crawlers

Images, alt text, and what AI crawlers do with them

Text inside images doesn't reach a text extractor. What alt text, captions, and HTML tables carry, what the vendors document about images, and how to audit it.

Text inside an image doesn’t reach a system that reads your page as text. A price table exported as a PNG, a chart with the numbers baked into the pixels, a heading set in a banner graphic: an extractor that turns HTML into text sees an <img> tag and, at best, its alt attribute. Whatever the image says is gone unless the page also says it in words.

That’s the whole problem, and most of the fix is old accessibility practice. Write alt text that describes what the image shows, put captions in the HTML, and give any image that carries data a text equivalent on the page. What’s new is that the reader you’re writing for now includes retrieval systems, and they’re less forgiving than a person who can at least look at the picture.

This post covers what is and isn’t documented about how AI crawlers handle images, what each piece of markup carries, and how to find image-locked content across a site without a terminal.

What do AI crawlers do with images?

For the assistants other than Google’s, it’s undocumented. OpenAI’s crawler documentation lists no image-specific crawler, and neither it nor Anthropic’s or Perplexity’s equivalents say whether images are fetched, interpreted, or ignored when a page is used to build an answer. Some assistants accept images that a user uploads. That’s a different path from the crawl, and nothing published says it applies to images on pages they retrieve.

Google documents more. Its image SEO best practices say Google “uses alt text along with computer vision algorithms and the contents of the page” to understand an image, and that it “extracts information about the subject matter of the image from the content of the page.” It also says Google “doesn’t index CSS images”: an image set as a CSS background isn’t found at all. Images are found through the src attribute of an <img> element.

Google also runs a separate crawler for images. Its common crawlers documentation lists Googlebot-Image, and says crawl preferences for that token affect Google Images, Discover, Google Video, and features across Google Search where images, logos, and favicons appear. A robots.txt rule that disallows Googlebot-Image reaches further than Image search.

The working assumption follows from what is and isn’t written down. Plan for an extractor that sees your text, your markup, and maybe your alt attributes, and nothing inside the image. If an assistant does interpret images, text on the page costs nothing. If it doesn’t, the text is the only route.

Where the information livesReaches a text extractor?Notes
Body text, HTML table, listYesThe reliable path
<figcaption>YesIt’s body text inside a figure element
alt attributeUndocumented for AI assistantsGoogle documents using it; extractors vary
title attribute on <img>UndocumentedDon’t rely on it
Text rendered inside the imageNo, unless the system interprets imagesNo vendor documents this for crawled pages
CSS background imageNoGoogle doesn’t index CSS images

What should alt text say?

Alt text should say what the image contributes to the page, in one sentence, as if the image failed to load. Google’s guidance asks for “useful, information-rich content that uses keywords appropriately” and warns against “filling alt attributes with keywords.” The accessibility test and the extraction test give the same answer: would a reader who can’t see the image lose anything?

Three cases cover most images:

  • Informative images. A product photo, a screenshot, a diagram of a process. Describe the content that matters: “ARM-X9 robotic arm mounted on a workbench, gripper closed around a 50 mm cube.” Skip “image of.”
  • Decorative images. Background textures, spacer graphics, a stock photo that only sets mood. Use an empty alt="" so screen readers skip it. An empty alt is a deliberate choice, not a missing one.
  • Images of text. A heading in a banner, a quote set as a graphic, a pricing card exported as an image. The fix is to make it real text. Alt text that repeats the words is the fallback when you can’t change the template.

Alt text is short by design. If the image holds more than a sentence can carry, it’s a complex image, and it needs a text equivalent on the page as well.

How do you handle charts and figures with data?

Give every chart or figure that carries data a text equivalent on the page: a sentence stating the finding, plus the underlying numbers as an HTML table. The W3C’s tutorial on complex images describes this as a two-part text alternative, a short description that identifies the image and a long description with “the essential information,” and recommends making long descriptions available to everyone.

