An AI assistant can turn an image into a sentence, but an editor still has to decide what that sentence should do for the reader. A backpack photo may identify the blue colour option in a shop, demonstrate the position of a pocket in a buying guide, or simply decorate a story whose text already explains everything. Before asking AI for alt text, decide why the image is there. Then check the proposed wording against the final version of the page.
Identify the job the image does on this page
Start with the heading, nearby paragraph and any caption. Ask what a reader would miss if the image disappeared. On a product page that compares backpack colours, the exact colour and any visible contrast panels may matter. In a packing guide, the relevant detail may be that the side pocket holds a water bottle. A detailed sentence about the stitching would add little if the guide is explaining where to put liquids.
The W3C Web Accessibility Initiative decision tree helps distinguish informative images, decorative images, images with meaningful text and images that function as controls. A decorative image can have an empty alt attribute. If an image is the only content inside a link, give the link an accessible name that describes its destination or action. The same picture can serve a different purpose on another page, so review it in each placement.
Give the assistant the image and the relevant context
Use the final crop, or describe the visible details yourself if the approved tool cannot accept images. Include only page context that the tool is permitted to process. A prompt for the packing guide could be: Draft one brief text alternative for this photo. The paragraph explains why a bottle should be easy to reach during a walk. Convey the visible detail that helps the reader understand the placement. If you cannot identify it confidently, tell me what to check. Avoid guessing the bottle’s capacity or the brand of the bag.
If the assistant says, “A 750 ml insulated bottle slides into a mesh side pocket,” inspect the materials and any visible label before keeping those details. If capacity and insulation are unknown, a better version is “A water bottle sits upright in the backpack’s exterior side pocket.” Each adjective should either communicate useful information or help identify the relevant feature. Leave out details you cannot establish from the image or a trusted source.
Edit for meaning and check the published page
Read the nearby text in order with the proposed alt text in place. If the sentence merely repeats the caption, look for information that is still missing. If the image is redundant in that context, follow the decorative guidance. If the image gives instructions or shows a key relationship, keep the detail the reader needs to follow the instructions. For a chart with multiple series, use a short identifying alternative and provide the important values and trends in nearby text or a table. The W3C images tutorial gives examples for these different roles.
Preview the page where the image actually renders. Check whether the alt attribute survived the publishing process and whether the image got cropped differently on mobile. If you can, read the page with a screen reader or inspect the accessible name of any image-only link. A description that correctly names an object still leaves a gap if the link’s action is unclear. Our guide to testing AI-generated design concepts with real readers explains how to give people a real task and observe where they get stuck.
Handle batches one image at a time
For a gallery or catalogue, keep a simple table with one row per image placement. Record the page, the image’s purpose, the wording you approved, and anything still to verify. Start with one example of each main role. For instance, test a product photo, a decorative banner and an image-only link before reviewing a long list. That small sample can reveal a template problem, such as the editor’s text being ignored or the same text appearing in every image of a gallery.
When a page changes, revisit the images whose purpose changed with it. A cropped photo may no longer show the side pocket mentioned in the earlier description. A thumbnail that used to appear beside a text link may now be the entire link. Update the description or accessible name where it is used. The final check is whether a reader can get the same useful information or take the same action from what the page provides in text.
