Overview
Every accessibility scanner, app, and overlay is built on automated checks, so the honest question is: how much can automation actually see? The consistent answer across studies is a minority of the problem, and the less important minority at that. This page explains why, mechanically, without hand-waving.
Where the numbers come from
Several independent measurements converge on the same picture:
- Deque, the maker of axe, the engine inside most tools, found automation catches about 57% of issue instances. But instances count repetition: one missing alt attribute on a 1,000-product catalog counts as 1,000 instances. Coverage of distinct WCAG criteria is far lower.
- The UK’s Government Digital Service tested scanners against a deliberately broken page and found even the best tools caught roughly 30 to 40% of the planted barriers.
- WebAIM’s annual million-site scan detects failures on about 95% of home pages using automation alone, and manual audits of the same sites always find more on top.
So “machines miss most of it” is not a marketing line from audit vendors. It is the tool vendors’ own published data.
Why the ceiling exists
A machine can verify presence, not quality. It can check that an image has an alt attribute; it cannot know whether “IMG_2041” describes the product. It can check that a button contains text; it cannot know whether the focus order that reaches the button makes any sense.
Concretely, a scanner cannot answer:
- Does this alt text describe the image usefully? (1.1.1)
- Does focus move in a logical order through checkout? (2.4.3)
- Can a keyboard user actually escape this popup? (2.1.2)
- Do these error messages explain what went wrong? (3.3.1)
- Does the heading structure reflect the real page organization? (1.3.1)
Each question requires understanding intent, what the page is for. That is judgment, and every item on that list appears constantly in real complaints.
The skew that makes it worse
The missed share is not random. It skews toward the barriers that actually stop purchases. A slightly-low-contrast footer, which machines catch, inconveniences users. A checkout that cannot be completed with a screen reader, which machines cannot catch, excludes them entirely. Serial plaintiffs demonstrate the second kind, which is why overlay-equipped and scanner-clean stores still get sued.
The right division of labor
Use machines for what they are good at: catching regressions in the detectable subset, at scale, continuously, without getting tired. Use humans for what only humans can do: verifying that the store actually works with assistive technology, and producing the documented report that has legal weight. The combination is the method behind every serious audit practice, including ours.
Example
A store’s scanner dashboard shows zero issues for six straight months. An auditor then spends one hour with a screen reader and finds: a variant picker that never announces the selected size, a cart drawer that opens without moving focus into it, and error messages that only change a border color. Three purchase-blocking failures, zero of them machine-detectable, all three the kind of thing a complaint would describe in its first paragraph.
Common mistake
Interpreting “0 issues found” as “0 issues.” A clean automated scan means the machine-checkable third is clean. It says nothing about the two-thirds where the lawsuits live. Run the keyboard walkthrough and the difference becomes obvious within minutes.