Why AI-built sites fail accessibility, and what that costs in Ontario

Accessibility is the part of a website that a generated build gets wrong most reliably, and the part an owner is least likely to notice — because everything works perfectly well for the person checking. It is also, in Ontario, the part with a law attached. Both of those are worth understanding before you accept a site that was produced in an afternoon.

Do AI-built websites meet accessibility standards?

Usually not, and the gap is measurable. WebAIM’s 2026 survey of the top million home pages found 95.9% had detectable WCAG 2 failures, up from 94.8% in 2025 — the first reversal after six years of improvement — with an average of 56.1 errors per page, a 10.1% rise.

What accessibility actually means for a small business

Strip away the standards language and it is one question: can somebody who does not browse the way you do still become your customer? Somebody using a screen reader, somebody who cannot use a mouse, somebody reading your prices in bright sunlight on a phone, somebody whose hands shake.

In Canada that is not a small group, and it is not a group that complains — it is a group that leaves. An unlabelled form field does not produce an angry email. It produces an enquiry you never receive and never know about.

The six failures that account for almost everything

WebAIM found that six issues alone accounted for 96% of every error detected across the million home pages they tested, and that the same six have topped the list for seven years running. That consistency is the useful part: it means the fix list is short and known. Their 2026 figures:

Three of those six sit directly on the path between a visitor and an enquiry. None of them is expensive to fix. All of them are invisible unless you go looking, which is precisely why a build that nobody reviewed ships with them.

The ARIA trap, and why it matters here

WebAIM reported something in 2026 that sounds backwards at first reading. Pages using more ARIA, the very attributes meant to make complex interfaces usable with assistive technology, averaged 59.1 detected errors, against 42 on pages that used no ARIA at all. More accessibility markup, more failures.

That is the signature of accessibility applied as decoration rather than as structure, and it is exactly what you get when a model is asked to make something accessible without understanding what the component does. Correct-looking attributes on the wrong elements are worse than none, because assistive technology believes them.

Where Ontario law actually bites

Ontario sets a specific bar. Under the Integrated Accessibility Standards Regulation made under the AODA, designated public sector organisations and businesses or non-profits with 50 or more employees must make public websites and content published after January 2012 meet WCAG 2.0 Level AA. That deadline was 1 January 2021 and has passed.

Be clear about the threshold, because plenty of marketing copy is not: if you have fewer than fifty employees, the website provisions of that regulation do not bind you. Most of the businesses we work with are under it.

What does not change with headcount is the customer. A twelve-person clinic loses the same booking to an unlabelled field that a five-hundred-person one does. The law is a floor under the largest organisations, not a description of who benefits.

Checking your own site without a specialist

Four tests, none of which need software you do not already have, and none of which need anybody technical. Each one takes a couple of minutes and each maps onto one of the six failures above. Do them on a phone if you can, because that is where most of your visitors are:

1. Unplug the mouse. Press Tab repeatedly. Can you reach every link and button, and can you always see which one you are on? If focus disappears, so does a customer.
2. Turn images off in your browser settings and reload. Anything that becomes meaningless was carrying information in a picture with no text alternative.
3. Read it outside in daylight on a phone. Grey text on a white card is the most common failure on the web and the easiest to feel.
4. Complete your own form using only the keyboard, and submit it wrong on purpose. See whether it tells you what went wrong.

Our nine-point audit checks the labelled-field, alternative-text and document-language items automatically and publishes what it found. It is free and it will run on any site, including one we did not build.

The order to fix things in

Do the three that sit on the money path first: form labels, button text, contrast on anything showing a price or a phone number. Those are usually an afternoon, and they are the ones converting visitors into enquiries right now.

Alternative text next, because it is also how image search and AI answer engines read your pages. Then the rest. If a site needs the ARIA unwound before any of this is meaningful, that is a rebuild conversation rather than a fix conversation, and what it costs to fix a website AI built covers where that line sits.