by Serguey Shinder
In the summer of 2024 a buyer at one of our trade customers emailed to say that our product pages were close to unusable for her. She is blind and orders using a screen reader, and she had been doing it for years with some effort. What had made it worse, she said, was a recent change.
The change was mine. Our catalogue had about six thousand product images and almost none of them had alternative text, the short description a screen reader speaks in place of a picture. We had been meaning to fix that for years, it was an obvious job for a model, and one afternoon in May we generated all six thousand. I read about forty, they were accurate, and we shipped.
Accurate was the problem. A typical one said: a grey metal bracket with four holes on a white background. That is true. It is also true, nearly word for word, of the forty brackets either side of it. Before the change her reader had skipped the images. After it, every product took an extra sentence to announce something that told her nothing, and a page of fifty brackets had become fifty sentences longer.
The model had done exactly what it was asked, and done it well. What it had never been given was the thing alternative text is for, which is not what the image looks like but what the image tells somebody that the rest of the page does not. For most of our product photos the honest answer was nothing at all, because the name, the dimensions and the finish were already written beside them. For a few, the photograph was the only place a detail appeared, the position of a fixing hole or the shape of a profile, and those were precisely the ones where a generic description was worst.
We redid it with her help. Images whose information was already in the text were marked as decorative, so a screen reader passes over them, which covered roughly four in five. For the rest, the model was given the product name, the listed attributes and one instruction: describe only what a buyer could learn from this photograph that is not stated in the text below it. And each month somebody who uses a screen reader goes through fifty pages, because I had checked forty descriptions by reading them, which is not how anybody who needed them would ever meet them.
What I took from it goes beyond images. A model can tell you what something is, fluently and correctly, and that is a different job from telling a particular person what they need to know about it. If the purpose is not in the prompt, you get an answer to a question nobody was asking, and it will be accurate enough to pass every check you thought of.
– Serguey Asael Shinder
Leave a Reply