AI can write your alt text. It doesn't know which detail the shopper needs.

What the field is for

Someone who's blind or has low vision often shops with a screen reader, which reads the page aloud. When it reaches a product photo, it reads the alt text instead. If that's empty, they hear nothing, or a filename. If it's vague, they hear something like "green bag," and have to guess at everything the photo would have told them.

The accessibility standard most sites work toward, WCAG, puts the requirement simply: images need "a text alternative that serves the equivalent purpose." That word purpose is the whole game. Alt text isn't a description of the photo. It's the reason the photo is there, in words.

It isn't only screen reader users who meet it, either. Alt text shows when an image fails to load on a slow connection, and search engines read it too. But write it for the person listening, and the rest tends to follow.

AI tools are now very good at the description part. Point one at a product photo and it'll tell you what's in it, accurately, in a tidy sentence. What it can't know is why you included that photo, and which detail a shopper needs from it. So let's take one product, a canvas weekender bag in a color the brand calls olive (a made-up example), and look at three versions of alt text for each kind of image: good, better and best.

The main photo

This is the front view on a plain background, the photo at the top of the page.

Good, what a vision model typically writes on its own: "A green canvas bag with brown leather handles on a white background." Accurate. But "on a white background" tells the shopper nothing, and "green" isn't the color they'll be choosing from.

Better, with the product name and color from your catalog: "Olive canvas weekender with tan leather handles." Now it matches the color the shopper will pick.

Best, with the detail that distinguishes this bag from the others on the page: "Olive canvas weekender, front view: tan leather handles, brass top zip, detachable shoulder strap." It's still short. It skips the background. And it front-loads what someone comparing three similar bags would want to know.

One more nuance. If the product name already sits right next to the image as a heading, you don't need to repeat the whole thing. Screen reader users hear the heading too. Lead with what the heading doesn't say.

The detail shot

Detail shots are where AI drafts are weakest, because the reason for the photo is usually a feature, and the model has no idea which feature you're showing off.

Good: "Close-up of a zipper." True, and useless.

Better: "Close-up of the brass zip pull."

Best: "Close-up of the two-way brass zip, which opens along the full top of the bag." This is the version that needs your product data, or a person who knows why the photo was taken. A model looking at pixels can't tell that the zip is two-way or that it opens fully.

The same goes for fabric texture, stitching, the inside pockets, the tag. If you photographed it, there was a reason. The alt text should say that reason.

The photo that shows size, and the ones to leave empty

Lifestyle photos usually exist to show scale or use: how big the bag is on a person, how it sits on the shoulder.

Good: "A woman carrying a bag in a train station."

Better: "A model carrying the olive weekender over her shoulder."

Best: "The weekender on a model's shoulder, reaching from shoulder to hip." That answers the question the photo was there to answer: how big is it?

Packaging and "what's in the box" photos are the other type worth slowing down for. Good: "A bag and a small pouch." Best: "What's included: the weekender, a detachable shoulder strap and a cotton dust bag." A sighted shopper counts the items in a second. A shopper listening to the page only knows what you tell them, and this is exactly the kind of detail that ends up as a return when it's missed.

Then there are images that carry no information at all: a background texture, a decorative divider, a repeated brand pattern. These should have empty alt text so screen readers skip them. Many tools will dutifully write "abstract beige pattern" for each one, which is just noise. WCAG specifically lists placeholder text and filenames as failures, and the same thinking applies to describing decoration.

Getting the AI to "best" more often

The trick is to stop asking the model to describe a photo and start asking it to write alt text for a product:

• Give it the product record. Name, color name, materials, key features. Most of the gap between good and best is information the model can't see.
• Tell it the image type. Main, detail, lifestyle, packaging, decorative. Each has a different job, and you can give the model a short rule for each.
• Tell it what to skip. Backgrounds, "image of," repeating the heading word for word.
• Never let it invent. If the record doesn't mention a feature, the alt text shouldn't either. A model will happily guess "waterproof." The same caution applies to AI-generated product imagery and to catalog claims in general.
• Route the risky types to a person. Detail shots, sizing images, anything showing a claim. Main shots and simple angles can often go straight through with a spot check.

Check it by ear

The best test costs nothing. Turn on the screen reader built into your phone or computer, close your eyes, and try to choose between two of your own products. You'll hear right away whether the alt text helps, and where it's just filling a field.

Do it for a handful of pages after the tool runs, and again whenever you change the prompt or the tool. It's the same habit as checking which words matter in AI translation: the model handles the bulk, and people check the details that change a buying decision.

We build AI applications that work from your own product data, and ecommerce operations around them. If you're about to let a tool fill in thousands of alt text fields, we can help you set it up so the answers are actually useful.

Related work: Howcast.