We audited our own site. Eight defects.

Anyone who sells accessibility and does not deliver it on their own site loses the conversation in the first sentence. So on 23 August 2026 we audited our own site — against the same framework we apply for clients, and with the same findings report. This text is the short version. The full findings are open in the source of this site.

What was audited

Seventeen routes in both languages, in light and dark appearance, at 1440 × 900 and 390 × 844 pixels: 68 automated passes with axe-core across WCAG 2.1 A and AA. Plus, by hand, whatever a machine cannot decide — the whole path with the keyboard, the accessible names of every control, the heading and landmark structure, the behaviour of the appointment assistant when the step changes.

Result: eight findings, seven of them serious, none blocking. Every function was reachable — but at seven points made considerably harder. Three of them are here because they show the pattern.

1 · The focus was invisible

Three controls in the header — language switcher, appearance, WhatsApp — showed no visible outline when tabbed through. Measured: outline-style “none”, no shadow, nothing. They appear on every page. The cause was not a forgotten rule but one that lost: the global focus definition sat in a cascade layer and was overridden by a utility class that switches the outline off and replaces it with a ring this colour palette does not draw. Fixed by putting the focus rule outside any layer. Measured again: 2 and 3 pixels solid, all three visible. (WCAG 2.4.7)

2 · Eight stops before the content

There was no skip link. Anyone working with the keyboard alone landed first on the logo, then on four menu items, the language switcher, the appearance switch, WhatsApp and the main button — eight stops, on every page again, before the content began. That is not a cosmetic flaw; it is the difference between a site people use and one they leave. The skip link is now the first stop, it is visible as soon as it has focus, and it moves focus to the content. (WCAG 2.4.1)

3 · Pressing “next” and hearing silence

In the appointment assistant, clicking “next” swapped the content, scrolled to the top — and did not move the focus. Measured, it landed on the document body. For a keyboard user that means starting from the front again. For a screen reader user it means nothing at all: there was not a single region on the whole site that would have announced a change. You pressed “next” and heard silence. Now the focus moves to the heading of the new step, and a politely announcing region says which step of how many has begun. (WCAG 4.1.3, 2.4.3)

The remaining five

• The golden text colour fell below the required contrast on tinted surfaces at five points — values between 4.26 and 4.34 to 1, where 4.5 is needed.
• Text colours at 70 and 80 per cent opacity: down to 2.85 to 1 in the light appearance, seven places in the code.
• Placeholders in form fields at 60 per cent opacity: 2.40 to 1. Not reported by the machine, because a placeholder is only visible in an empty field — found by hand.
• For a screen reader, the 31 calendar days in the appointment assistant were named only “31, button” — no weekday, no month, no “selected” state.
• On the Turkish appointment page two labels had stayed German, hard-coded in the markup instead of coming from the dictionary. A Turkish screen reader reads “Nächster Monat” with Turkish pronunciation — nobody understands the result.

The most serious defect was older

It comes from an earlier pass and belongs here nonetheless: anyone who had switched on “reduce motion” in their operating system did not see parts of the site at all. Sections that fade in stayed invisible — 33 blocks on the services overview. The page was empty for exactly the group the setting is made for. A findings report that hides its own history is worth less than one that names it.

And a measurement error

After remediation the machine still reported two contrast values — both on pages whose German twin passed with identical markup. The cause was not in the page but in the measurement: the audit run caught the fading sections in the middle of the 0.9-second animation and measured a semi-transparent surface — a state no person ever sees. The run now measures with reduced motion, that is, the final state. It would have been convenient to write the two values off as outliers. A findings report that presents a measurement error as a result is exactly as wrong as one that hides findings.

What that means — and what it does not

All eight defects are fixed, in the code, without an overlay and without an add-on tool. Across 68 passes the automated run now reports no machine-detectable violation of WCAG 2.1 AA; before, there were eleven places. The manual review has no open item.

“No machine-detectable violation” does not mean “accessible”. By their own developers' account, automated tools find about a third of barriers. We did no pass with a blind user, tested no magnification software and no voice control. And this text is not a legal assessment — a lawyer makes that, not us.

That is why we do not put an automated scanner on this website, although it would be the best lure: a traffic light that a machine hangs over somebody else's page promises the same as an overlay — a result nobody stands behind. Anyone selling a traffic light cannot explain why the neighbour's traffic light is worthless.