Multimodal content: images, video, and AI answers

AI is going multimodal: the engines increasingly understand images and video, not just text, which changes what visibility means for visual content. For years, images and video were effectively invisible to the systems that decided visibility — understood only through the text around them. That is changing as the engines develop genuine understanding of visual media. But the change is partial and uneven, which means the old disciplines — alt text, transcripts, captions, surrounding context — still matter enormously, even as genuine visual understanding emerges. The practical upshot is a both/and: make your visual content genuinely good and genuinely understandable, using both its inherent qualities and the text that anchors it. This is a guide to making images, video, and audio work for AI visibility in a multimodal world — and to why context and transcripts remain the foundation.

AI is going multimodal

The defining shift is that AI is becoming multimodal: able to understand and reason about images, video, and audio, not just text. Where earlier systems treated non-text content as opaque, the engines increasingly develop genuine understanding of what an image shows or a video contains. This is a significant change for visual content, which is becoming genuinely visible to the machines that determine visibility, rather than being understood only through surrounding text. Understanding that the engines increasingly comprehend visual media directly is the starting point for thinking about multimodal visibility — it means your images and video are, increasingly, content the engines can engage with on their own terms, not just metadata to be read. The multimodal shift opens genuine visibility for visual content that text-only systems never offered.

What multimodal means for visibility

For visibility, multimodal AI means visual content can increasingly be understood, surfaced, and valued on its own merits, while the text that anchors it remains important. A genuinely good, relevant image or video can contribute to how the engines understand and surface your content, rather than being dead weight understood only through captions. But the understanding is still developing and uneven, so the text signals — alt text, transcripts, captions, surrounding context — remain crucial for making visual content legible and reliable to the engines. The practical meaning of multimodal visibility is therefore layered: create genuinely good visual content that the engines can increasingly understand directly, and anchor it with the text signals that ensure it is understood reliably. Both the visual quality and the text anchoring matter.

Images and AI understanding

For images, AI understanding is improving, but the text signals that describe and contextualise them remain essential for reliable comprehension. Alt text — the text description of an image — has always mattered for accessibility and for machine understanding, and it continues to matter as a reliable signal of what an image shows. The surrounding content, captions, and context all help the engines understand an image's meaning and relevance. So while the engines increasingly understand images directly, the disciplined use of accurate alt text, meaningful captions, and relevant surrounding context remains the reliable way to ensure your images are understood correctly. Good image practice — genuine, relevant images with accurate descriptive text — serves both the emerging direct understanding and the established text-based understanding, which is the safe approach.

Video and the transcript imperative

For video, the single most important practice for AI visibility is the transcript, because text remains the most reliable way for the engines to understand and index video content. While the engines increasingly understand video directly, a transcript makes the full content of a video available as text — reliably understandable, indexable, and citable. A video without a transcript locks its content in a form the engines understand less reliably; a video with a good transcript makes all its content available as text the engines can fully process. For any video you want to contribute to your visibility, a transcript is close to essential. The transcript imperative is the clearest practical rule of multimodal content: if you produce video, transcribe it, so its content is available as the text the engines understand most reliably.

Images that get surfaced

Beyond being understood, images can be surfaced and shown by the engines, which makes genuinely good, relevant, original images valuable. Images that genuinely illustrate, explain, or add value — rather than decorative stock imagery — are more likely to be useful to the engines and to the people they serve. Original images, diagrams, and visualisations that genuinely convey information are particularly valuable, because they offer something specific and ownable, as we discuss below. Ensuring your images are genuinely relevant, high-quality, and well-described makes them candidates to be understood and surfaced rather than ignored. The images that work for multimodal visibility are the ones that genuinely add value and are clearly described, not the generic decorative images that fill much of the web and signal nothing.

Original visuals as defensible content

Original visual content — genuine diagrams, data visualisations, and informative graphics — is a form of the defensible, summarization-resistant content we described in our guide to content AI can't easily summarize. A genuinely insightful diagram or data visualisation conveys something specific that the engines can point to but cannot fully replace, and that competitors cannot trivially copy. This makes original visuals a doubly valuable form of content: they contribute to multimodal visibility, and they are defensible assets whose value survives. Investing in genuine original visuals — real diagrams that explain, real visualisations that reveal — rather than generic stock imagery is both a multimodal-visibility and a defensibility strategy. The original visual is content with genuine value, which is exactly what both the engines and the summarization paradox reward.

