Alt text says what an image contributes, not what it shows

Alt text exists to say what an image contributes to the page. For most generated hero images, the honest answer is nothing: they are there to make the page pleasant, and the argument is in the words. The correct alt text for an image like that is empty. Until this morning, every hero on this blog described itself anyway, and the reason is worth more than the fix.

A screen reader reads every word you put there

The W3C's Web Accessibility Initiative is plain about this. Its tutorial on decorative images defines them as images that add no information to the page, and tells authors to give them a null alt attribute, an empty pair of quotes in the markup, so that assistive technology skips them. A description, it says, would add audible clutter to screen-reader output, or distract when its subject differs from the text around it.

The same page closes the obvious loophole. Leaving the attribute out entirely is worse than leaving it empty, because some screen readers then announce the file name instead. Empty is a deliberate statement: there is nothing here you need.

Among the cases the tutorial lists as decorative is the image that illustrates the adjacent text without contributing information, which it calls eye-candy. That is a fair description of a metaphorical hero. It is supposed to set a mood. It is not supposed to be read.

Our own heroes got this wrong

Every hero on this blog has carried alt text built from the scene we asked the image model to paint, prefixed with "Editorial illustration" and cut at 120 characters. On our post about what Search Console shows, a screen-reader user heard the title, then a sentence about grey stepping stones across a lawn and a shortcut worn through the grass, which stopped halfway through a word. Finnish readers heard the same English sentence, because the field is not translated. Then the argument began.

To a sighted reader that picture takes half a second and quietly rhymes with the post. Read aloud, it is a riddle with no answer and a broken last word. We built the description because a missing alt text looks like an accessibility failure in every audit tool, and a description looks diligent. Neither appearance was the point.

Today's hero is the first to ship with an empty alt attribute. The archive behind it still carries the old descriptions, and correcting each one is a single field edit. The mistake was never expensive to fix. It was expensive to notice.

Google's advice is about a different image

The strongest objection comes from search. Google's image SEO guidance calls alt text the most important attribute for giving an image metadata, says Google uses it alongside computer vision to understand what an image shows, and offers a descriptive example as best practice. Leave it empty and the image contributes nothing to image search.

That advice is written for images you want found. A product photograph, a diagram of how something works, a picture of a place a customer will visit: those deserve to rank in Google Images, and describing them serves the reader and the index at once. Nobody searching for pictures of a garden path is looking for an article about Search Console, so ranking our hero there would help no one. The same guidance warns that alt attributes stuffed with keywords can get a site treated as spam, and points readers to the W3C for accessibility. It is the rule we argued for the prose in Google doesn't penalize AI content: write for the person, and the crawler follows.

The machine already knows why the image exists

The WAI tutorial adds one more line that matters here. Whether an image is decorative or informative is a judgment only the author can make, based on why the image was included. On a hand-built site, that reason often lives in someone's head and leaves with them. A stock photo chosen in a hurry has no recorded purpose at all.

A generated hero is the rare case where the reason is written down at the moment of creation. The engine asked for a metaphor, in a deliberately printed style that, as we argued in illustrations are not evidence, refuses to look like proof. The role is known before a single pixel exists, so the alt decision costs nothing and can be made by rule rather than by mood.

We should concede where this goes the other way. Machines are good at describing pixels and poor at saying what an image means for an argument. A vision model will tell you a chart has five blue bars. The useful alt text says what the chart shows and why the paragraph above it needs it, which requires knowing the data and the claim. For informative images, a generated description is a draft for a human, never the finished text.

Decide the role before writing the words

So give every image a role before it gets any alt text. Decoration gets an empty attribute and costs no one anything. Evidence comes from the client, carries a description of what it contributes, and is checked like any other claim on the page. If a blog cannot say which of its images are which, the alt text was never the problem. It had not yet decided what its pictures are for.