What Image and Video Formats Can AI Engines Understand?

Which formats does each system support?

The point to take from the table is that JPEG, PNG and WebP work everywhere images are accepted, and MP4 or WebM work wherever video is accepted; Claude takes images but not video files. Newer formats such as AVIF are fine for Google Search. Use a picture element with a JPEG or PNG fallback if you want every system and browser to get a usable file.

What do AI search systems actually read from your images?

There is a difference between what a model can do with an image you hand it and what a search system does with images on the web. Google says it uses alt text, computer vision and the content of the page to understand what an image shows. It also draws on captions, image titles and, lightly, the filename. It does not index images set as CSS backgrounds.

The AI assistants' own crawlers do fetch images. In a 2024 study of crawler traffic to one large site, Vercel and MERJ found images made up about 35% of Claude's crawler fetches, while ChatGPT's crawler spent most of its fetches on HTML. None of the companies documents how, or whether, image content feeds into answers. Treat the words around an image as what counts, and the pixels as a bonus.

What that means in practice:

• Alt text that says what the image shows and why it matters: "Line chart of heat pump installations in [region], 2019 to 2025, rising from 30,000 to 98,000 a year", not "chart".
• Captions under important images, since captions are read by people and machines alike.
• Nearby copy that states the point the image makes.
• Descriptive filenames such as heat-pump-installations-uk-2019-2025.png rather than IMG_4412.png.
• Standard img elements in the HTML, not images injected by script or set as backgrounds.

Can AI read text inside images?

Models can read text in an image they are given, and Google's image systems use computer vision. But search pipelines that retrieve passages to answer questions work from page text. A price list, a specification sheet or a chart saved only as an image is at best a weak signal. Put the facts in HTML as well:

• Charts: add a short summary of the finding and a data table below.
• Infographics: repeat the key figures in the text.
• Screenshots of settings or steps: write the steps out.
• Menus, price lists and timetables: publish them as text, not only as images or PDFs of scans.

This also helps people using screen readers, which is the original reason for alt text.

What do AI search systems read from your videos?

For video, the transcript is what matters most. Google's systems can process many video file types. But its video indexing needs a watch page, a page whose main purpose is showing a single video, and that page must itself be indexed. The video also needs a stable thumbnail URL. The AI assistants' crawlers generally do not render JavaScript. So a video player loaded by script, with no text on the page, gives them only the page title.

Publish each important video with:

• A dedicated watch page with a descriptive title and a summary of what the video covers.
• A full transcript in the page's HTML, or at least a detailed summary with the key points and figures.
• Chapters with timestamps and headings, which also help Google show key moments.
• VideoObject structured data with name, description, thumbnail, upload date and duration, as covered in what video structured data does for SEO and AI visibility.

YouTube is a special case. Google's AI features cite YouTube often, and Gemini can work with public YouTube links directly. Captions, titles, descriptions and chapters on YouTube are therefore worth the same care as your own pages. How to prepare YouTube videos for AI answers is covered in how to optimise YouTube videos so AI answers cite them.

Why does this matter more as search becomes visual?

People increasingly search with images. Google said in May 2026 that more than one in six searches in AI Mode in the US used voice or images, with image searches growing quickly. A shopper photographs a product and asks where to buy it. A homeowner photographs a damaged roof tile and asks what it is. Systems answering those questions match the photo against images they have indexed and the text around them. Clear product photos from several angles, each with accurate alt text and product names nearby, make that match more likely. For online shops, the images in your product feed matter as much as those on the page. Whether images and videos can appear inside Google's AI answers is covered in whether images and videos can appear in Google AI Overviews.

Which format choices still matter?

• Speed. WebP and AVIF are smaller than JPEG and PNG for the same quality, which helps page performance. Keep a fallback in a picture element.
• SVG for diagrams and logos. It stays sharp at any size, and text inside SVG is real text rather than pixels.
• Avoid animated GIFs for information. Claude uses only the first frame of an animation, and GIFs are large. Use a short MP4 or WebM with a text summary.
• Stable URLs. Changing image and thumbnail URLs often can stop them being indexed, as Google notes for video thumbnails.
• Matching extensions. Google recommends that the file extension matches the actual file type.

What might an illustrative case look like?

Consider an illustrative garden equipment retailer whose buying guides relied on comparison charts saved as images and on how-to videos embedded without text. AI answers about choosing a lawnmower cited competitors' text-based guides instead. The retailer added a data table and a two-sentence summary below each chart and rewrote its alt text. It gave each video a watch page with a transcript and chapters. Within two months, AI answers began quoting figures from its comparison tables. Its videos started appearing in Google results for questions they answered.

What do people ask about images, video and AI?

Is WebP or AVIF better for AI visibility?

Neither format affects visibility directly. Both are supported by Google Search. Choose for speed, and keep a JPEG or PNG fallback.

Do AI engines watch videos on my site?

Rarely in a way you can rely on. They depend mainly on transcripts, titles, descriptions and structured data.

Should every image have alt text?

Every meaningful image should. Purely decorative images can use empty alt text so screen readers skip them.

Are PDFs with images read by AI?

Text in PDFs can be read. Scanned pages without a text layer are much harder to use, so publish key content as HTML.