Diagrams and infographics

Diagrams and infographics are a particularly valuable form of original visual content, because they convey information in a way that is both genuinely useful and ownable. A clear diagram that explains a process, a well-designed infographic that presents data, a visual that makes a complex idea understandable — these add genuine value and are specific to you. As the engines understand visual content better, genuinely informative diagrams and infographics can be understood and surfaced, while also serving human readers. The key, as always, is genuine value: a diagram that actually clarifies, an infographic that actually informs, rather than decorative graphics dressed up as information. Original, genuinely informative diagrams and infographics are among the most valuable visual assets you can create for multimodal visibility, combining usefulness, defensibility, and increasing machine-understandability.

Video as answer content

Video is increasingly a form of answer content, as people turn to video to learn and the engines increasingly understand and surface it. A video that genuinely answers a question — a clear explanation, a genuine demonstration, a useful walkthrough — can be valuable answer content, especially with a transcript making its content available as text. This extends the answer-first thinking to video: a video that clearly and genuinely answers a real question, transcribed so the engines can understand it, can be surfaced as an answer. For topics where video genuinely serves the question better than text — demonstrations, walkthroughs, visual explanations — producing genuinely useful video, transcribed, extends your answer content into a format the engines increasingly understand and people increasingly prefer for certain questions.

Audio, podcasts, and transcripts

Audio content, including podcasts, follows the same transcript imperative: text remains the reliable way for the engines to understand and index audio. A podcast or audio piece without a transcript is largely opaque to the engines; with a transcript, its full content becomes available as text, indexable and citable. For any audio content you want to contribute to your visibility, transcription is essential. Beyond visibility, transcripts make audio content accessible and searchable for humans too. The discipline is the same as for video: produce genuinely valuable audio content, and transcribe it so its content is available as the text the engines understand most reliably. Audio and podcasts can contribute to visibility, but almost entirely through their transcripts, which is why transcription is non-negotiable for audio you want the engines to understand.

Image schema and markup

Structured data has a role in helping the engines understand images, as part of the honest schema we described in our guide to schema that matters. Image-related markup can help the engines understand what an image is and how it relates to the content, formalising the context that the alt text and surrounding content provide. As with all schema, the value is in accurately describing genuine content: markup that helps the engines correctly understand a real, relevant image. Using appropriate, honest image markup is a sensible part of making visual content legible, complementing the alt text and context. Like all schema, it does not substitute for genuinely good, relevant images well-described, but it is a reasonable technical complement that helps the engines understand your genuine visual content.

Multimodal and entity understanding

Multimodal content can contribute to the engines' entity understanding, as images and video associated with your brand and subject feed the picture the engines build. Genuine visual content that clearly relates to your field — diagrams of your subject, video explaining your topic, images of your genuine work — contributes to the engines' understanding of what you are about, complementing the text signals. This connects multimodal content to the entity and knowledge-graph work we have described: your visual content, properly understood and contextualised, is part of how the engines understand your entity. As multimodal understanding improves, the visual dimension of your presence increasingly feeds the engines' model of you, which is one more reason to invest in genuine, relevant, well-described visual content that reinforces what you are about.

Accessibility equals AI-readability

A useful principle is that the practices which make content accessible to people with disabilities are largely the same ones that make it understandable to AI. Alt text, transcripts, captions, clear structure — these serve both accessibility and machine-readability, because both depend on content being available in forms beyond pure visual or audio presentation. This alignment is convenient: doing the accessibility work well simultaneously does the multimodal-visibility work well. Treating accessibility not as a compliance afterthought but as a genuine practice — providing real alt text, real transcripts, real captions — serves your human audience inclusively and your AI visibility simultaneously. The overlap between accessibility and AI-readability means the right thing to do for people is also largely the right thing to do for the engines, which is a rare and welcome alignment.

Video platforms and AI

Video platforms are themselves increasingly part of how people find answers, and the engines draw on video content including from major platforms. This means video published on major platforms, with good titles, descriptions, and transcripts, can contribute to your visibility both on the platform and in the broader engines that draw on it. The disciplines are familiar: genuinely useful video, clearly titled and described, transcribed or captioned, answering real questions. Being present with genuinely valuable video on the platforms where people seek it, properly described and transcribed, extends your multimodal presence into an important channel. As the engines increasingly understand and surface video, a genuine video presence — useful content, well-described and transcribed — becomes a real part of a complete visibility strategy, not a separate silo.

Don't lock content in images only

A critical warning: do not lock important textual content inside images where the engines cannot reliably read it. Text embedded in an image — a chart with crucial numbers only in the graphic, important information presented only as a picture of text — is far less reliably understood than the same content available as actual text. While the engines increasingly read text in images, relying on that for important content is risky. The rule is to ensure any genuinely important information is available as real text, not locked only in images, with the visual as a complement rather than the sole carrier. This is the visual counterpart of the extractability principle: important content should be available in the form the engines read most reliably, which for text content is actual text, not pixels.

