Multimodal GEO: How to Get Your Video & Images Cited by AI

Multimodal GEO is the discipline of building a machine-readable text layer around your video and images — VideoObject/ImageObject schema, an on-page visible transcript, and descriptive alt text — so AI engines can extract, attribute, and cite your media. The core problem: ChatGPT, Perplexity, Google AI Overviews, and Gemini do not "watch" your video or "see" your image; they cite the words attached to the pixels. If those words don't exist in extractable, server-rendered HTML, your best-produced video is invisible to citation.

This guide is for marketers, SEOs, and GEO practitioners who publish videos, demos, charts, and blog images and want to know why that media never surfaces in an AI answer — and the repeatable stack that fixes it. The outcome: a media page that moves from mentioned to cited.

The hero proof comes from our own audit. Seven Indexable pages carried FAQ schema with no visible FAQ text — and earned zero measurable AI-citation benefit until we added the visible, answer-first text (Indexable B2B AI-Visibility Audits, July 2026). That is the law of multimodal GEO in one finding: schema alone is invisible to citation; schema plus visible, extractable text is what gets cited. A VideoObject tag with no on-page transcript fails for the same reason.

What is multimodal GEO?

Multimodal GEO is generative engine optimization applied to non-text media — video, images, audio, and charts — by optimizing the machine-readable text layer AI engines read in place of the media itself. It differs from traditional video SEO in its target: classic video SEO chased the rich snippet and a blue-link click, while multimodal GEO chases citation inside an AI-generated answer, where there is often no ten-result list at all — the unit of success moves from ranking position to being the attributed source.

The strategic bet is timing. "Generative engine optimization" went from 726 searches per month in January 2025 to 14,343 by July 2025 before settling near 7,000–8,000 in 2026 (Ahrefs, via Indexable B2B AI-Visibility Audits, July 2026). Multimodal-GEO terms sit today where GEO sat in early 2025 — near-zero volume — so owning the category now is an entity play ahead of the demand curve.

Why don't my videos and images get cited by AI?

Your videos and images don't get cited because AI engines index the machine-readable text around the media, not the media itself — and most embeds and images ship with almost no such text. A JS-injected YouTube iframe, an image named IMG_4823.jpg with empty alt text, or a chart saved as a flat PNG all present a crawler the same thing: a rectangle with no meaning attached. Nothing to chunk, nothing to quote, nothing to cite.

This is the same failure mode as the schema-only-FAQ finding, and media is worse off because a video frequently has neither schema nor visible text. The mention-to-citation gap is measurable: Indexable was mentioned in 40 ChatGPT answers but cited in only 9 — a 22.5% citation rate, with 31 prompts naming us without linking us (Indexable B2B AI-Visibility Audits, July 2026). Multimodal GEO closes that gap.

Share of Voice makes the same point at category scale. Profound leads ChatGPT Share of Voice at ~70%, with Indexable second at ~26% — ahead of AthenaHQ (~23%) and Search Atlas (~8%) (Indexable B2B AI-Visibility Audits, Ahrefs Brand Radar, July 2026). But Share of Voice counts mentions, not citations, and the whole category is barely cited — being named in an answer is not the same as being linked as its source. That mention-to-citation gap, not raw mention volume, is the real battleground, and closing it — the action and citation layer, not share of raw mentions — is exactly what multimodal GEO does.

Do AI engines actually "watch" video or "see" images?

No — production AI search engines do not watch video or view images to source most citations; they read proxies. For video, the proxies are the automatic speech recognition (ASR) transcript, captions, the title and description, and on-page copy near the embed. For images, they are alt text, the filename, the caption, the surrounding paragraph, and ImageObject schema. Some frontier models can analyze a keyframe, but that vision pass is expensive and inconsistent, so citation leans on the cheap, reliable text layer.

This is why a rendering audit matters. We quantify raw-versus-rendered word count with Screaming Frog rather than trusting a boolean "JS-dependent" flag, which repeatedly catches media pages where the transcript exists only after JavaScript runs — invisible to a crawler reading raw HTML (Indexable B2B AI-Visibility Audits, July 2026). A JS-injected player with no server-rendered transcript is, for citation, a blank page.

How does VideoObject schema get your video cited?

VideoObject schema gets your video cited by handing engines a structured, labeled summary — name, description, thumbnail, upload date, and, critically, the transcript property — so the model can attribute a specific claim to a specific video URL. Google's video structured-data docs and Schema.org define these; the transcript and description fields hold the citable language.

The minimum citable set is name, description, thumbnailUrl, uploadDate, contentUrl, and a populated transcript. The full JSON-LD appears in the schema block below — this page carries a live VideoObject with a real transcript, practicing what it preaches. Deploy it through your technical SEO agent alongside the visible transcript, never in place of it.

Why does an on-page visible transcript beat schema alone?

An on-page visible transcript beats schema alone because engines cite the extractable text a page renders — schema is a hint, not a substitute for words on the page. This is the direct read-across from our hero finding: the seven schema-only FAQ pages produced no citation benefit until the visible answer text went live (Indexable B2B AI-Visibility Audits, July 2026). Schema tells an engine what a thing is; only the transcript gives it a sentence to quote.

Our directional multimodal-citation observation reinforces this. Method: within the Indexable video-engine rollout we compared two page states — (A) VideoObject schema plus a server-rendered visible transcript versus (B) schema only, with no transcript — and tracked which appeared as a cited source (Indexable B2B AI-Visibility Audits, July 2026). Stated honestly as a case observation, not a controlled A/B: state-A pages surfaced as cited media sources while state-B pages stayed mentioned or absent. The takeaway is atomic: a VideoObject tag without a visible transcript underperforms one with a transcript, because AI cites the transcript, not the tag.

