AI alt text for a catalogue of 90,000 product photos

Accessibility rules need a text description for every image. For a store with about 90,000 product photos, writing them by hand is not an option.

The owner had already tried plugins and services. They started bringing the site down after a dozen or so tags, with memory use spiking on the server.

So he generated texts outside the site and imported them back by hand. In his words, "a long procedure, a lot of work". And the deadline was less than a year away.

The owner already had a Google Sheet of products and his own prompt. We built on that instead of starting over.

Every 15 minutes the scenario takes 50 products that are not done yet. It sends each photo to the model separately, together with the product's name, description and attributes: period, style, finish, material, colour.

The model writes one sentence per photo: only what is visible, one visible detail, no marketing words, up to 180 characters, following WCAG 1.1.1.

Up to 16 photos per product get their own text. The status and any error are written next to each row, so the next run continues where the last one stopped.

Our first version sent all of a product's photos to the model in one request. It was cheaper, but the texts were worse.

We rebuilt it to describe one photo per call. The cheaper version stays in the account, switched off, in case cost ever matters more than quality.

The scenario lives in the client's own Make account, not on our servers. The model, the prompt and the sheet are settings, so they can change them without rebuilding anything.

We tested on 500 real rows before the full run. Then we wrote a plain-language guide for the team: how to switch it on and off, what to check once a month, what to do if a run fails.

The sheet and the manual import were the last manual step. In phase two the pipeline talks to the site directly.

Every few minutes it finds product photos without alt text in the media library, writes the text and saves it back. The owner uploads a photo, and a few minutes later its alt text is in place.

It takes only product images, through the WooCommerce API, so blog images stay untouched. A photo used in several products is processed once. Access goes through WordPress's own application passwords.

We deliberately skipped the off-the-shelf connector plugin: it is known for running out of memory on large sites, the exact problem that started this project.

Phase two passed testing in June 2026.