How we measure, and where what we can establish comes to an end.
A report is only worth something if it holds when an opponent contests it line by line. This page says how your report is written, what no machine can verify on your site, and the mistake our own tool made.
Every criterion returns three outcomes, never two.
| Outcome | What it means for you |
|---|---|
| C | Checked, your page passes. Nothing to do. |
| NC | Checked, your page fails. A fact, readable in the code your server sends. Indisputable, and therefore fixable. |
| NT | Not testable. Our tool cannot see this criterion. It tells you so instead of counting it as a pass. |
About 30% of the criteria can be measured automatically. The remaining 70% need a human.
| Beyond any tool | Why |
|---|---|
| Real colour contrast | The page must be rendered to compute it. |
| The keyboard path | Someone must press Tab and watch where they land. |
| What your JavaScript adds afterwards | The page must be executed, not merely read. |
| Whether an alt text is apt | “image1.jpg” is an alt text. It describes nothing. Only a human sees the difference. |
| What a screen reader says | Someone must listen to it read your page. |
That figure is written on every report we deliver. An automatic measurement establishes failures; it never establishes compliance. Anyone selling you compliance on a tool's word alone is selling you a document that will not hold.
Your own code can contradict you.
The French order of 20 September 2019 requires the wording “Accessibility: fully / partially / not compliant”. We read it and set it against the measurement.
A site declaring itself fully compliant while its code carries an automatic failure contradicts itself in its own words. If that is yours, better to hear it from us than from a complainant: the statement then becomes the proof of the breach.
Because we publish our mistakes.
On its first run against the live web, our tool reported three failures on a large public-service site that declares itself fully compliant. All three were wrong: a link whose name lived inside an image, a hidden frame no screen reader ever meets, and two top-level headings the standard allows.
We do not name that site. We fixed all three causes and added fifteen tests that stop them coming back. A false accusation costs more than a missed failure: it makes you mobilise a team for nothing, and the day you find out, it is the whole report you stop believing.
Our own pages go through the tool and the result is published, failures included: our accessibility statement (FR). It says “not compliant”, because no human audit has taken place here either.
You can re-run the measurement without us.
The tool that produces your results is published. Your IT department, your agency or a competitor can run it against your site and get the same result — today, in six months, at your desk as at ours. No artificial intelligence takes part in the measurement: that is what guarantees two runs give the same figure.
A report whose computation cannot be replayed is not proof: it is a billed opinion. The measurement code · its fifteen tests.
Have your sites measured
Tell us how many sites, and by when. You receive the dated file within five working days: the record, the drafted statement, the remediation plan in order of cost, and the tool to contradict us.