TechnicalAI SEOEntity SEO

Image SEO for AI Search: How Multimodal Models Read Your Visuals

Alt text used to be the whole conversation about images and search: describe the picture, include a keyword, move on. Multimodal models changed the underlying situation. They process the image itself. They can tell you what is in a photograph, read the text inside a screenshot, and interpret a chart, without ever seeing your alt attribute.

That does not make alt text obsolete. It changes what it is for.

What multimodal models can and cannot get from an image

They can determine content. Objects, scenes, layout, colours, and text rendered inside the image, including words in a screenshot, a diagram label, or a slide.

They cannot determine identity or provenance. A model can see a modern office building. It cannot know it is your office, in which city, photographed by whom, or that the chart in it is your original research rather than someone else’s.

What a multimodal model can and cannot read from an image An annotated diagram of a chart image on a page. The model can read the shapes, the colours and any text baked into the image, including an outdated price label, which is flagged as a risk. It cannot determine whose chart it is, when it was made, or whether the data is original. Alt text, the visible caption and ImageObject markup supply that missing identity and provenance layer. The model sees the picture. It cannot see whose it is. What pixels give away, and what only your text can supply Q3 pricing update starts at $3,500 text baked into the image is now read READSshapes, colours, layout no alt text needed READS"$3,500" — even if that price is stale an outdated graphic is now a contradictory fact about you CANNOT TELLwhose chart this is, or when identity and provenance are invisible in pixels ALT TEXT + CAPTION names the entities and dates the claim IMAGEOBJECT attributes the asset to your Organization Alt text stopped being a description for a machine that cannot see. It is now the fact layer a machine cannot infer.

Scroll the diagram sideways to see all of it.

How to read this: an illustrative annotation of multimodal image handling. The figures on the mock chart are placeholders. The behaviour shown, that in-image text is read and that identity is not inferable, follows from how multimodal models process visual input.

That gap is the entire job of the text around the image. You are no longer describing the picture for a machine that cannot see. You are supplying the facts a machine cannot infer from looking.

What this changes in practice

ElementOld jobNew job
Alt textDescribe the visual for accessibility and crawlersName the entities and supply context the pixels cannot carry
FilenameKeyword signalWeak but free context signal
CaptionOptional design elementHigh-value, visible, extractable context
Surrounding copyLoosely relatedThe primary source of the image’s meaning
ImageObject markupRarely bothered withAttaches the asset to your entity graph
Text inside imagesInvisible to searchReadable, and now a liability if wrong

That last row deserves attention. Text baked into images used to be simply unreadable. Now it is read. Which means an outdated price on a graphic, an old company name on a slide, or a stale statistic in an infographic is no longer harmlessly invisible. It is a contradictory fact about your brand, sitting on your own site, undermining the consistency that entity SEO depends on.

The practical checklist

1. Write alt text that names things

Not “team photo” and not “web design agency team photo Austin.” Write “Olga Kunger, founder of Ambeltek, at the studio’s Austin workspace.” Named entities, real relationships, plain language. This serves screen reader users properly and gives the model the identity layer it cannot see.

2. Use visible captions for anything that carries a claim

Captions are text on the page, so they chunk and get retrieved like any other passage. For charts, screenshots and data visuals, a caption stating what the image shows, the source and the date makes the visual quotable. An uncaptioned chart is a picture; a captioned one is evidence.

3. Audit text inside your images

Go through your graphics for outdated prices, old branding, superseded statistics and wrong dates. Anything contradicting your current site text is now actively read. Fix or replace.

4. Add ImageObject where it matters

For original photography, diagrams and research charts, ImageObject with caption, creator pointing to your Organization or author Person, contentUrl and licence information attributes the asset to your entity. This slots into the connected graph described in Organization schema for AI search and is part of our schema markup service.

5. Put the image near the text it supports

Retrieval works on passages. An image separated from the paragraph explaining it loses the context that makes it meaningful. Keep visual and explanation in the same section.

6. Replace stock photography with something specific

Generic stock contributes nothing a model can associate with you, and it looks the same as every competitor using the same library. Original photography, real screenshots of your work, and custom diagrams are all entity signals. Stock is not.

7. Do not let images break the technical basics

Descriptive filenames, correct dimensions, modern formats, lazy loading below the fold, and images that do not tank your Core Web Vitals. Speed still gates everything.

Where original visuals actually earn citations

The highest-value images for AI visibility are the ones that carry information nobody else has:

  • Original data charts. A model summarizing a topic wants a source for the number. If your chart is the source, you get named.
  • Process diagrams. A clear diagram of how something works is frequently the most quotable asset on a page.
  • Real screenshots. Concrete, verifiable, and impossible for a competitor to duplicate.
  • Comparison visuals. Well-structured, extractable, and matched to the comparison queries that fan-out generates.

These are the same properties that make text citable: specific, attributable, self-contained. Images are not a separate optimization discipline. They are the same discipline applied to a different asset type, which is the point made throughout our complete guide to generative engine optimization.

Where this leaves you

Stop thinking of alt text as the accessibility box you tick and start thinking of the whole image context as fact supply. The model sees the picture; you provide the names, dates, sources and relationships it cannot see. Then audit what your existing images are silently asserting, because whatever text is baked into them is now part of what AI systems believe about your brand.

If you want your visual assets audited alongside the rest of your entity signals, that is included in our AI SEO service.

Frequently asked questions

Do AI models actually look at images or just read alt text?

Multimodal models process the image itself, so they can describe what is in a picture without any alt text at all. Alt text, captions and surrounding copy still matter because they supply the context an image cannot carry on its own, such as who or what it depicts by name.

Is alt text still worth writing for AI search?

Yes, for two reasons. It remains a legal and accessibility requirement, and it supplies the naming and context that visual analysis cannot infer. A model can see a building; only your text tells it which building and whose.

What is ImageObject schema for?

ImageObject lets you attach structured facts to an image: its caption, creator, licence, and the URL it represents. It connects a visual asset to your entity graph so an image can be attributed to your organization rather than floating unattributed.

Do images help my brand get cited in AI answers?

Indirectly. Distinctive original imagery strengthens entity recognition and gives visual surfaces something to retrieve, while generic stock photography contributes nothing a model can associate specifically with your brand.

OK

Olga Kunger

Founder & Lead Strategist, Ambeltek

Olga leads Ambeltek's web development, AI SEO, and GEO work — helping brands rank on Google and get cited by AI engines. More about Olga →

Ready to get found on Google and in AI?

Tell us about your project and we’ll send a free visibility audit plus a tailored proposal, usually within one business day.

Start a project