations by platform: Perplexity 38.7%, Google AI Overviews 36.6%, Google AI Mode 19.6%, ChatGPT 4.4%, Microsoft Copilot 0.5% and Gemini 0.2%. That is where YouTube citations land, not how often each engine reaches for video. An engine contributing few citations to the sample lands low in that table for reasons that have nothing to do with its appetite for video.

#Popularity is not the currency

The useful part of the Otterly study is the correlation analysis, because it separates the attributes that track with citation frequency from the ones that do not.

The ones that do not are the ones a marketing dashboard reports. Video views came in at r = -0.03, likes at -0.02 and channel subscribers at -0.03. Those are zeros with noise on them.

The ones that do are textual. Description length reached r = 0.31, hashtag presence 0.20, and recency around 0.3. Weak to moderate, but in a dataset where the popularity signals are flat, they are the only things moving.

The distribution of cited videos says the same thing more bluntly. 40.83% of cited videos had fewer than 1,000 views and 36% had fewer than 15 likes. On the channel side, 35% of cited channels had fewer than 10,000 subscribers and 50% had fewer than 41 total videos, with a median channel total view count around 2.2 million.

A small channel with a thorough description is competing on even terms with a large one. That is the opposite of how the YouTube recommendation surface works, and it is why treating an AI citation strategy as a subscriber-growth strategy misreads the mechanism.

Format matters in one direction only: 94% of citations went to long-form videos and 5.7% to Shorts, with playlists, channels and livestreams making up the remaining 0.3%. A sixty-second clip does not carry enough transcript to answer anything.

#Timestamps are a Google-surface multiplier

Chapter markers are the one structural feature with a clear repeat effect. 31% of cited videos contained timestamp signals, and 78% of those were cited more than once, usually across two to five chapters. One video, several citable units.

The effect is not universal. Timestamped citations in that dataset appeared only on Google surfaces, 73% in AI Overviews and 27% in AI Mode, and were absent from ChatGPT, Perplexity, Gemini and Copilot.

That asymmetry lines up with Google's own documented feature rather than with anything model-specific. Google generates key moments automatically where it can, and publishers can control them explicitly with Clip markup or let Google derive them via SeekToAction, per Google Search Central. On YouTube, the equivalent input is the timestamp list in the description, formatted one per line with the label on the same line, as described in Google's key moments documentation.

The chapter is doing the same job for video that a well-scoped section does for a page. It defines the unit an engine can lift.

#What this data cannot tell you

Otterly states the limitation directly: the dataset contains only videos that were already cited, so the findings explain repeated citation behavior rather than initial eligibility. Conditioning on the outcome is a real constraint. A correlation of -0.03 between views and citation frequency among cited videos does not prove that view count plays no part in getting into the pool in the first place.

What survives that caveat is the negative result about optimization effort. Among videos that already cleared the bar, spending on reach did not buy more citations, and the text attached to the video did.

#The practical version

Write the description as a document, not as a caption. Three hundred and thirty-four words was the average description length among cited videos, which is closer to a short article than to the two-line blurb most brand channels ship.

Add chapters, and label them with the question a viewer would type rather than with section names like "Setup" or "Part 2". The label is the retrievable string.

Publish a transcript somewhere you control. The citation for a YouTube video goes to youtube.com, which means the answer names your brand but the link points at a platform you do not own. A companion page on your own domain carrying the transcript, the chapter headings and the same claims is the only version of that content that can earn a citation pointing back to you.

And check whether your video content is reaching engines at all. The same mechanism that governs how images reach AI answers as text applies here: an embedded player with no surrounding copy is a blank space to a text retriever, however good the footage is.