Making Images and Diagrams Legible to Generative Systems

Most images on the web carry almost no retrievable information, not because images are inherently opaque to generative systems, but because almost nothing around them was built to make their content legible. Alt text gets filled in as an afterthought or skipped entirely. Filenames stay as the camera or export tool left them. Captions, when they exist at all, restate what a human viewer can already see rather than adding anything a system reading only the surrounding text would otherwise miss. This sub-chapter covers what actually changes that: treating an image's descriptive context as real content, not accessibility overhead.

Alt Text as Description, Not a Keyword Opportunity

Alt text exists to describe what an image actually shows, to a person who can’t see it and to a system reading the page’s underlying markup rather than rendering it visually. Treating alt text as a place to work in a target keyword phrase, regardless of whether that phrase accurately describes the image, produces text that fails both audiences at once: a screen reader user gets an inaccurate description, and a generative system reading that text gets a false signal about what the image actually contains.

The correct standard is simple to state and genuinely effortful to apply consistently across a large image library: write what the image shows, specifically enough that someone who can’t see it would understand its actual content. A product photo’s alt text should name the product and its relevant visible attributes. A diagram’s alt text should describe what relationship or process it depicts, not just label it generically as “diagram.” This isn’t a technical SEO checkbox distinct from the framework’s broader anti-manipulation stance; it’s the same honest-description principle applied to a different field on the page.

Descriptive Filenames: A Small, Frequently-Skipped Signal

A filename like IMG_4471.jpg carries no information. A filename like wide-toe-box-running-shoe-side-profile.jpg carries real, specific signal about what the file actually contains, independent of and in addition to whatever alt text ultimately gets applied to it.

This is a small fix with almost no cost, which is exactly why it’s worth naming directly rather than assuming it happens automatically. Export tools and content management systems default to generic, sequential filenames unless someone deliberately renames files before upload, and that deliberate step gets skipped constantly in practice, not because it’s difficult, but because it’s easy to forget when a dozen other things need attention during a content push. A quick, systematic pass renaming files descriptively before they go live costs a few minutes per image and adds a real, durable signal that persists for the life of the file.

Why Surrounding Copy Matters More for Diagrams Than Photos

A purely illustrative photo, one that shows a general scene already fully described in the surrounding text, carries relatively low retrieval stakes. If a system can’t process the image directly and the alt text is thin, little real information is lost, because the text around the photo already covers the same ground.

A diagram is a different case entirely. A diagram expressing a specific relationship, process, or comparison often exists precisely because that information is easier to communicate visually than in prose, which means the surrounding text frequently doesn’t restate it in full. This makes the quality of a diagram’s alt text and surrounding explanatory copy substantially more consequential than it is for a purely illustrative photo, since a thin description here risks losing information that exists nowhere else on the page.

When an Image Is the Only Source for a Specific Fact

The highest-stakes case is a diagram or chart that expresses a relationship, sequence, or comparison not stated anywhere else in the surrounding text. This happens more often than teams typically account for: a workflow diagram showing a process’s actual sequence, a comparison chart summarizing options a reader would otherwise have to piece together from scattered paragraphs, an architecture diagram showing how components connect.

In every one of these cases, the image isn’t illustrating already-explained content. It’s the primary source for that specific piece of information, and if a generative system can’t extract meaning from it, either through direct processing or through sufficiently detailed alt text and surrounding copy, that information is effectively unavailable to any answer the system generates. Identifying which images in a content library fall into this category, versus which are purely illustrative, is the single highest-value distinction to make when deciding where to invest description effort.

Captions as a Distinct Signal From Alt Text

A caption serves a genuinely different purpose from alt text, even though both describe an image in some sense. Alt text exists primarily for accessibility and for systems reading markup without rendering images visually. A caption exists primarily for a sighted human reader scanning a page, providing context, attribution, or a specific detail that adds to what the image visually shows rather than substituting for it.

Both get read by generative systems processing a page’s content, but they shouldn’t be treated as redundant or interchangeable. A caption that simply repeats the alt text wastes an opportunity to add genuinely distinct information, a source, a date, a specific detail about what’s shown that wouldn’t naturally belong in a pure accessibility description. Writing them as two distinct pieces of content, each doing its own job, produces more total retrievable signal than writing one and copying it into the other field.

A Practical Audit Approach for Existing Image Libraries

For a site with an existing library of images accumulated over time, a systematic audit is more productive than trying to fix everything at once. The practical approach starts by identifying which images fall into the high-stakes category covered above, diagrams and charts expressing information not stated elsewhere in the surrounding text, and prioritizing those first, since they represent the clearest cases of currently-inaccessible information.

From there, a broader pass through remaining images can check for the basics: generic or missing alt text, non-descriptive filenames, and captions that either don’t exist or simply duplicate the alt text without adding distinct value. This audit doesn’t need to happen all at once across an entire library; treating it as an ongoing practice applied to new content by default, with periodic passes through older content prioritized by which images carry the most unique information, is more sustainable than a one-time comprehensive fix that’s unlikely to actually get finished.

Making Images Carry Real, Retrievable Meaning

Michael Rubinstein has flagged image description as one of the most consistently under-invested areas across the sites he’s audited, not because it’s technically difficult, but because it sits in exactly the kind of low-visibility, easy-to-defer category that gets skipped when a content team is under time pressure, even though the actual cost of doing it well is genuinely small.

ScribePress generates descriptive alt text, filenames, and captions as a default part of its publishing pipeline, treating this as a required content field rather than an optional accessibility pass applied inconsistently after the fact.

Learn more about the work behind this framework at michael-rubinstein.com.

Put the framework to work