AI-written image alt text: review every one or spot-check a sample?

Say you run a small ecommerce site and have an AI write alt text for a few hundred product and blog images. Generic stuff like "blue running shoe" is fine for most of them, but the model sometimes describes details that aren't in the photo, or repeats the same keyword in every line. My instinct is to fully review alt text on pages that get organic traffic and only spot-check the rest. Decorative images don't need descriptions anyway. Others argue a bad alt can sit there for years, so a full pass before publishing is cheap insurance. Where do you draw the line? And do you give the model a rule like "describe only what's visible, no keywords," or let it run and clean up after?

Agree on the split, but I'd sort by template rather than by traffic. If a product gallery uses the same shot layout, the model tends to fail the same way across all of them, so one bad alt usually means a hundred. Spot-checking one per template catches that faster.

On the prompt side, "describe only what's visible" isn't enough on its own. Feed the model your catalog fields (color, material, size) and tell it to use only those plus what it can see. Most invented details, like "leather strap" on a canvas bag, are the model guessing, and a fixed attribute list cuts that down a lot.

The catalog-fields idea works, but watch what happens when a field is blank. The model doesn't leave a gap, it fills it. A canvas bag with no material set is exactly where "leather strap" shows up. So the prompt should say "if a field is empty, leave that detail out," not just "use only these fields." That one line matters more than the list itself.

For deciding which pages to review first, I'd use Search Console. In the Performance report, filter Search type to Image and you'll see which pages actually pull image-search clicks and impressions. That's a better priority list than overall organic traffic, because those are the pages where alt text is doing real work in Google Images. The report doesn't break things down by individual image, so you still have to open the pages, but at least you know where the review time pays off.

The Image filter counts pages, not pictures, so a product page with eight photos shows up as one row and you still can't tell which image earned the clicks. And if the shop sells through Shopping or Performance Max, the page's alt text doesn't feed those campaigns anyway. They pull from the Merchant Center feed title and image. That means a page can look important in Image search and still matter less for paid traffic, so check both before deciding where the review hours go.

Splitting the review by channel instead of by traffic alone makes sense. I'd push back on one thing, though. Keyword stuffing usually hits feed titles hardest, because the model is writing for a search box and tries to fit in every term. A title like "blue running shoe men lightweight mesh sport trainer" can read fine to the model but performs badly in Shopping. So if paid traffic matters, review feed titles first and treat page alt text as the lower-stakes pass, mostly a sample check for invented details.

Is the shop running Shopping or Performance Max at all, or is it mostly organic right now?