What each layer gives the AI, and whether it is citable alone:

How do you write image alt text and ImageObject schema for AI?

You write image alt text and ImageObject schema for AI by treating alt text as extraction fuel — descriptive, entity-rich, full-sentence context — not a short keyword string for accessibility. Good alt text names the entities and the claim ("Bar chart of ChatGPT Share of Voice — Profound leads at ~70%, Indexable second at ~26%," a real figure from our own Brand Radar — Indexable B2B AI-Visibility Audits, July 2026) — the sentence an engine can lift. Pair it with ImageObject schema (contentUrl, caption, description), a human-readable filename, and a visible caption. Charts need extra care: a flat PNG is unreadable, so every chart needs a text equivalent — the numbers restated in a nearby sentence or table, which is what actually gets cited.

How do you get cited by Perplexity, ChatGPT, and Google AI Overviews with media?

You get cited across engines by ranking the media page in Google's top 10 first, then building the extractable text layer on it — top-10 organic ranking is the shared citation gateway. The per-engine nuances sit on that base:

• Google AI Overviews: 38% of AI Overview citations come from pages already ranking top-10 (Ahrefs, 2026).
• ChatGPT: the #1 Google page is cited 43.2% of the time by ChatGPT (AirOps, 2026), and it fans out into sub-queries, so broad topic coverage raises your odds.
• Perplexity: answers overlap 82% with Google's top-10 (Semrush, 2026) and favor recent content — fresh dateModified plus top-10 rank is the combination.

What is the multimodal GEO checklist for a video or image page?

The multimodal GEO checklist is a 10-step sequence that makes a media page extractable and citable; work it top to bottom for every video or image page you want AI to cite:

• 1. Server-render the transcript — visible in raw HTML, not JavaScript-injected.
• 2. Add VideoObject schema with the transcript property populated.
• 3. Format the transcript for extraction — timestamps, speaker labels, question-style H3s.
• 4. Write entity-rich alt text for every image — full context, not a keyword.
• 5. Add ImageObject schema with caption and description for key images and charts.
• 6. Ship captions as an SRT file and confirm the ASR transcript is accurate.
• 7. Place answer-first copy near the embed — a one-sentence atomic answer by the player.
• 8. Give every chart a text equivalent — restate the numbers in a sentence or table.
• 9. Keep dateModified fresh — refresh inside the ~70-day AI weighting window (Zyppy fan-out analysis, 2026); AI weights updates on that window versus ~13 months for organic.
• 10. Measure in Brand Radar — track the page from mentioned to cited across engines.

Steps 1, 2, and 7 do the heaviest lifting; a page shipping only those three already beats most competitors' media pages. At scale, an enterprise AI SEO agent runs the full checklist across an entire library — the case for AI SEO agents over an agency hand-editing every embed.

How do you measure whether your video or image is getting cited by AI?

You measure media citation by tracking each engine's answer for your target prompt cluster and recording whether your page is mentioned (named, no link) versus cited (linked as a source) — the same split we use to measure AI citation in Brand Radar, part of the wider AI brand visibility loop. Our baseline is concrete: a 22.5% citation rate against 40 ChatGPT mentions (Indexable B2B AI-Visibility Audits, July 2026), and the goal of any multimodal fix is to lift the media page above that line. Watch the prompt "what content ranks in AI Overviews," where we register as mentioned-not-cited, and confirm the page appears as a linked source within the ~70-day recency window after you ship the transcript and schema.

Frequently asked questions

How do I get my videos cited in AI Overviews?
Rank the video's page in Google's top 10 — 38% of AI Overview citations come from top-10 pages (Ahrefs, 2026) — then add VideoObject schema with a populated transcript property and a server-rendered, visible transcript. AI Overviews cite the transcript text, not the video file.

Does a video transcript help AI citation?
Yes — a visible, on-page transcript is the single highest-leverage change for video citation, because engines quote the transcript rather than watching the video. In the Indexable video-engine rollout, pages with VideoObject schema plus a visible transcript were the ones that surfaced as cited sources (Indexable B2B AI-Visibility Audits, July 2026).

How do I get cited by Perplexity AI?
Perplexity overlaps 82% with Google's top-10 results and favors recent content (Semrush, 2026), so rank the page top-10, keep dateModified fresh inside a ~70-day window (Zyppy fan-out analysis, 2026), and give any media a server-rendered transcript or text equivalent. Perplexity cites the extractable text.

Do I need a video schema markup generator?
A generator produces valid VideoObject JSON-LD faster but is optional, and one that omits the transcript property still leaves you uncitable. What matters is that the schema includes the transcript and the same transcript is visible and server-rendered on the page.

How long should a transcript be to get cited by AI?
Publish the full verbatim transcript, not a summary — more accurate extractable text means more atomic claims an engine can lift. Format it with timestamps, speaker labels, and question-style H3s so individual claims survive the 100–200 word sliding window ChatGPT extracts from (Petrovic, 2026).

What content ranks in AI Overviews?
Content that ranks in AI Overviews is text-first content on a page already in Google's top 10 — 38% of AI Overview citations come from top-10 pages (Ahrefs, 2026) — with extractable, answer-first passages the engine can quote. Video and images rank only through their text layer: a server-rendered transcript, entity-rich alt text, and VideoObject or ImageObject schema. The media file itself does not rank; the words attached to it do.