Alessio Di Rubbo

Italian Linguistic QA, independent review with documented findings

Service · Page published October 5, 2026 · Updated October 20, 2026

I check Italian texts that already exist and tell you, with evidence, whether they hold up. Agency translations, machine-translated or AI-generated copy, legal and checkout content. What you get back is a list of findings with a severity and a fix each, the corrected text, and a sign-off document you can file.

Who it is for

The service is for people holding an Italian delivery with no way to judge it.

  • Localization and content managers who received an Italian file and want an independent review with evidence, not a “looks fine”.
  • Brands producing Italian content with machine translation or AI who want to know where it gives way.
  • Legal and compliance teams in e-commerce, fintech or insurance who need documented review of Italian texts.

The recurring case is a grammatically sound delivery that nobody inside the company can assess, because nobody is a native speaker of Italian.

What the review measures, and what it does not

The review measures the linguistic quality of the Italian, from grammatical correctness through to naturalness. It does not verify facts and it does not replace a technical SEO or GEO audit.

There are six error categories, fixed in advance: agreement, register, calque, syntax, terminology, fluency. There are three severity levels: critical, medium, minor. The full method, with examples per category and its limits stated, is public: How I assess Italian quality (in Italian).

That the method is public is the point. You can read how I assess before trusting me with anything, and you can contest a finding by citing the criterion.

What you receive

  • QA report with findings grouped by content tier and severity, plus counts per category, because a single score hides where the problem is.
  • Corrected text in Excel: original, issue, fix, rationale. One row per finding, reusable by your team.
  • Compliance note for legal and regulatory text, where a terminology choice has consequences beyond style.
  • Formal sign-off document recording what was reviewed, what was found, and that the text is cleared for release.

Machine translation post-editing

If your Italian text comes from DeepL, another MT engine or a language model, the review doubles as machine translation post-editing (MTPE): every finding is corrected and you get the edited text back. Unlike a standard post-editing pass, you also see which errors the machine makes systematically, so you can adjust your glossary, prompts or MT settings.

I know machine translation from the evaluation side as well: since 2016 I have assessed the output of MT engines and language models for Translated, including on the ModernMT project. The recurring errors of ChatGPT and DeepL in Italian are described in this article (in German).

If you need full post-editing of every line, low-risk content included, we agree on it before starting, and the tiered review below then runs in full at every tier.

How I work

Triage and deduplication. I clean the file and group identical or near-identical strings. A typical batch of 650 strings becomes about 480 unique review items, and the work concentrates where it is needed instead of repeating itself.

Tiered review. Critical content is reviewed first and in full: payment, legal text, checkout. Account, onboarding and product content follow. Low-risk content I scan for systematic patterns rather than line by line, and I say so in the report.

Findings and corrections. Every issue carries the fragment, the category, the severity, an explanation and a concrete fix. The rationale is there so your team does not repeat the same error at the next release.

Sign-off. A formal document closes the work and stays on file.

What I need from you

  • The file in whatever format you have, raw is fine. Excel, XLIFF, a TMS export or a folder of strings all work.
  • The German or English source, if it exists. Without it I review the Italian alone and some error categories stay invisible: a threshold that disappears in translation only shows up when source and target are aligned.
  • The usage context of the strings where the file does not make it obvious. A button and a heading with the same text call for different choices.

Timing

Typically fifteen working days from input delivery, for a batch of ordinary size. For large volumes or fixed deadlines we agree the priority tiers before starting, so critical content is reviewed first even if the rest arrives later.

What I do not do

I do not rewrite texts the way I would have written them. The review corrects, documents and proposes, but it respects brand decisions already made: if the register is Lei, I flag inconsistencies within Lei rather than proposing tu.

I do not verify facts. If a product page states a wrong measurement and the Italian is flawless, my report will not catch it.

I do not issue a numeric score as the final measure. The score in the demo on this site is an estimate produced by a language model, useful as a quick indicator and not calibrated. The reference remains the list of findings.

Frequently asked questions

How do I know the findings are grounded and not opinions?

Every finding names the category and the criterion applied, and the method is published. The syntax and fluency categories depend in part on the reviewer's judgement, and I say so: for those I always propose a rewording, so the discussion happens between two concrete texts rather than over an impression.

Can you review AI-generated text?

Yes, and it is a growing part of the work. Italian produced by language models has recognisable recurring patterns: generic masculine where explicit agreement is required, subjunctive replaced by indicative, excessive gerunds, punctuation imported from English.

Do I need this if an agency translated it with internal review?

It depends what you need to demonstrate. A review internal to the agency that did the translation is not independent. If the check has to stand up to a legal department or to a client, the reviewer's independence is half the value.

What does it cost?

Agreed per batch, based on volume after deduplication and the share of critical content. A raw string count is not a good pricing basis, because part of it repeats.

Why not just correct the text with a free AI tool?

For a quick check before publishing it may be enough. It is not enough when the check has to prove something: an AI tool does not record what it checked or against which criterion, and the same request made twice can return two different answers. An independent review produces findings with severity and the criterion applied, which a legal department or a client can verify. AI-generated text adds a further problem: a model reviewing Italian written by a model risks letting through the very patterns it produces.

Send me one Italian page and I will tell you what a native reader notices. Book a free review.