Write alt text for the actual image and its role in the page, not for the prompt that requested it. A prompt can mention details the model did not produce, hidden technical instructions or unrelated stylistic goals. Alternative text should help someone understand the delivered visual in context. It is not a place to stuff keywords.
Start from the image's job
Decide whether the image conveys information, illustrates a story, performs a link's function or is purely decorative. Context affects the appropriate description. Use established accessibility guidance for the actual publication format rather than giving every image the same long catalogue of visible objects.
For an informative image, identify the essential subject and relationship. If the picture is a concept or reenactment and that distinction matters, include it in the surrounding caption or description. Do not invent factual provenance merely because the generation brief requested realism.
A fictional café transformation
An invented article compares a real café photograph with a doodle-style interpretation. The original prompt requests several tiny characters, warm light and a retro lens. The output contains only two drawn figures, and no camera provenance is known.
A suitable description focuses on the actual scene and transformation: a café table reinterpreted with two small drawn figures around a cup. The surrounding caption can identify it as AI-assisted artwork. Copying the prompt would describe absent characters and a lens that has not been established.
Keep exact information outside the artwork
If an image carries essential text, supply that information in accessible page text as appropriate. Do not make someone depend on decorative raster lettering to learn an event's date. For complex charts or diagrams, a short description may need a separate detailed explanation containing verified data or relationships.
Check descriptions after every meaningful visual revision. A caption for an old image can become incorrect even when the filename is unchanged. Keep the description attached to the delivered version rather than the initial creative idea.
Include descriptions in the handoff
Imagild's library helps manage generated assets, but publication accessibility depends on the destination and its supported fields. Deliver a small text record with the approved file, proposed alt text, caption and any important disclosure. The recipient can adapt it to the actual page context.
Use the final image for review. A reverse prompt may help inventory visible features, but its analysis still needs checking and is not automatically suitable alt text. Keep private instructions, names or sensitive production details out of public descriptions.
Should decorative images have long descriptions?
Not necessarily. Follow the destination's accessibility implementation and the image's purpose. Describing every decorative detail can add noise rather than useful information.
Does mentioning AI improve search ranking?
Do not assume a ranking benefit. Use accurate disclosure where relevant and write useful descriptions for people; discoverability is not a guarantee.
Sources and related reading
The W3C WAI Images Tutorial explains how image purpose affects text alternatives. Apply that guidance to the actual publication context, not to private generation instructions.