An image library that maintains itself: alt texts and photo review by AI

Ask a website manager where the time goes and the answer is rarely the copy. It is the photos. Collecting, sorting, resizing, naming, placing them in the right spot, and then thinking about the description once more for every language. With hundreds of photos that is no longer a chore, it is a day job. That is exactly the work I handed over to AI in a recent project. In this article I show what becomes possible.

The silent problem: missing alt texts

An alt text is the description of a photo in the page's code. Visitors using a screen reader hear that description read out. Google reads it to understand what is in the photo, and uses it for Google Images and increasingly for AI search results.

In practice the alt text is missing almost everywhere, or it says something like "IMG_4832" or "photo1". Understandable, because nobody enjoys describing hundreds of photos one by one. Let alone in three or four languages.

AI describes what is actually in the photo

That work can be done differently now. The AI looks at every photo in the library and writes a description of what is really in it: not "image", but for example "wooden jetty on a lake at sunrise". Per language, because a German visitor deserves a German description.

The photo itself is stored once in the library and belongs to all language versions of the page. Only the alt text differs per language. That keeps management simple: replacing a photo is one action, not four.

And because the texts come from the content of the photo, they are correct. The visitor with a screen reader notices, and so does Google.

Photo review: which photo is good enough?

A second task AI handles well: judging whether a submitted photo is usable. Customers and partners deliver photos in all shapes and sizes. Sharp and blurry, light and dark, portrait and landscape, and sometimes ten near-identical shots.

The AI reviews every submitted photo on sharpness, exposure and composition, and gives a verdict on a four-point scale with a short explanation. Below a configurable quality threshold a photo is rejected. Above it, advice is added: this photo is strong enough for the large header image at the top of the page, that other one fits better as a card in an overview.

The work shifts from "look at everything yourself" to "judge a proposal". The human still decides, but no longer has to search.

The technology underneath: web formats without thinking about it

Under the hood the library automatically creates the right web formats from every original: modern file types such as AVIF and WebP, in multiple widths for phone, tablet and desktop. The original is kept, the website always serves the smallest variant that is still sharp.

That is not a detail. On most websites photos are the biggest source of load time, and load time directly costs revenue. A library that handles this itself keeps the site fast without anyone ever having to think about it.

What this delivers

For a site with hundreds of photos this means, concretely: alt texts that are correct everywhere in every language, a consistent quality bar for the imagery, and a website that stays fast. The manual work that remains is the enjoyable part: choosing which photos tell the story.

And as with everything I build with AI: the AI proposes, the human decides. Every text and every verdict can be adjusted. In practice that rarely happens, and that is exactly the point.

Read also

• From WordPress to a flat-file CMS: switching without losing your Google rankings, the migration this image library was part of
• What AI really means for your SMB (without the hype), more practical AI applications
• View all completed projects, web platforms and automation solutions

Do you also have a folder with a thousand photos waiting? Book a no-obligation call. Chances are this work can be automated for you too.

Bert