YouTube

Otterly's 2026 study, drawn from more than 100 million citation instances tracked over 30 days including roughly 5.5 million social media citation instances^[4], found social and video platforms supplying 5.54% of all AI citations^[1]. Inside that slice Reddit takes 46.4%, YouTube 31.8% and LinkedIn 13.0%, so two platforms carry 78.2% of everything social an answer engine quotes^[2]. One vendor's panel is one measurement, and the figures move; the ordering has been stable across citation studies for longer than the percentages have.

YouTube citations do not spread evenly. Perplexity accounts for 38.7% of them, Google AI Overviews 36.6%, AI Mode 19.6%, ChatGPT 4.4%, Copilot 0.5% and Gemini 0.2%^[3]. Timestamped citations, which drop the reader at a moment inside a video rather than at the video, appear almost only in Google's own surfaces: 73% in AI Overviews and 27% in AI Mode, with none observed in ChatGPT, Gemini, Copilot or Perplexity^[5]. Google said in March 2025 that it was building more visual responses into AI Mode, images and video among them^[6], and its guidance for site owners asks for text supported by high-quality images and video where applicable^[7].

94% of YouTube citations point at standard long-form videos, 5.7% at Shorts and 0.3% at playlists, channels and livestreams together^[8]. Otterly's reading of its own data is that answering the question well, rather than view count or subscriber count, decides whether a video is cited^[9]. That matches what the same studies find for text: retrieval selects for a passage that answers, which is the argument behind content structure work generally.

An engine reads a video as words. YouTube produces automatic captions by speech recognition across a published list of languages, and warns that they can misrepresent what was said through mispronunciation, accent, dialect or background noise, so creators should review them^[10]. An uncorrected auto-caption is the text a retrieval system indexes and may quote. The Data API exposes only caption-track metadata by default, and the transcript itself needs a separate captions.download call^[11], which is why third-party visibility tools read transcripts unevenly.