Three errors that proofreading the Italian will not find, from a QA on marketplace documentation

Case study · Version 1.0 · Published October 9, 2026 · Updated October 9, 2026 · by Alessio Di Rubbo

Three errors found in a DE→IT QA on seller documentation for a European marketplace have one thing in common: none of them shows up when you read the Italian. All three are grammatically correct and fluent, and two of the three would pass any automated check. They only become visible when the German source is aligned with the Italian text, element by element.

This page describes the method and the findings. It does not report results: references to the client have been removed and no performance data is published, for the reasons given at the end.

The context

Documentation for the sellers of a European marketplace, translated from German into Italian. Regulatory and procedural texts: terms, fees, legal obligations by country, return and take-back procedures. The kind of content where an error does not produce an odd sentence. It produces a different rule.

The review took place after translation, on files that had already been delivered and were partly published.

Finding 1: the threshold that disappeared

Severity: critical. The German source defines two consecutive tiers in a two-item bulleted list, and the higher tier carries a double fee.

The second Italian item is not a translation of the second German rule. It is an almost literal paraphrase of the first, with a single verb changed: different enough not to trigger an automated duplicate check, similar enough to give the copy away. The stricter tier, and the higher fee that comes with it, do not exist in Italian.

Both Italian items are grammatically perfect. Reading the Italian on its own reveals nothing. The gap only shows when the German list is aligned item by item and one threshold turns out to be missing.

I reread the raw HTML around that point to confirm that both items belonged to the same list and were not an artefact of my alignment script.

Finding 2: two translations of the same page, never reconciled

Severity: medium. For a single German source there were two independent Italian files: same title, same modification date, same source content.

They do not diverge through an oversight. They diverge systematically, as if two people who never spoke to each other had produced them.

One of the two versions puts the English and the Italian acronym side by side in the same brackets, where the Italian regulatory abbreviation, RAEE, is correct and sufficient on its own.

A typo is an accident. Two complete, independent translations in circulation are a process problem, and exactly the scenario that translation memories and binding glossaries exist to prevent. Anyone comparing two help pages sees two legal terminologies for the same concept.

Finding 3: the link that vanished from one market

Severity: minor. The source mentions two markets in the same sentence, both as active links. The Italian keeps the link on the first and loses it on the second: same string, plain text, no tag.

I did not find it by reading. The two versions are grammatically and semantically identical. The count of structural elements between source and target did not match, twenty-four blocks against twenty-three, and I traced the missing element by hand.

It is the only one of the three found through a structural difference rather than through reading. It takes the tools and the patience to compare the HTML element by element rather than the text line by line.

What I ruled out

The part of an audit that you do not see is the part that produces no findings.

The same pass checked and dismissed several suspects: a VAT table covering eight countries, a holiday calendar covering six, a fee table across twelve product categories, nine pairs of country and IBAN. All correct, row by row.

Two structural count mismatches in other documents remain unresolved and are logged as open, not as findings. A check that closes everything is a check that has stopped looking.

The time spent ruling things out is what separates a real audit from a list of suspicions, and it is also why a serious review costs more than proofreading.

What to take away

Three points that apply beyond this client.

• Checking the Italian on its own is not enough for procedural and regulatory content. Two of the three findings are invisible without the source alongside.
• Counting structural elements between source and target catches a class of errors that no proofreading finds: lost links, missing list items, skipped cells.
• Terminology has to be fixed beforehand, not checked afterwards. Finding 2 does not come from a wrong translation. It comes from the absence of a binding glossary at the moment two people worked on the same source.

Why there are no result figures

This page publishes no client metrics: no error rates, no volumes, no effect on support requests after the corrections.

There are two reasons. The first is consent: a client’s performance data is not mine to publish, and an anonymised case study that reports specific figures stops being anonymous much sooner than people think. The second is that I would have no honest way to attribute a change to the review: the corrections went into production together with other changes, and a figure presented as their consequence would be a reconstruction, not a measurement.

What I can show is the method and the findings. It is also the only part that carries over to your case.

Further reading

• How I assess the quality of Italian text: audit methodology
• German-Italian glossary: keeping your site’s terminology consistent for AI systems (in Italian)