Skip to content

Review

A scan settles what a machine can check. Most WCAG criteria also need a person: whether alt text describes its image, whether focus moves in an order that makes sense, whether a screen reader announces what changed. Review turns those into plain questions about real pages of your site, with the steps to test each one beside it, and records who answered what and when.

Review is on every plan. Your answers become the verdicts that the readiness reports, statement and VPAT are built from.

  • A finished scan. The scan picks the pages to check and settles what it can on its own. Until a site has one, Review says “Run a scan first.”
  • The right role. Owners, admins and members answer checks and record verdicts. Viewers can read the review but not answer. See Team and roles.

Open a site, then Review in its sidebar. The Sites list shows every site with how many criteria are ready, how many fail, and when a result was last recorded; sort it by Most review work left to see where review work waits first.

A site’s first full scan writes the checks (no AI is involved), and Review writes any still missing when it is opened. It offers a First session on one page, with an estimate of how long it takes. Press Start, or pick any check from the list.

Review tests a sample of your pages, chosen the way WCAG-EM recommends:

  • the home page;
  • one page for each template, recognised from the page’s address and its measured layout;
  • pages with the content a criterion is about, such as forms, video or sign-in fields;
  • pages your journeys visit;
  • as many other kinds of page as fit, then, with room left, a few at random.

Each criterion is checked on up to 6 sampled pages. A criterion about something only some pages have (forms, say, or video) is checked on every page that has it, up to 25. Language versions of the same page share one set of checks, except where the words matter.

Some parts are checked once, not page by page:

  • The site header, footer and site-wide scripts, once for each distinct header: the answer stands for every page that shares it.
  • Criteria about consistency across pages (2.4.5 Multiple Ways, 3.2.3 Consistent Navigation, 3.2.4 Consistent Identification and 3.2.6 Consistent Help), as one set of checks comparing pages.
  • A form or video that repeats on many pages, on the first page it appears on.

The Sample in Evaluation shows every scanned page, grouped by template, and why each sampled page is in the review. Add pages (up to 50) or take them out, then Update review. Take a sampled page out and another page of the same template takes its place; take out a template’s only page and the report says that template was not tested.

A check is one requirement on one page, asked as a plain question: what to try, and what you should see. Checks are asked only where they can apply: a check about captions appears only on pages where the scan found video or audio.

Open a check and How to test this gives the steps, for each tool that applies: the keyboard, the browser, NVDA on Windows, VoiceOver on the Mac, developer tools, or the live review panel. It also says when the check passes, what usually makes it fail, and links to the full guide to that criterion.

Checks are grouped by page. Two other kinds sit alongside them:

  • Scan findings to confirm. Things the scan flagged but could not decide. Your answer here is the issue’s answer in Issues too. Findings from checks that are not yet reliable enough are left out, and do not count as work.
  • Checks that need something you may not have, such as an account to sign in with, a test order or a phone. Review lists them so you can Set aside the ones you cannot do yet.

Review has three modes. Today is the checks to answer, with chips for AI can check, To retest, Problems, Couldn’t check and Done. Criteria is where each criterion stands, as a list or as a grid of criteria against pages, with chips for Waiting, Failing and Ready. Evaluation is what the review rests on and what goes out from it: its scope, sample, method and signature.

Filter, at the top of Today and Criteria, narrows them to the checks for a Developer, a Designer or Content, or to those assigned to you, and says so under the title while it does. It is only a filter: it changes nothing about who may answer, and Review’s progress and Evaluation always count the whole review.

Start a session asks one question at a time, in an order that keeps similar work together: keyboard checks first, then forms, navigation and so on.

Answer When What happens
Looks good It does what the check asks Recorded as a pass
Problem It does not You describe it, and it becomes an issue
Couldn’t check Something stopped you You say what; the criterion waits until someone can
Doesn’t apply here The page has nothing the check is about You say why (the report prints it), unless the scan’s own evidence shows it

From the keyboard, 1 to 4 answer, j and k move through the list, and ? shows every shortcut.

A Problem asks for what is wrong in one line, what you saw, how bad it is (critical, serious, moderate or minor) and, optionally, evidence: a PNG, JPEG, GIF, WebP or PDF up to 10 MB, pasted, dropped or chosen. It becomes an issue you can assign and track.

On a scan finding, the two answers are worded for the question (for contrast, Readable or Too faint): the problem answer confirms the finding, and the other dismisses it as not a problem.

After a Looks good, you can attach proof (a screenshot or a file); it goes on the report.

Found something no check asks about? Log a problem on the page, choose the criterion it fails and describe it. If the page has an open check for that criterion, Review takes you to it, so it is one issue, not two.

Set the browser, assistive technology and platform you are testing with in the bar at the top. It is sent with every answer you give until you change it, and the readiness reports cite it (“Firefox 131, NVDA 2025.1, Windows 11”). Without it they say “Environment not recorded.”

