The most common barriers in German online shops — 792 findings across 20 sites

The 792 findings are not spread evenly. Sorting the 20 sites by their number of findings gives this sequence:

0 · 2 · 5 · 7 · 9 · 10 · 13 · 13 · 22 · 23 · 27 · 35 · 35 · 39 · 55 · 57 · 58 · 73 · 152 · 157

The median is 25, the mean 39.6. The two largest sites alone account for 39 per cent of the entire measurement with 152 and 157 findings. Two sites of comparable size, with the same number of pages, can therefore be a factor of 20 apart — the number of findings hangs on the state of the theme, not on the size of the company.

One site had none. A mail-order house for artisan goods stayed at zero machine-detectable findings across seven tested pages — with 206 passed checks. Compliant sites exist, and they are not the largest or the most expensive ones. That is why we do not work with “you are all failing” but with “we will tell you where you stand”.

By severity, the 792 findings break down into 104 rated critical and 688 rated serious. No lower severities occurred in the entire run — which follows from the fact that the active rule sets target precisely those criteria whose violation actually impedes use.

430 of 792 findings are contrast failures. They occurred on 19 of 20 sites; the only one without a contrast finding was the same one that had no findings at all. Across all sectors, all sizes, all budgets — that is the clearest result of the entire measurement.

The requirement is not a matter of interpretation. WCAG 2.1 success criterion 1.4.3 requires a contrast ratio of at least 4.5:1 against the background for normal text, and at least 3:1 for large text. A contrast ratio is a number you calculate, not one you judge. So there is no argument about taste here.

Why it happens everywhere anyway: light grey on white has been a design fashion for years. It hits almost always the same places — secondary labels under product names, placeholder text in search fields, disabled button states, footer links, the small print under prices and the copy line in promotional banners. None of those stands out when designing on a bright, expensive screen. Every single one stands out on a phone in the sun.

There is a second figure that is rarely mentioned: on all 20 sites, additional contrast questions remained open as “not automatically decidable” — text over images, over gradients, over video. Including on the site that had zero findings. In those cases a tool cannot determine the background unambiguously and therefore reports not an error but an open question. Anyone looking only at the error count misses them.

62 findings on 5 of 20 sites are images without a text alternative. It is the only one of the four big rules for which no correction can be generated — not even with a language model.

The reason is banal and constantly ignored: whoever has not seen the image cannot describe what it shows. And a wrong alternative text is worse than none. It makes the finding disappear in the next test run without solving the problem — screen-reader software then confidently reads out something false. That is exactly the mechanism behind the poor reputation of automatically generated alternative texts in overlay tools.

There are exactly two cases that can be decided by machine: tracking pixels and spacers. An image of 1×1 pixel, or with “spacer”, “pixel” or “blank” in the file name, is safely decorative and gets an empty alt attribute. Everything else needs a human who looks at the image and knows why it is on the page.

For a quote this means: the “alternative text” item is the only one that grows linearly with the number of images. It cannot be optimised away and cannot be bundled into a batch correction.

An automated test covers only part of the criteria; about a third is the figure usually given. What this measurement fundamentally did not test:

• keyboard operability of the ordering process — the basket is opened but not filled, and no order is placed
• focus visibility and whether keyboard operation gets stuck anywhere
• reading and tab order — whether the order in the code matches the visible one
• the substantive correctness of existing alternative texts — a wrong one passes the test
• the understandability of error messages and whether they are tied to the field they refer to
• captions and audio description in videos
• and the statutory documentation itself: the information under Annex 3 and a working feedback route — documentation questions, not code questions

The overlap with what the statute demands of the authority is striking. Annex 1 no. 1 (a) BFSG requires that the sample examine all steps of a procedure, at least in the standard order followed by a typical user, and letter (b) expressly names forms, dialogue boxes, input confirmations and error messages. That is precisely the area an automated test does not cover — and precisely why a report without findings is not evidence of conformity. That sentence also appears in every report we deliver.

Two further sites could not be measured at all, for a reason that has nothing to do with the scope of testing: the request already failed the browser’s certificate check — on one the address requested did not match the certificate, on the other the issuing authority was unknown to the browser. We do not click past a certificate warning. On a third site only the home page was reachable: there a protective rule in the upstream network refused the automated browser, and overcoming an access barrier is a different thing from reading a publicly reachable page. All three are therefore missing from this analysis rather than represented by estimated values.