Roughly half of what an audit finds was decided before any code existed: a palette that fails contrast, a focus state nobody drew, an error pattern that uses colour alone, a tap target too small for a thumb. Those are design decisions, and retrofitting them costs several times what deciding them once would have. This page is the design half of the work, and it is written for the designer and the business paying for the redesign, who is usually the same person who received the letter.
What accessibility in web design actually means
Not a separate accessible version and not a widget. It means the palette meets the contrast ratio, every interactive element has a drawn focus state, information is never carried by colour alone, text reflows at 200% zoom and at 320 pixels wide, and touch targets are large enough to hit. Six decisions, made once in a component library, that remove most of a future audit's template findings.
Where a redesign helps and where it hurts
A redesign is the cheapest moment to fix accessibility and the most common moment to break it. New templates mean new template issues: on the worked example nine templates carry 126 issues, and a redesign replaces all nine at once. Put the checklist into the design review rather than into a post-launch audit and the same work costs a fraction.
What to hand engineering
The palette with the ratios recorded, the focus states, the error pattern, the heading scale mapped to real heading levels, and annotated designs for anything interactive: what it is called, what role it plays, what changes when it is used. Those annotations are what stop a developer inventing ARIA to describe a picture.
Questions people ask about accessibility web design
Does accessible design limit the visual language?
Barely. Contrast and focus are the two constraints, and both have solutions that keep a brand intact.
Who signs it off?
Whoever signs the design off already, with the checklist in the review. Adding a separate approval step is how it gets skipped.
What about a design system?
It is the highest-leverage place to fix this: one conformant library removes the same finding from every future project.