After each scan, on every plan, AI tries the open checks it can with no press, so you open Review to suggestions as well as questions. A pre-test works as Check with AI does below: a real browser, a proposal with screenshots, and nothing recorded until a person answers. It uses none of your AI checks.

  • They start with the site’s first full scan, which writes its checks. A site that has no checks yet for another reason gets them when someone opens Review, and pre-tests run after the scan that follows.
  • A check AI proposed an answer for in the last 30 days is not tried again.
  • Each site pre-tests at most twice the pages its review asks about in a calendar month, the pages with the most to try first.
  • Pre-tests may use only a share of the workspace’s monthly AI limit, so what people ask for always has the rest.
  • A site that scans behind a login is pre-tested only once someone turns on Pre-test signed-in pages in Review in its settings. The setting says what it sends, and to whom.

Suggestions arrive on Today, under AI can check.

On Growth, Pro and Enterprise, you can also ask AI to try a check and propose an answer. A person accepts it or answers differently; nothing counts until they do.

Plan AI checks a month
Starter Not included (pre-tests still run)
Growth 50
Pro 300
Enterprise 1,000

Choose Check with AI on one check or on a page’s group, or Check all n with AI at the top of Today. You confirm before it runs, with an estimate of the time and of the allowance it uses.

It opens the page in a real browser with the site’s scan settings (including any sign-in), tries the check, and comes back with screenshots of what it did. Any form it submits is stopped before it leaves the browser, so nothing is sent to your site.

The proposal says what AI suggests, how sure it is, and how often reviewers have agreed with AI on that check. Accept records it in your name; Disagree (or the other answer) records yours. “Suggested, not recorded. Your answer is the one that counts.” On Today, AI’s passes open where they are listed, so you can go through them together and accept the ones you agree with. The VPAT says how many of a verdict’s answers were accepted and how many of those checks were opened first, and a tick given without opening a check does not help that kind of check earn Tick all.

Some checks AI never tries, because they need a person: watching and listening to media, flash analysis, real sign-ins, payments, gestures and judgement calls among them. They stay in your list.

When every check of a criterion has an answer, Review offers a verdict:

  • Fails if any check found a problem;
  • Passes if every check looked good, or did not apply with at least one that looked good;
  • Not applicable if none applied.

A Couldn’t check holds the criterion back until someone can.

Record outcome saves it, with a note written from the checks; Review each lets you look first, and Later leaves it for now. In Criteria you can also set a criterion’s status yourself. A verdict has to agree with the evidence: a Fails needs a confirmed problem behind it, and a Passes cannot be recorded over open findings or a failed check. If a criterion is assigned to someone, only they can record its verdict.

An answer stays until a person changes or clears it. A scan never removes, moves or overrides one: when the checks are rewritten after a scan, every answer carries over to its check. Review shows a hint when an answer is getting old (after 90 days by default; change it in Done) or when the page’s layout has changed since, but the answer still counts until someone checks again.

A scan can still put a recorded verdict in question, for example with new issues on a criterion that passed. The reports then mark it requires reassessment, and the VPAT stops reading it as Supports.

When someone marks a problem you logged as fixed, its check comes back under To retest: “Looks good verifies the fix, Problem reopens it.”

  • Looks good closes the issue as verified.
  • Problem reopens it and adds what you saw as a comment.

Issues the scan found are verified by the scan instead, once two scans in a row do not find them. The readiness reports count each kind of retest.

A site’s Overview adds it up in one place, with nothing estimated and no time-saved figure:

  • Found: the problems found in the site’s standard, open or closed, on how many pages, and how many are new this month. What the scan could not decide, what an imported audit found and what someone dismissed are not in the number.
  • Fixed and proved: problems two scans in a row no longer find, or a person retested, and how many came back. A finding closed because we changed the check, or because its page was removed, is not a fix. Neither is one marked verified by hand with no retest.
  • Tested by people: checks answered and on how many pages, and criteria decided, as Evaluation counts them, with how many the scan decided on its own and how many answers accepted a suggestion without opening the check.
  • Machine went first: where pre-tests have run, the checks they pre-answered, how many of those people answered, and how many they agreed with. Agreed is not a measure of accuracy, and a tick nobody opened is left out.

Below them, Left is the criteria not tested and the problems to fix, and Next is the next useful thing with its estimate, in the order Today uses. On a new site it starts instead with what the scan did and the first few things worth doing, then the size of the rest. It also says whether pre-tests ran after the last scan, or why not.

Evaluation shows what a report on the site would rest on today: how many criteria are settled, failing and waiting, the scope, the sample (its pages, journeys and documents), the method, the signature and the documents built from it. Each criterion’s verdict is in Criteria. Export makes a readiness report. Only a verdict a person records is an assessment.