Blog / AI & search /

AI image alt text: useful shortcut or risk?

Most online stores have the same folder of shame: hundreds, sometimes thousands, of product images with no alt text. Or worse, alt text that just says IMG_4021 or repeats the product name on every photo.

Nobody wants to write alt text for 2,000 images by hand. So when AI tools started being able to look at images and describe them, it felt like the answer had arrived. ChatGPT can now take an image as input, image description has been built into accessibility tools for years, and plugins and apps promise to fill in every empty alt attribute on your site with one click.

I think AI alt text is genuinely useful. I also think it’s easy to use in a way that makes things worse. Here’s where I’ve landed.

What alt text is actually for

It’s worth starting here, because most bad alt text comes from forgetting it.

Alt text is a short description of an image for people who can’t see it: people using screen readers, people whose images didn’t load. Search engines use it too, to understand what an image shows and what the page is about.

But the first job is accessibility. Good alt text tells someone what the image means in that context, not just what’s in it. I covered the basics in image SEO on WordPress: file names, alt text and size and image SEO for online stores.

Where AI alt text helps

Getting from nothing to something

An image with no alt text tells a screen reader user nothing at all. A reasonable AI description is a real improvement over that. If the choice is between AI alt text and none, for a large back catalogue, AI wins.

Speed on large catalogues

A store with a few thousand product images can get a first draft for all of them in an afternoon instead of weeks. That changes whether the job happens at all.

Describing what’s visually there

Modern models are good at the literal: colour, shape, material, setting, people, objects. “A navy linen shirt laid flat on a white background, with rolled sleeves” is the kind of thing they do well.

Where it quietly fails

It doesn’t know the context

AI describes the image. It doesn’t know why the image is on the page. A photo of a woman laughing at a laptop could be:

  • a stock photo that needs no description at all (decorative),
  • a product shot where the laptop is the point,
  • an illustration for a post about remote work.

Each needs different alt text. The AI writes the same description for all three.

It doesn’t know your product

It can see a blue shirt. It can’t see that it’s your “Harbour” shirt in slate, in 100% linen, with mother-of-pearl buttons. The details that matter to a shopper, and the words that might match a search, are often exactly the ones it can’t know.

It can be confidently wrong

Models sometimes guess: the wrong material, the wrong colour name, a detail that isn’t there. On a product page, “leather” when the product is faux leather is not a small error. It’s a misleading description of something you’re selling.

It over-describes

Left alone, AI tends to write long, flowery descriptions of every image, including logos, dividers and background patterns. For a screen reader user, that’s noise. Decorative images should have empty alt text (alt="") so they’re skipped.

The keyword temptation

Some tools offer to “optimise” alt text with keywords. That’s how you end up with “best cheap linen shirts men linen shirt summer” on every photo. It’s bad for accessibility and doesn’t help SEO either.

How I’d use it

AI as the first draft, a human as the editor. In practice:

  1. Sort images first. Product images, content images, decorative images. Decorative ones get alt="" and no AI at all.
  2. Give the AI context. Don’t just pass the image. Include the product name, colour and material from your product data, and ask for a short, factual description of under about 125 characters. The result is far better than an image-only description.
  3. Review the important ones by hand. Your top products and top pages, at least. Check for wrong details.
  4. Spot-check the rest. Read a random sample. If the error rate is high, fix the prompt and run it again.
  5. Keep it short and plain. One sentence, describing what matters, no “image of” at the start.

A simple prompt that works better than most:

Write alt text for this product photo. Product: [name]. Colour: [colour].
Material: [material]. Describe what the photo shows in one plain sentence
under 125 characters. Don't start with "image of". Don't add details you
can't see or that aren't in the product information.

Is it an SEO risk?

Not in the sense of a penalty. Google cares about whether alt text is helpful, not who or what wrote it. That’s consistent with what Google has said about AI-generated content in general: quality is what’s judged, not the method.

The real risks are quieter:

  • Inaccurate descriptions of products you sell.
  • Worse accessibility if decorative images get long descriptions and important ones get vague ones.
  • Generic, identical alt text across similar images, which tells nobody anything.

Those are reasons to review the output, not reasons to avoid the tools.

My view

Use it. But treat “AI-generated alt text” like “AI-generated meta descriptions”: a fast first draft that still needs someone who knows the product and the page. The time saved writing is real. Spend some of it checking.

Here’s a small test: pick ten product images, generate alt text with and without the product details in the prompt, and compare. I’d be curious which version you’d rather a screen reader read aloud.