# Accessibility testing for web applications: what differs from a website
URL: https://pandus.online/en/accessibility-testing-web-applications
Updated: 2026-10-21

How to test the accessibility of a web application or SaaS: the states a crawler cannot reach, axe in CI, checks behind the login, a keyboard script and live user signals.

On a website the page is the unit: you open a URL, the content is there, a checker reads it. In a web application the **state** is the unit. The same URL shows an empty list, a loading spinner, a filtered table or an error, depending on what the user did. The sign-in page is the only thing a crawler sees, and it is rarely where the problems are.

This guide is for teams that build SaaS products, portals and internal tools. It assumes you know the [WCAG 2.2 checklist](https://pandus.online/en/wcag-2-2-checklist-online-shops) and want to know where to point it.

## Where web apps break

Six places account for most of the blockers we see in applications:

1. **Route changes without focus management.** In a single-page app the URL changes, the content changes, but focus stays on a link that no longer exists. A screen reader user hears nothing. The fix is to move focus to the new view’s heading or main region and to update the document title.
2. **Dialogs.** A dialog must take focus when it opens, keep Tab inside, close on Escape and return focus to the control that opened it. Most custom dialogs fail at least one of these; see the [modal dialog pattern](https://pandus.online/en/accessible-modal-dialog).
3. **Toasts and live updates.** A message that appears for three seconds in the corner is invisible to a screen reader unless it is in a live region. Status messages are a WCAG 2.2 criterion (4.1.3).
4. **Custom controls.** Comboboxes, date pickers, sliders and toggles built from `div` elements have no role, no name and no keyboard behaviour unless someone adds them. The [select and combobox pattern](https://pandus.online/en/accessible-dropdown-combobox) shows the expected behaviour.
5. **Tables and lists with actions.** A row of icon buttons (edit, delete, download) with no accessible name reads as “button, button, button”. Virtualised lists that render only the visible rows confuse screen readers that read the whole list.
6. **Forms with inline validation.** Errors that appear next to a field but are not linked to it, or are announced before the user finishes typing. The [form errors pattern](https://pandus.online/en/accessible-form-validation-errors) covers the wiring.

None of these are exotic. They are the normal building blocks of an application, and every one of them is invisible to a check that runs on the sign-in page.

## Layer 1: automated checks in CI on real states

Run the open axe-core engine in your end-to-end tests, in the state you want to check: after signing in, after opening the dialog, after applying the filter. Playwright with `@axe-core/playwright` is the common setup; our guide on [accessibility tests with Playwright](https://pandus.online/en/accessibility-testing-playwright) has the code and the rule set that matches the Pandus scanner, so CI and the weekly check agree.

Three habits make this useful rather than noisy:

- Fail the build only on **serious and critical** results; keep the rest as a report.
- Test **states, not pages**: one test per dialog, per empty state, per error state.
- Check the **document title and focus** after each route change in the same test.

## Layer 2: the signed-in screens in a browser

CI covers what you thought to test. The [Pandus browser extension](https://pandus.online/en/accessibility-checker-browser-extension) covers what you have open right now: a staging environment, the admin area, a customer’s account with real data. It runs the same rules, explains each problem in plain words and copies a note for the developer or a prompt for an AI assistant. The tab-order and focus tools show the order keyboard users will meet, which is the single most common thing to be surprised by in an app.

## Layer 3: ten minutes with the keyboard

Automation does not judge whether the order makes sense or whether the toast was heard. A short script does, and it fits in a sprint review:

1. Unplug the mouse, open the sign-in page and sign in with Tab and Enter only. Can you see where you are at every step?
2. Open the main dialog of the product. Does focus move into it? Does Escape close it and bring you back?
3. Trigger a success message and an error. With a screen reader running (NVDA, VoiceOver), are they announced?
4. Use the main table or list: sort, filter, open a row, use a row action. Did you need the mouse anywhere?
5. Change a route from the navigation. Where is focus now, and what does the title say?

Wherever you got stuck, users with a keyboard, a switch device or a screen reader get stuck too. Write the five answers down; they are the manual part of your accessibility page.

## Layer 4: live signals from real users

The three layers above test what you decided to test. The [Pandus SDK](https://pandus.online/en/accessibility-analytics-sdk) reports what users actually met: keyboard focus trapped on a screen, focus lost after a dialog closed, a control without a name that people kept clicking, a form error that was never announced, a layout that broke when someone zoomed. Each signal carries the page, the element and the WCAG criterion, and the signals are counted, so the first fix is the one that hits the most users. On a sample of page views the SDK also runs axe-core on the real state of the app, with every text replaced before anything leaves the browser.

After a fix, the scanner checks again and the live signal is watched for a set number of days. “Confirmed by real visitors” is a better status than “closed”.

## What to write down

Buyers of web applications ask for evidence. Keep three things current:

- **An accessibility page** with the date of the last check, the standard (WCAG 2.2 AA), the known limitations and a way to report a barrier. The [accessibility statement guide](https://pandus.online/en/eaa-accessibility-statement) has a template; Pandus drafts it from a questionnaire.
- **An Accessibility Conformance Report** in [VPAT format](https://pandus.online/en/vpat-accessibility-conformance-report) when public bodies or large organisations are among your customers. Pandus produces a draft from the automated check on Agency plans; the manual columns need a person.
- **The list of open problems** with owners and dates, exported to Jira or GitHub, so the next sprint has it.

Automated checks, the extension and the SDK find part of the problems; the keyboard script and a manual review find the rest. We say on every report that it is an automated check and not a full audit, and nothing here is legal advice.

Q: Can an online accessibility checker test my web app?
A: Only the pages it can reach without signing in, in their initial state. For the app itself you need a check that runs where the state is: in CI with a browser that signs in, in the browser extension on the open screen, or in the users’ browsers with the SDK.

Q: How much does automation find in a web app?
A: Automated rules find a part of the problems, the same as on a website: contrast, names, labels, roles. Focus order, focus management in dialogs and route changes, and whether a toast is announced need a person with a keyboard and a screen reader.

Q: Where do custom components fail most often?
A: Dialogs that do not trap and return focus, date pickers and comboboxes that work only with a mouse, tables with icon buttons that have no name, and toasts that disappear before anyone hears them.

Q: Do we need a VPAT for a web app?
A: If you sell to public bodies or large organisations, buyers will ask for an Accessibility Conformance Report in VPAT format. Pandus produces a draft from the automated check on Agency plans; a complete ACR needs a manual review.
