Images, video, and audio can help answer questions that text alone explains poorly. Their value depends on the task and the product’s capabilities. Evaluate the content for its audience and measure its observed use in AI answers separately.
What video evidence can show
A video citation, a retrieved transcript, and direct video processing are different observations. Count the one your test actually exposes. A low citation count cannot establish that the system did not use video information.
Gemini’s API documentation describes video inputs and public YouTube URLs. That establishes a supported API capability. It does not show that every consumer Gemini or Google Search answer watches the video it cites.Gemini video understanding. Documents supported video input methods. Consumer search behavior must be checked separately.
For an existing video, compare a question answered from its transcript with one answered from the video input in a supported tool. Keep the prompt fixed. Record whether visual details missing from the transcript are identified correctly.
When media helps
Choose media where it helps explain the subject. The following uses are practical examples.
- Visual questions: images can show a product, setting, or design detail more clearly than a description.
- Instructions: a demonstration can explain a physical task. Include accurate captions or a transcript where useful.
- Identification: clear product images and labels help users distinguish products and variants.
- Audience use: evaluate viewing, engagement, and customer outcomes independently of citation frequency.
Transcripts and metadata
Publish accurate captions, transcripts, and descriptions where useful. They improve access to the content and provide text for systems that use it. Multimodal models may also process visual or audio information directly.
Provide descriptive titles and relevant surrounding text. Use accurate alt text for informative images and supported video markup where appropriate. A transcript should represent what is said, including material corrections.
Planning media work
First improve the accessibility and accuracy of media you already publish. Create new media when it serves a demonstrated customer need. Measure audience outcomes alongside any citations or recommendations observed.
Do not infer a universal media strategy from one citation sample. Test the formats relevant to the category and retain the ones that justify their production and maintenance cost.