Measuring multimodal visibility

Measuring multimodal visibility is harder and less developed than text measurement, but worth attending to. The signals include whether your images and video are being surfaced and understood, whether your visual content appears in relevant results, and whether your transcribed video and audio contribute to your presence. This is less precise than text measurement, reflecting the earlier stage of multimodal understanding, so it should be held with appropriate humility. But attending to whether your genuine visual content is being understood and surfaced — and ensuring the foundations (alt text, transcripts, context) are in place — is part of a complete approach. As multimodal understanding and measurement mature, this dimension will become more measurable; for now, doing the fundamentals well and attending to whether visual content is surfacing is the sensible approach.

Common multimodal mistakes

The recurring mistakes undermine visual visibility. Treating images as decoration rather than genuine content, with no descriptive text. Publishing video and audio without transcripts, leaving their content opaque. Locking important textual information inside images where it is unreliably read. Neglecting alt text, captions, and context. Using generic stock imagery that signals nothing rather than genuine, relevant visuals. And assuming multimodal understanding is complete and skipping the text anchoring that remains essential. Each wastes the potential of visual content or leaves it unreliably understood. The fix is to produce genuine, relevant visual content and anchor it with the text signals — alt text, transcripts, captions, context — that make it reliably legible, while never locking important content in images only.

A practical approach

As an approach, multimodal content for AI visibility combines genuine visual quality with reliable text anchoring. Produce genuine, relevant, valuable visual content — original diagrams, data visualisations, useful video — rather than decoration. Anchor every image with accurate alt text and meaningful context. Transcribe every video and audio piece, so its content is available as text. Use honest image markup where appropriate. Never lock important textual content inside images only. Treat accessibility as genuine practice, which simultaneously serves AI-readability. Be present with genuine, well-described video on the platforms where people seek it. And attend to whether your visual content is being understood and surfaced. Done this way, your multimodal content contributes genuinely to your visibility as the engines increasingly understand visual media, while remaining reliably legible through the text that anchors it.

The bottom line

AI is going multimodal — increasingly understanding images and video, not just text — which makes visual content genuinely visible to the engines for the first time, while the text that anchors it remains essential. The practical approach is both/and: produce genuine, valuable visual content (original diagrams, data visualisations, useful video) that the engines can increasingly understand directly, and anchor it reliably with the text signals (alt text, transcripts, captions, context) that ensure it is understood correctly. Transcribe all video and audio, because text remains the most reliable way for the engines to understand them. Never lock important content inside images only. Treat original visuals as the defensible, summarization-resistant assets they are. And remember that the accessibility practices which serve people also serve the engines. Done this way, your images, video, and audio become genuine contributors to your visibility in a multimodal world — understood both on their own merits and through the text that makes them legible.

Frequently asked questions

Does AI understand images and video now?

Increasingly, yes — the engines are becoming multimodal, able to understand and reason about images, video and audio, not just text. But the understanding is still developing and uneven, so the text signals that anchor visual content — alt text, transcripts, captions, surrounding context — remain essential for reliable comprehension. Treat it as both/and: genuine visual quality plus reliable text anchoring.

Do I need transcripts for video and audio?

Yes — it's close to essential. While the engines increasingly understand video and audio directly, a transcript makes the full content available as text, which remains the most reliable form for the engines to understand, index and cite. A video or podcast without a transcript locks its content in a less reliably understood form. Transcribe everything you want the engines to understand.

How do I make images work for AI visibility?

Use genuine, relevant, valuable images (original diagrams and visualisations, not decorative stock), and anchor each with accurate alt text, meaningful captions, and relevant surrounding context. Appropriate honest image markup helps too. And never lock important textual information inside an image where it's unreliably read — keep important content as real text, with the visual as a complement.

Are original visuals good for AI visibility?

Yes — genuine diagrams, data visualisations and informative graphics are defensible, summarization-resistant content that AI can point to but not fully replace, and that competitors can't trivially copy. They contribute to multimodal visibility, serve human readers, and become increasingly machine-understandable. Investing in original informative visuals beats generic stock imagery on every dimension.

Why shouldn't I put important text inside images?

Because text embedded in an image — crucial numbers only in a chart graphic, important information only as a picture of text — is far less reliably understood by the engines than the same content as actual text. While they increasingly read text in images, relying on that for important content is risky. Keep genuinely important information available as real text, with visuals as a complement.