Multimodal GEO: Images and Videos in AI Answers

AI engines no longer just cite text. Google AI Mode, available in France since July 22, 2026 (Abondance, see also our article on Google AI Overviews), generates answers with images, videos and interactive maps. ChatGPT now displays visuals in its answers. Perplexity embeds YouTube videos directly in its summaries. Multimodal GEO means optimizing your visual content so that it is selected and displayed in AI answers, not just your text.

Why multimodal changes the game in 2026

Three major developments have accelerated the trend:

• Google I/O, May 2026: Google Search is "completely reimagined with AI". Users can attach images, documents and videos to their searches. According to Google, AI Mode queries are on average 3 times longer than traditional searches, and increasingly include visual context.
• YouTube has become the #1 source cited in AI Overviews: according to BrightEdge (September 2025), YouTube accounts for 29.5% of citations in Google AI Overviews, far ahead of the second most-cited domain (Mayo Clinic, 12.5%). Videos are no longer a complement; they are a main source.
• ChatGPT generates and displays visuals: GPT-4o, which introduced native image generation, was retired from ChatGPT on February 13, 2026 (OpenAI), but its successors kept these capabilities: ChatGPT can generate images in its answers and include visuals from the web when relevant.

The direct consequence: a text-only website misses out on a fast-growing surface of AI visibility. Text is still the foundation, though: captions, transcripts and descriptions follow the same rules as in our guide to sourcing your content to get cited by AI.

How AI engines select visual content

AI engines don't "see" your images the way a human does. They rely on metadata and textual signals to understand and select visuals:

1. The image alt attribute

Alt text matters as much for GEO as meta descriptions do for traditional SEO: it's the text AI engines read to understand your visual.

It's the #1 signal. An empty alt="" or an alt="IMG_4523" makes your image invisible to AI. An alt="Comparison table of home insurance rates in France 2026" makes your image citable. The alt text should describe what the image shows and why it is relevant in the context of the page.

2. The text surrounding the image

RAG systems extract the text surrounding an image to understand its relevance. An image placed after a descriptive H2 with an explanatory paragraph will be better understood than an isolated image with no context. The ideal pattern:

• Descriptive H2 (e.g. "Price comparison by region")
• Context paragraph (30-50 words)
• Image with descriptive alt text
• Caption or <figcaption> adding extra information

3. ImageObject and VideoObject schemas

The ImageObject schema in JSON-LD structures an image's metadata (author, date, description, license). The VideoObject schema does the same for videos. These schemas help engines read your media, but they are not a condition for appearing in AI answers: for its AI Overviews and AI Mode, Google states that no special markup is required. The text around the image or video (alt text, caption, transcript) remains what matters most. To learn more about implementing schemas, see our Schema.org guide for AI visibility.

4. Format and technical performance

Heavy images slow the page down, which degrades Core Web Vitals. AI engines favor fast pages. Use modern formats (WebP, AVIF), lazy loading, and explicit dimensions (width and height) to avoid Cumulative Layout Shift (CLS).

YouTube: the most powerful multimodal channel

YouTube is now the leading source of citations in Google AI Overviews (29.5%, BrightEdge). It's also a major source for Perplexity, which embeds videos directly in its answers.

Why AI engines cite YouTube so much

• Automatic transcripts: every YouTube video generates a text transcript that AI engines can index and extract. Your 10-minute video becomes a 2,000-word text document, rich in citable information. It's the same principle as long-form content performing better for AI citability.
• Trust signals: view count, likes, comments, channel age. These signals help AI engines assess the credibility of the content.
• Built-in VideoObject schema: YouTube automatically generates the structured metadata AI engines use to understand a video's content.
• Timestamps and chapters: YouTube chapters let AI engines cite a specific moment in the video, not just the whole video.

Optimizing your videos for GEO

1. Question-style titles: "How do you choose home insurance in 2026?" rather than "Our insurance offer"
2. Rich description: the first 200 characters of the YouTube description matter most. Put a 30-50-word answer capsule there that answers the title's question.
3. Chapters with timestamps: split your video into logical sections. Each chapter = a separate citation opportunity.
4. Reviewed transcript: YouTube's automatic transcript is often approximate. Correct it manually so technical terms are accurate.
5. Embed videos on your site: a YouTube video embedded on your page adds a multimodal signal that AI engines detect (YouTube iframe).

Images and infographics: GEO best practices

The types of images AI engines cite most

• Comparison tables: AI engines love structured comparisons. An HTML table with figures will be extracted as-is.
• Infographics with sourced data: an infographic summarizing key statistics with their sources is highly citable. But the infographic alone isn't enough: the equivalent text must be on the page (AI engines don't "read" images).
• Screenshots: for tutorials and technical guides, screenshots with text annotations (descriptive alt text + captions) are very useful.
• Charts with data: a chart showing a trend is citable if the surrounding text explains the data. The chart alone is invisible to AI.

The "all image" trap

Some sites, especially in e-commerce and restaurants, put most of their content in images (menus as images, product sheets as images, prices as images). AI engines cannot read text inside an image. If your menu, prices or specifications only exist in visuals, they don't exist for ChatGPT, Perplexity or Gemini.

The rule: every important piece of information must exist as text on the page. Images illustrate and enrich; they don't replace text.

Podcasts and audio content

Podcasts are a growing format for AI visibility, mainly thanks to transcripts:

• Publish the full transcript of each episode on your site. It's long, rich content with expert quotes and hard data, exactly what AI engines value.
• Structure the transcript with an H2 per topic covered, timestamps, and links to the sources mentioned.
• Add a PodcastEpisode schema in JSON-LD to structure the metadata (guests, topic, duration, date).

How Detekia evaluates multimodal signals

Detekia's free GEO score takes several multimodal-related signals into account in its analysis:

• Content depth criterion (15 pts): the word count of the main content is measured. An "all image" page with no text will score low, which signals an AI citability problem.
• Answer structure criterion (40 pts): text answers under headings and data tables are what count. Information locked inside an image or a video is not read as text.

A 4-week multimodal action plan

Week 1: audit your existing images

Review every image on your main pages. Check that each image has descriptive alt text (not just the file name), that the surrounding text provides enough context, and that images don't contain important textual information with no HTML equivalent.

Week 2: optimize your YouTube videos

Add chapters to your existing videos, correct automatic transcripts, rephrase titles as questions, and enrich descriptions with answer capsules.

Week 3: create targeted multimodal content

Create an HTML comparison table (not an image) on your main page, embed your best YouTube video, and add a VideoObject schema if you host videos on your site.

Week 4: measure and iterate

Run Detekia's free GEO score again to measure the impact of your optimizations on your score. Compare it with your initial score and identify the next actions. The Detekia score evaluates your page, not AI answers. To check whether your videos and visuals show up in AI answers, Beeleven offers an AI visibility audit covering ChatGPT, Gemini, Claude and Perplexity.

Key takeaways

Multimodal GEO is not a "nice to have". With YouTube at 29.5% of AI Overview citations (BrightEdge, 2025) and AI engines displaying more and more visuals in their answers, ignoring multimodal means missing out on a growing share of the AI visibility surface. The three most impactful actions: descriptive alt text on all your images, a YouTube presence with optimized videos (question titles, chapters, transcripts), and text that accompanies every visual. AI engines don't see your images; they read what you write around them.