How do images actually appear in AI answers?
Two routes. First, the text route: retrieval systems index the words around an image — alt text, caption, nearby paragraphs — and those words decide whether the visual surfaces as a supporting card or gets described in the answer. Second, the multimodal route: models like Gemini 3 can interpret images directly when they are part of a retrieved page, but they still rely on your labels to know what they are looking at and who it belongs to. Both routes reward the same discipline, detailed in multimodal GEO: no unlabelled image ever earns a citation.
What does the 2026 context demand?
AI Overviews serves more than 2.5 billion monthly users (Alphabet, Q2 2026) and the Gemini app over 750 million monthly users, both of which render visual answers with image cards. With roughly 68% of search sessions ending zero-click (Ahrefs, February 2026), the image a user sees inside the answer may be their only contact with your brand. If that image is a competitor's chart because yours had no alt text, you lost an impression you never knew existed.
How do you write alt text and captions for retrieval?
- Alt text: describe the content, not the aesthetics. Name the subject, the data and the takeaway: “Bar chart of median response times by vendor, 2026, Vendor C fastest at 41 ms” beats “chart”.
- Caption: state the insight. Captions are read more often than body copy, and engines extract them as standalone passages. Put the number and the date in the caption.
- Filename: descriptive and stable. vendor-response-times-2026.png helps; IMG_4021.png does not.
- Surrounding copy: co-locate the claim. The paragraph beside the image should make the same point in words, so text-only retrieval still captures the insight.
Which images deserve the most investment?
Original charts from your own data are the highest-value visual asset, because they can only be sourced to you — the visual counterpart of original research. Product diagrams, comparison matrices and process flows also extract well. Decorative stock photography adds nothing to retrieval and can be deprioritized. Every chart should have an HTML table twin beneath it: engines that cannot parse the pixels still cite the numbers.
Why does licensing matter to AI engines?
Engines prefer images they can legally surface. Mark your originals with clear licence information — ImageObject structured data with a licence field, plus a visible credit line — so answer engines can attribute and display them confidently. Conversely, hotlinked or ambiguously licensed images are often excluded from visual answers. If you want your diagrams quoted with attribution, say so explicitly: an attribution-request licence line turns every reuse into a brand mention, compounding the unlinked-mention value described in unlinked mentions.
What technical markup helps?
Use ImageObject structured data with contentUrl, creator, license and caption fields; keep images on crawlable URLs rather than behind scripts; and serve modern formats with explicit width and height so rendering is reliable. See structured data for AI for the markup patterns that define passage and asset boundaries.
What are the pitfalls?
The most common failure is charts that exist only inside images with no text equivalent — the insight dies with the pixels. Second is generic alt text written for compliance rather than comprehension: “image” and “graphic” describe nothing. Third is unsigned original work: if your watermark and metadata do not identify you, engines that display the chart cannot credit you, and you lose the citation even when you win the impression. Finally, teams forget localization — a chart with English-only labels underperforms in multilingual answers, which matters if you operate across markets per multilingual GEO.
Audit quarterly: pick your twenty most important images, check alt text, caption, licence and table twin, then test visual questions in your fixed measurement set and screenshot which images the engines actually show.