The pattern that works for both readers:

<figure>
  <img src="/img/response-times-2026.png"
       alt="Line chart of median API response time by month, January to June 2026.">
  <figcaption>
    Median response time fell from 420 ms in January to 180 ms in June
    after the cache rollout in March. Monthly values are in the table below.
  </figcaption>
</figure>
<table>
  <caption>Median API response time by month, 2026 (illustrative data)</caption>
  ...
</table>

The caption gives the claim in one sentence, which is the passage an answer can lift. The table gives the numbers in a form that survives extraction; Tables and lists that survive extraction covers how to build it, and Numbers, units and dates that survive extraction covers how to write the values. The alt text identifies the chart and leaves the data to the page.

The same rule applies to pricing. A pricing grid rendered as an image is one of the most common ways a page hides its most-asked answer; Pricing pages on answer surfaces treats that case in detail.

Where does image-locked text hide?

Image-locked text tends to come from a handful of templates, not from individual pages. Check these first:

  1. Hero banners with the page’s real headline baked into the graphic and a generic H1 in the HTML.
  2. Pricing and comparison grids exported from a design tool.
  3. Infographics and charts in blog posts and reports, with no caption or table.
  4. Screenshots of documentation, such as configuration panels or code, where the settings exist only as pixels.
  5. Quote cards and testimonials set as images for social sharing and reused on the page.
  6. Spec sheets where the product’s key figures exist in a downloadable image or PDF and nowhere in the HTML.

A related case is text that exists in HTML but is only shown after a script runs. That’s a rendering problem rather than an image problem; JavaScript content and what never makes the trip covers it.

How do you check one page by hand?

Turn images off and read the page. Every browser can do it through its settings or a developer-tools option, and a reader view gets close. Whatever information disappears when the images go was image-locked.

Then view source and search for <img. For each one, look at the alt attribute: present and descriptive, present and empty on purpose, or missing. Search for background-image in the inline styles too, since those images aren’t indexed by Google at all. Finally, check that each chart has a <figcaption> or adjacent paragraph and, where it carries numbers, a table.

What do AI crawlers extract from your page? walks through the rest of the page’s elements the same way.

How do you audit images across a whole site?

Start with a crawler’s image report, then sample by template. Most desktop crawlers list every image with its alt attribute, so a filter for missing alt gives you the first list in a few minutes, and you can export it to a spreadsheet and pivot it by the page’s directory to see which templates produce the problem.

Missing alt is the easy part. Image-locked text doesn’t show up as a missing attribute, so for that:

  • Custom extraction. If your crawler supports it, extract the count of <figure> elements without a <figcaption>, and the count of images inside elements you know hold data (a pricing section, a chart container). Pages with a high count go on the review list.
  • Word count by template. In the spreadsheet, sort pages of the same template by word count. A pricing or product page with far fewer words than its siblings is often carrying its content in images.
  • Sample by eye. Open three to five URLs from each template with images turned off. Template problems repeat, so a small sample finds most of them.

If you’re comfortable in a terminal, this lists <img> tags with no alt attribute on one page:

curl -s https://example.com/pricing/ | grep -o '<img[^>]*>' | grep -v 'alt='

It won’t catch alt attributes that are present but useless, and it won’t find text inside images. Those still need a person looking at the page.

In GEOBee, image alt text is one of the scored SEO signals (8 of 100 points, listed on the three lenses), and each page’s detail panel has an Images tab for inspecting the evidence behind that score. The scan checks attributes, not pixels: it can tell you an image has no alt text, but not whether the picture contains a price.

What about robots.txt and image directories?

Check that a robots.txt rule meant for one purpose isn’t blocking images you want indexed. Disallowing Googlebot-Image affects images across Google Search features, per Google’s documentation, not only Image search. A blanket disallow on /assets/ or /wp-content/uploads/ written years ago can remove product images from every surface that shows them. robots.txt patterns for AI covers how to scope rules by agent and path.

For the AI crawlers, there’s no documented image-specific token to manage, so the decision is the same as for any other path: if a crawler is allowed on the page and you want its images treated the same way, don’t disallow the image directory for that crawler. The AI crawler access guide sets out which crawlers matter for which surface.

None of this makes an image cited. It makes sure the information the image carries is also available as text, which is the only form every reader of the page is documented to handle. If you want missing alt text and thin pages flagged across a whole crawl, download GEOBee and run the 7-day trial on your own site.

On this page