How to use AI to write alt text that actually helps
A practical workflow for using AI vision models to write alt text: sort your images, give the model context, and review before you publish.

Most websites we audit have the same quiet problem: images with no alt text, or alt text that says "image1.png", "hero" or "photo of a laptop". Screen reader users get nothing useful, search engines get nothing useful, and the designer who spent hours choosing those images gets no credit for them.
AI vision models are now good enough to describe an image in a sentence or two. That makes alt text fast to write, but it doesn't make it good by default. A model sees pixels; it doesn't know why the image is on the page. This guide shows you how to use AI for the boring part and keep the judgement calls for yourself.
What alt text is actually for
Alt text replaces an image for people who can't see it. That includes screen reader users, people on slow connections where images fail to load, and crawlers that index your page. The WCAG success criterion 1.1.1 (Non-text Content, Level A) asks for a text alternative that serves the same purpose as the image.
The key word is purpose. The same photo of a person typing can mean "our team works remotely" on an About page and "a customer using the dashboard" on a product page. The alt text should follow the purpose, not just the pixels. This is exactly where AI needs your help.
Sort your images before you prompt anything
Before you open any AI tool, go through the page and put every image into one of four groups. This takes ten minutes and saves you from generating text that should never exist.
- Decorative: background textures, dividers, abstract shapes, stock photos that add mood but no information. These get an empty alt (alt="") so screen readers skip them.
- Informative: photos and illustrations that show something the text doesn't say. These need a short description.
- Functional: images that act as links or buttons, such as a logo linking to the homepage. The alt text describes the action or destination, not the picture.
- Complex: charts, diagrams and product screenshots with data. These need a short alt plus a longer explanation in the surrounding text.
Only the informative and complex groups really benefit from AI. Decorative images need no text, and functional images need a word or two you can write faster than you can prompt.
Give the model the context it can't see
If you upload an image and ask "write alt text for this", you'll get a generic description: "A person sitting at a wooden desk using a laptop in a bright room." It's accurate and almost useless. The fix is to tell the model where the image lives and what job it does.
A prompt structure that works well for us:
- The page and section: "This image is in the hero of a pricing page for a Framer template shop."
- The surrounding copy: paste the heading and the paragraph next to the image.
- The purpose: "It shows what the dashboard looks like after setup."
- The constraints: "Write alt text under 125 characters. Don't start with 'image of' or 'picture of'. Don't repeat the heading."
With that context, the same model will usually give you something like "Dashboard showing three active projects, weekly revenue chart and an empty task inbox." That tells a screen reader user what a sighted user learns from the image.
Rules to include in every prompt
We keep a short block of rules and paste it into every request. You can save it as a custom instruction or a text snippet.
- Describe what matters for this page, not everything visible.
- Skip "image of", "photo of" and "graphic of"; screen readers already announce it as an image.
- Include any text that appears in the image if it carries meaning, such as a sign or a button label in a screenshot.
- Don't guess people's identity, age, ethnicity or emotions unless the page depends on it.
- Don't add keywords for SEO that aren't true of the image.
- End with a full stop so screen readers pause naturally.
The rule about guessing is important. Vision models will confidently describe a stranger as "a happy young woman" when the page has no reason to say that. If you don't know, and it doesn't matter, leave it out.
Handle screenshots and charts differently
Product screenshots and charts are where AI helps the most and where it also makes the most mistakes. Models can read the labels in a chart, but they sometimes misread numbers or invent trends that aren't there.
For these images, ask for two outputs: a short alt (one sentence on what the image shows and its main takeaway) and a longer description you can place in the body text or a caption. Then check every number against your source data. If the chart says revenue grew 18%, and the model says 81%, the alt text is now spreading a wrong fact to exactly the people who can't check it.
Treat AI-written alt text for data like a junior's first draft: the structure is useful, the numbers need checking.
Batch the work without losing quality
On a site with a blog, a portfolio and a CMS full of case studies, you may have hundreds of images. Doing them one by one in a chat window is slow. A better workflow:
- Export a list of images with their page URL, the nearby heading and any existing alt text. A simple spreadsheet is enough.
- Mark each row as decorative, informative, functional or complex.
- Send the informative and complex rows to the model in groups of ten to twenty, with the context columns included.
- Review the output in the spreadsheet. Edit, don't just approve.
- Paste the final text into your CMS or design tool.
In Framer, image settings include an alt text field, and CMS collections can hold alt text in a dedicated field that you bind to the image. Setting up that field once means every new post or case study asks for alt text at the moment you add the image, which is when you know its purpose best.
Review checklist before you publish
Spend thirty seconds per image on this check. It catches most of the problems AI introduces.
- Does the alt text match the purpose of the image on this page?
- Is it shorter than about 125 characters, unless it's a complex image with a longer description elsewhere?
- Does it avoid repeating the caption or heading word for word?
- Are all names, numbers and labels correct?
- Is any decorative image still carrying text it doesn't need?
Then turn on a screen reader (VoiceOver on Mac, Narrator on Windows) and listen to one page from top to bottom. Hearing your alt text read aloud between headings shows very quickly which descriptions are too long, too vague or redundant.
Where AI alt text goes wrong
A few patterns we see again and again:
- Over-description: three sentences about lighting and furniture for a mood photo that should have been decorative.
- Keyword stuffing: "best Framer template for agencies modern website design" as alt text for a screenshot. It doesn't help rankings and it annoys screen reader users.
- Confident errors: the model names a product, a place or a brand that isn't actually in the image.
- Same text everywhere: one alt reused across a gallery, so every image sounds identical.
All of these come from skipping context or skipping review. Neither step takes long.
Conclusion
AI can take alt text from the task everyone skips to a task that takes a few minutes per page. The model handles the description; you handle the purpose, the facts and the decision about what doesn't need text at all. Sort your images first, give the model the page context, keep a fixed set of rules, and always review before you publish. Your site becomes more usable for real people, and your images finally get described the way you intended.


