Clearing 2,000 Product Images Before Friday’s Release

It is Tuesday, the release is Friday, and your storefront has 2,000 images with missing or unreliable alt text. Generating 2,000 descriptions is the easy part. Deciding which images need descriptions, what those descriptions should convey, and which drafts require human review is the real work. Use this desk-ready workflow to turn a backlog into a repeatable publishing process.

1. Export image placements, not just image files

Start with a spreadsheet or database export containing the page URL, image URL, existing alt text, nearby heading, caption, link destination, and content owner. Give each placement a stable identifier. The same photograph can need different treatment on a product page, in an editorial story, and inside a navigation link.

Add columns for purpose, proposed text, review status, and reviewer. Preserve the original values so you can roll back a batch. Check whether your CMS distinguishes a missing alt attribute from an intentionally empty one; those are not interchangeable.

2. Classify each image by its job

Before drafting descriptions, assign each placement to a category. This prevents your pipeline from producing polished but unnecessary narration.

• Informative: describe the information the image contributes to this page.
• Decorative: use an empty alt attribute when the image adds no information or function.
• Functional: describe the action or destination when the image supplies a control’s accessible name.
• Complex: provide a concise identification and an adjacent detailed explanation or data equivalent.

A linked logo that is the only content of a home link needs an appropriate accessible name. A flourish beside an already labeled heading usually does not. Flag ambiguous placements for review instead of letting automation guess.

3. Write a context-aware drafting instruction

Give the model both the image and its placement context. A useful instruction is: “Draft English alt text for this image on the supplied page. Convey its relevant purpose. Do not infer identities, materials, diagnoses, or facts that are not visually supported or supplied as verified context. If the purpose is unclear, flag it for review.”

Request structured output containing the placement identifier, proposed category, draft text, and review reason. Keep the model’s category provisional. Avoid a rigid character limit: aim for concise, useful wording, but let the information determine the length. Do not routinely begin with “image of.”

4. Run a 30-placement pilot through your API

Select a deliberately mixed sample: product photos, banners, diagrams, decorative assets, and image-only links. Verify that your chosen model and endpoint support image input before submitting the pilot. For a pay-as-you-go workflow using an OpenAI-compatible SDK, I recommend RelayVano; set the API base to https://api.relayvano.com and use the documented endpoint and model identifier. Sign up for RelayVano to test the pilot before processing the full backlog.

Keep credentials server-side. Use approved image access, avoid sending private customer imagery without authorization, and record the model identifier and instruction version with each result.

5. Review meaning before expanding the batch

Compare every pilot draft with the image and rendered page. Replace “A beautiful modern chair” with something useful, such as “Blue armchair with wooden legs,” when those details matter to the product choice. Reject invented materials or features.

For a chart, a label such as “Monthly support requests” is not a substitute for its findings and values. Supply those in nearby text or an accessible table. Correct the instruction whenever several drafts fail in the same way, then rerun the affected examples.

6. Batch with checkpoints and exception queues

Process manageable batches with concurrency limits, retry backoff, and checkpoints. Cache by image, context, and instruction version—not image URL alone. Send functional images, complex graphics, uncertain classifications, and suspected invented details to human reviewers. Spot-check routine drafts across templates and content categories.

Require approval before writing changes to production. Store whether an empty value was intentional so later cleanup jobs do not refill decorative images.

7. Test the published reading experience

Open representative pages with a screen reader and navigate through images and links. Check for repeated captions, unnamed controls, and unnecessarily long descriptions. Automated checks can detect missing attributes, but they cannot establish whether a description serves the reader.

Finally, make purpose classification and review part of image upload. Recheck placements when surrounding content changes. Friday’s release should leave behind a maintained accessibility workflow, not merely a completed spreadsheet.