Alt Text and Multimodal Answers: When Images Get Read

This post is about what images can realistically contribute right now, and what still depends on the text around them.

What crawlers actually do with images

Most AI crawlers fetch HTML first and images rarely or not at all. When an engine answers a question, it is usually working from the text of the page, including alt attributes, captions, and surrounding paragraphs. Whether the image itself was ever downloaded varies by engine.

In our testing, the text around an image tends to do more work than the image. A product photo with the alt text "Acme Model 3 desk lamp, matte black, 40 cm tall" contributes those facts to the answer. The same photo with alt text "image1.jpg" contributes nothing, regardless of how capable the engine's vision model is.

What alt text should say

Alt text was designed for screen readers, and that remains the right way to write it: describe what the image shows, in plain language, for someone who cannot see it. That description happens to be exactly what a text-first engine needs.

Good alt text names the subject, the relevant attributes, and the context. "Comparison table of Acme and Beta pricing tiers, showing Acme starts at $29 per month and Beta at $49" is useful. "Pricing comparison" is thin. "Acme pricing best value affordable SaaS" is keyword stuffing and reads as such.

Keep it to a sentence or two. If an image needs a paragraph, that paragraph belongs in the visible text.

Captions and surrounding text

A visible caption in a figcaption element is read by every crawler and seen by every human. For charts, diagrams, and screenshots, the caption should state the one fact the image is there to show. "Response time dropped from 800 ms to 120 ms after the migration" is a sentence an engine can quote. The chart is supporting evidence a human can check.

The paragraph before and after an image also matters. Engines tend to associate the image with the text nearest it, so put the image next to the claim it supports, not three paragraphs away.

Images that contain text

Infographics, screenshots of tables, and images of quotes are the worst case. The information is there, visually, and absent from the text layer. If an engine does download and read the image, it may extract some of it. If it does not, the content is gone.

The fix is the same as for PDFs: put the content in HTML. A table should be an HTML table. A quote should be a blockquote. An infographic should have its key figures stated in the text below it. The image can stay as a visual, but it should never be the only copy of a fact.

Image file names and structured data

File names are a small signal and cost nothing. acme-model-3-matte-black.jpg beats IMG_4821.jpg. ImageObject schema with a caption gives a crawler one more machine-readable copy of what the image shows. Neither is a priority over alt text and captions, but both are easy.

Where multimodal answers show up

Shopping queries and "what does X look like" queries are where image-aware answers appear most. For those, a clear, well-described product image on a crawlable page is worth doing. For a B2B service page, images are rarely the reason you get cited, and the time is better spent on text.

How we treat images in an audit

At AEO GEO Labs, images are part of the content review rather than a separate workstream. We flag images with missing or useless alt text on priority pages, identify any facts that exist only inside an image, and move those facts into the HTML. The goal is that nothing an engine should know about you is locked in a format it might not read.

A fast check

Open your most important product or pricing page and disable images in your browser. Read what is left. Every fact that disappeared is a fact an engine may never see. Write it back in as alt text, a caption, or a sentence. The free on-page SEO checker at AEO GEO Labs flags missing alt attributes on a page, which is a reasonable place to start the list.