Context-Aware Alt Text Generation and Accessibility Auditing

The problem with alt text on real websites

Screen reader users depend on alt text to know what an image is doing on the page. On many sites it is missing, or it is the file name, or it is a generic description that is technically accurate and useless: "a person standing near a building". Content teams at news sites, universities and online shops know they should fix it, but auditing thousands of images by hand never gets prioritised.

Good alt text depends on purpose. The same photograph of a vice-chancellor needs a different description in a news story, on a staff profile and as a link to the leadership page. That context is exactly what most tools ignore.

What already exists and the gap

• axe-core and WAVE detect missing alt attributes and many other failures, but do not write alt text or judge whether existing text is useful.
• Siteimprove and similar platforms audit accessibility across large sites for organisations with budgets.
• Overlay products such as accessiBe promise automatic fixes, including AI-generated alt text, and are widely criticised by accessibility practitioners for inaccurate results.
• CMS features and vision models can caption images, usually from the image alone.

The gap is context-aware generation with a mandatory human review, plus honest measurement of whether the output helps screen reader users. Detecting missing alt is solved. Writing alt text that serves the image's purpose, and never publishing it unreviewed, is not.

What you build

Two roles: the editor who reviews suggestions, and the site owner who reads the audit report.

1. Crawl a site within a page limit, capturing images and their DOM context.
2. Run axe-core for your chosen WCAG subset.
3. Classify each image's purpose: decorative, informative, functional (inside a link or button) or complex (charts, diagrams).
4. Generate alt text per purpose, from the image plus its context.
5. Review, edit and export.

The review step is not optional. Publishing generated descriptions unreviewed is how you make a site worse in a way nobody notices.

The technically interesting part

The hard problem is turning page context into the right kind of description. A functional image inside a link should describe the destination, not the picture. A decorative divider should get empty alt. A chart needs a summary of its point, and a longer description elsewhere. Build the purpose classifier first, then prompt per purpose using the heading, caption and paragraph, and show your supervisor side-by-side examples where context changed the output for the better.

How to evaluate it

• Blind comparison: for 100 images, raters see context-aware, context-free and human-written alt text in random order and rate usefulness in context. Automated similarity metrics are close to meaningless here; say so in your report.
• Screen reader users: if at all possible, involve blind or low vision reviewers. Their ratings outweigh everyone else's.
• Purpose classifier accuracy against hand labels, especially decorative versus informative.
• Audit accuracy: compare your WCAG findings against a manual audit of 10 pages.
• Editor time: minutes per 50 images with and without suggestions.

Suggested plan

1. Weeks 1 to 2: choose the WCAG subset and three test sites (with permission or public sector sites); hand-label 100 images.
2. Weeks 3 to 5: crawler, context capture, axe-core checks.
3. Weeks 6 to 8: purpose classifier and context-aware generation.
4. Weeks 9 to 10: review queue and export.
5. Weeks 11 to 13: blind rating study and audit comparison.
6. Week 14: report and demo.

Path to a product

The first customer is a university web office or a digital agency that maintains many client sites and keeps being asked for accessibility compliance. Charge per site per month for audits and the review queue, or sell it as an agency tool. Accessibility regulation is tightening in several markets, which creates demand, but the pitch must stay honest: this helps editors write better alt text faster; it does not make a site compliant on its own.

Scope warning

• Pick a subset of WCAG criteria and cover them properly. The whole specification produces a shallow checker.
• Do not build a browser overlay or inject fixes into live pages.
• Do not claim compliance. Claim findings and suggestions.
• Do not crawl sites without permission or ignore robots rules.

Learn what you need on Sefism

• Frontend roadmap : semantic HTML and the DOM your crawler reads.
• AI engineering cheat sheet : prompting vision models and structuring their output.
• axe-core : the rules engine behind your automated checks.
• How to Meet WCAG 2.2 : the success criteria your findings map to.
• Learn Accessibility : what good alt text looks like, with examples.

Get help with this idea

The strength of this project is the context-aware generation and a rating study with real users, so plan those early. Tauseef can help you pick the WCAG subset, design the blind evaluation and keep the scope away from overlay-style shortcuts. If you are new here, create a free account first.

• Discuss this idea with a mentor
• Send a query to Tauseef
• Book a 1:1 session