Multimodal: Video GEO — Optimizing for the Machine Gaze

A typical model parses a video in three separate streams (visual, audio, text), tokenizes each, and cross-references the layers. LLMs cite YouTube constantly, so video is both a retrieval surface and a social format — and search strategists have to treat video as a structured dataset a machine can read. This talk covers the mechanics: why your fast-cut product shot statistically doesn’t exist when the model samples at roughly one frame per second, how audio bolding makes speech models catch your brand names, and why caption contrast is a machine-readability problem on top of an accessibility one.