Video GEO

Video GEO is the practice of making video, in practice mostly YouTube video, more likely to be retrieved, cited and quoted by AI answer engines such as Google AI Overviews, AI Mode, ChatGPT and Perplexity. It is a branch of generative engine optimization that treats a video's title, description, chapters and transcript as the text an engine matches against a question. The evidence so far points one way: how closely that text fits the query predicts citation, and the size of the audience does not. All of it is correlational and most of it comes from vendors of AI-visibility software, so it describes what cited videos look like rather than proving what causes a citation.

What Predicts Citation

The strongest single finding comes from Gumshoe (September 2026), which compared 5,000 videos cited in AI Overviews with 207,494 uncited ones drawn from 1.1 million YouTube citations. A model using only the semantic similarity between the prompt and the video's title and description outperformed a model using every piece of video and channel metadata, placing the cited video in its top ten in 59.6% of tasks against 48.6%.

Format matters too. Otterly's study of more than 100 million citation instances (March 2026) found long-form video accounted for 94% of YouTube citations and Shorts for 5.7%, with videos of 10 to 20 minutes forming the largest duration cluster at 32.1%. Gumshoe reports a median length of 9.4 minutes for cited videos against 6.2 for uncited ones. Description length correlated modestly with citation in Otterly's data (r = 0.31), and the average cited video carried a 334-word description. That is a weak signal, not a rule.

What Does Not

Popularity is close to irrelevant. Otterly measured correlations of −0.03 for both views and subscribers, and 40.83% of cited videos had fewer than 1,000 views. Gumshoe's figures are slightly less flat: cited videos came from channels with a median of 30,700 subscribers against 10,900, yet a cited video had only a 52% chance of having more views or likes than an uncited one, which is barely better than a coin toss. Both studies imply that a small channel answering the question precisely can be cited ahead of a large one.

Chapters, Key Moments and Captions

AI Overviews frequently cite a moment rather than a video. BrightEdge data reported by MediaPost (August 2026) found that half of the queries citing a YouTube URL in AI Overviews point to a specific timestamp, often around 40 seconds in. Otterly found 73% of timestamped citations in AI Overviews and 27% in AI Mode. Google's video documentation says it will prioritize key moments set by the publisher, either through structured data or the YouTube description, which makes chapter timestamps the one lever in this field that Google itself documents.

Captions are the least-tested element. No study cited here isolates their effect. The case for them is inferential: Ahrefs' 75,000-brand analysis (December 2025) counted a brand as mentioned on YouTube when its name appeared in a title, transcript or description, and those mentions were the strongest correlate of AI visibility it measured (Spearman 0.74 for AI Mode). Since one vendor study of 264 videos (youtube-transcript.ai, August 2026) put the median word error rate of automatic captions at 9.9%, with a fifth of transcripts lacking sentence punctuation, corrected captions are a sensible precaution for brand and product names. Google's Gemini API documentation states that the model processes both the audio and visual streams of a video, including YouTube URLs. That describes a capability developers can call. It is not confirmation of how Google's search products select the videos they cite.

Engines Differ, and Shares Move

Video citation is concentrated on a few engines. Ahrefs' rankings of the most-cited domains give YouTube very different weight on each.

Other measurements disagree with this picture. Otterly's September 2026 index, which counts YouTube against all citations rather than the top 50 domains, put it at 0.148% for ChatGPT, 0.216% for Gemini and 0.196% for Perplexity, and recorded YouTube's share of AI Mode falling from 13.5% on September 1 to between 3.4% and 3.9% daily from September 16. The denominators are different, so the two sets of numbers cannot be compared directly, and the Perplexity figures in particular are hard to reconcile. What both support is that Google's surfaces are where video citation is densest, consistent with a SIGIR 2026 paper finding that AI Overviews favour Google-owned properties, and that any share can move by a multiple within weeks.

In Practice

The working method is unglamorous: title the video as the question it answers, write a description that states the answer and lists chapters with timestamps, fix the captions, and prefer one substantial video per question over many thin ones. Publishing the transcript as an article on the brand's own site gives engines that cite little video a text version to retrieve. Because citation sets change daily, results need repeated sampling rather than a single check; tracking visibility per engine is the province of AI visibility measurement tools like LLM Optimizer.