Guide 03 · Evidence
What makes a finding defensible.
Any tool can produce a list of warnings in thirty seconds. Whether that list survives an expert reading it is a different question. Here are the four things that decide it — useful for evaluating any vendor’s work, including ours.
Which page, exactly.
A finding must name the page that actually rendered it. This sounds obvious and is the most common quiet error in automated accessibility work.
Requests redirect. A tool asked to examine one address may be served a different page, and unless it follows the redirect through to attribution, the report cites a URL that never displayed the barrier. An expert who visits that address and finds nothing wrong has just been handed the argument.
Ask a vendor: does attribution follow redirects to the page that rendered?
When, and in whose time zone.
Websites change, often weekly. A finding without a date describes nothing in particular. A screenshot without a date is a picture.
Time zone matters more than it looks. A finding recorded at 11pm on the 30th in one zone is the 31st in another, and a document that renders dates in the wrong zone invites an argument about whether the examination happened when it claims.
Ask a vendor: is every finding dated, and in which time zone are those dates rendered?
With what, at what version.
Accessibility rule engines change between releases. A rule can be added, tightened, or withdrawn. A finding produced by an unnamed tool at an unrecorded version cannot be reproduced, and evidence that cannot be reproduced is difficult to defend.
The same is true of viewport size. A barrier that appears only on a phone-width layout is real, and it needs to be recorded as such, or a reviewer opening the page on a desktop will conclude the finding was wrong.
Ask a vendor: are engine and rule-set versions recorded per examination and printed in the document? Is the viewport recorded against each finding?
How many were discarded, and by whom.
This is the one that separates evidence from output, and it is the section an opposing expert reads first.
Automated evaluation produces candidates. Many are duplicates of a single underlying defect; some are wrong. Rigorous work normalizes them and then has a person confirm or discard each one. A document that shows all three counts — produced, normalized, confirmed — is telling you how much judgment went into it.
A report that shows only a final number has not told you whether anyone looked. A report where the three numbers are identical is telling you nothing was discarded, which in this field is not plausible.
Ask a vendor: how many candidates did the tools produce, how many survived review, and who reviewed them?
What good evidence still does not establish.
Automated evaluation finds a minority of real barriers. No tool judges whether alternative text is accurate or whether an error message can be understood. A page with zero automated failures has not been shown to be accessible, and any vendor claiming otherwise is overselling.
A confirmed finding is not a legal conclusion. It establishes that a barrier was present on a page at a recorded time, measured against a published standard. Intent, notice, harm and whether the failure amounts to a violation are legal questions that belong to the attorney, not to the vendor.
A vendor willing to state both limits plainly is more likely to be worth relying on than one who is not.
Hold us to this.
Every question in this guide has an answer in our methodology, and every brief we deliver prints its own review record.
This guide is general information and not legal advice. Aldermarq is not a law firm.