Journey audits
Per-page scanning proves pages are accessible in isolation. Journey audits prove a user can complete a task: a real browser walks your flow (sign up, search, add to cart, checkout) step by step, measures the state at the end of each step, and reports a step timeline, so the answer is “checkout blocked at step 3”, not just a violation list.
A journey is measured once per step, so the steps you write are the states that get checked. Read What a journey measures before writing one: it is the difference between a journey that finds what you were looking for and one that quietly finds nothing.
Available on Pro and Enterprise plans. Configure journeys per site: open the site, then Find and Journeys in its sidebar. Owners, admins and members edit and run; viewers can read results.
Anatomy of a journey
Section titled “Anatomy of a journey”A journey has a start URL and up to 12 labelled steps. Each step is a short list of actions from the same safe set interaction scenarios use (click, fill, press key, hover, wait for element, wait), plus two journey-only actions:
- Go to URL navigates within the site (same origin only).
- Assert checks the current URL and/or that the page contains given text, so a broken flow fails loudly at the exact step instead of timing out later.
Every step that completes gets an accessibility scan of the state it reached, a screenshot, and a violation count. Actions inside a step are never scanned individually: the scan happens at the step boundary. The run ends early at the first step that cannot complete, marked “journey blocked here”; the steps after it are not measured at all.
What results mean
Section titled “What results mean”- A run’s outcome is completed (every step passed) or blocked at step N.
- Step findings become tracked issues tagged with the journey and step, and with the WCAG criteria they fail. They get the full issue workflow (assignment, comments, tracker sync) and resolve automatically when a later run of the same step no longer reproduces them. They are counted apart from the page scan’s results.
- A step reports only what it introduced: the violations already on the page when the step began belong to that page’s scan. See What a journey measures for how axe, transition and probe findings are each attributed.
- A step that was blocked or skipped keeps its issues untouched: absence of evidence is not evidence of a fix when the step never executed.
Agentic journeys
Section titled “Agentic journeys”On Enterprise plans a journey can run in agentic mode: instead of exact actions, each step carries a goal (“add any product to the cart”) and on every run an AI model picks the actions from what is actually on the page. This survives redesigns that would break a scripted selector.
Guardrails, since an AI is choosing the clicks:
- The model can only pick from the same safe action set, and only target elements that exist on the page at that moment; it cannot navigate, run scripts, or invent selectors.
- At most 8 model calls per step; a step that cannot reach its goal is reported as blocked, exactly like a scripted failure. If the model returns something outside the safe action set we ask once more, and that retry counts against the same 8.
- Every run records the concrete actions the model took, so the timeline reads the same as a scripted run.
- Decisions are made deterministically where possible, but a page that changes between runs can produce different action paths. For flows that must be byte-for-byte repeatable, use scripted mode.
Journeys by ear
Section titled “Journeys by ear”A journey by ear is walked the way a screen reader user moves through a page. Each step carries a goal, as in an agentic journey. On the first run an AI model hears only what the virtual screen reader says, and presses only the keys a screen reader user has: next and previous item, headings, links, landmarks, buttons and form fields, find, Enter, Tab and Escape. When every goal is met, the journey keeps a checkpoint for each press: the line the reader said on the control, and what pressing it gave (an announcement, focus moving, or a new page). Every run after that replays the checkpoints with no AI.
- A replay fails where the ear cannot go on. The reader never says the control’s line (it lost its name or its role), or pressing it now says nothing where it used to be announced. Each step shows what the reader said.
- Check what it pinned. The walk keeps every press, including one it did not need (adding the same item twice, say). Open the step to see its checkpoints, and remove one you do not want.
- A failed checkpoint is walked again. If the walk gets through another way, it proposes new checkpoints for you to apply; if it cannot, the failure stands.
- It verifies nothing. The virtual screen reader is a model of what a screen reader is given, not NVDA, JAWS or VoiceOver. Use a journey by ear to spot obvious blockers on a route, and test the route with a screen reader to verify it.
- The walk is slow and uses AI. A model call takes about 14 seconds, so the first walk takes a few minutes, with at most 24 model calls a step and 48 a run. On Pro a journey by ear walks once and then replays; walking it on every run is an Enterprise feature.
Running journeys
Section titled “Running journeys”- Manually from the site’s Journeys page (up to 10 runs per site per day).
- Automatically after each full-site scan via the per-journey toggle. A journey with a run already queued or running is skipped, so scans can never stack runs.
- No new step starts more than 4 minutes into a run (8 minutes for a journey by ear, whose model is slower), and no single step runs longer than its own budget. A step already under way when that mark passes is allowed to finish rather than being cut off (a half-measured step would report nothing at all), so in the worst case a run comes to rest around 7 minutes in (about 11 for a journey by ear). Exceeding a budget ends the run cleanly as blocked. These are ceilings for a pathological site: a typical run finishes in well under a minute, and a first walk by ear in a few.
Safety
Section titled “Safety”Journeys perform real actions: forms get submitted, accounts can be created, orders can be placed. Treat them accordingly:
- Point journeys at a staging environment where possible.
- Never enter real payment data; journeys stop at cross-origin handoffs (payment providers) by design.
- Every journey request carries an
X-TAC-Journey: 1header so you can identify and filter this traffic in your analytics or WAF. - Logged-in flows can reuse the site’s encrypted authenticated- scanning credentials via the journey’s “use credentials” toggle.
- Runs are never retried automatically. If the infrastructure fails mid-run, the run is marked failed and you decide whether to run it again.