Xbench alternative for online translation QA

What's Wrong is an online translation QA tool: the form checks you know from Xbench, imported checklists included, then a reading of the translation for meaning. On 12 of our delivered jobs it found 16 real errors Xbench never flagged and left out 71% of its false positives. Keep Xbench for project-wide term searches and files that must stay offline.

I have used Xbench for more years than I like to count: first as a project manager at an LSP, then running operations, then in my own company. It carried a big part of our QA and it caught real errors: a wrong number, a broken tag, a term that drifted from the glossary.

It also flagged a lot that was not wrong. If you have ever scrolled through a 2,000-line Xbench report with the delivery due in an hour, you know what happens next. You skim. And skimming is when the one real problem gets through.

What Xbench checks has hardly changed in years, and nothing in it reads a sentence for meaning. That was fine when the main risk was a tired translator. Today much of what we review was drafted by a machine, and machine output fails differently. The tags are right and the numbers are right. The sentence reads well. And it says something else.

Xbench and What's Wrong side by side#

What we measured on our own files#

We did not build a test file for this page. We took 12 jobs we delivered in 2026 (14 files), Turkish and Arabic, that still had the Xbench report next to them. Half of the reports carry the reviewer's own label on each finding: false positive or fixed. Then we ran the same files, before the fixes, through What's Wrong on production with the AI review on. About 22,000 source words in all. On the two largest files the AI review read a sample of the segments, so the counts there are a floor.

• 16 real errors Xbench never flagged. A reviewer would have changed every one of them: wrong terms that read correctly, a dropped list item, a typo. Six more sat inside segments Xbench had flagged for something unrelated, so the reviewer had looked at the segment and moved on.
• 71% of Xbench's false positives did not appear at all. Of the findings the reviewers had labelled false positive in the six labelled reports, seven in ten were simply not in the What's Wrong report. Most of the rest came with the AI disputes label, which keeps them visible and takes them out of the way.
• With the client's checklist imported, nothing Xbench fixed was lost. The 18 errors the reviewers fixed from Xbench were client preferences and mechanical slips: straight quotes, end punctuation, an unclosed parenthesis, a glued word. With the client's own Xbench checklist imported into the profile, What's Wrong caught all 18 on the same files. Without it, our default profile missed most of them, which is the point of the import.

That is the honest shape of it. Xbench is good at form. What's Wrong is good at form and reads for meaning. Updated November 2026, from our own delivered files; the raw reports are not public because they hold client text.

Three errors Xbench did not flag#

The first comes from one of our delivered files, with the names changed and the error kept word for word. The other two are written for this page with the same kinds of error. None of them has a tag, a number or an untranslated word for Xbench to compare.

1. Tenants that became licences (English to Turkish)

• Source: Single workloads and tenants can be isolated, not only whole regions.
• Target: Yalnızca tüm bölgeler değil, tek tek iş yükleri ve lisanslar da yalıtılabilir.
• Xbench: flagged the segment for a checklist rule only; the reviewer marked it ignore.
• What's Wrong: "The term 'tenants' in a cloud computing context refers to 'kiracılar', not 'lisanslar' (licenses)." Suggested fix: Yalnızca tüm bölgeler değil, tek tek iş yükleri ve kiracılar da yalıtılabilir.

2. A safety step that went missing (English to Turkish)

• Source: Before cleaning the Brandel heater, unplug it from the mains and let it cool down completely.
• Target: Brandel ısıtıcıyı temizlemeden önce tamamen soğumasını bekleyin.
• Xbench: nothing to flag. The sentence that is left is correct; the instruction to unplug is simply not there.
• What's Wrong: "The instruction to unplug the device from the mains is missing from the target." Suggested fix: Brandel ısıtıcıyı temizlemeden önce fişini prizden çekin ve tamamen soğumasını bekleyin.

3. A warranty nobody offered (English to Turkish)

• Source: The Brandel charger works with every Brandel Pro battery.
• Target: Brandel şarj cihazı tüm Brandel Pro bataryalarıyla çalışır ve ömür boyu garantilidir.
• Xbench: nothing to flag. The added clause promises a lifetime warranty the source never mentions.
• What's Wrong: "The target adds information about a lifetime warranty that is not present in the source text." Suggested fix: Brandel şarj cihazı tüm Brandel Pro bataryalarıyla çalışır.

Where Xbench is still the better tool#

• Searching a whole project. Looking up how a term was translated across 40 files at once is where Xbench shines, and we do not try to compete with that.
• Files that must never leave your computer. Xbench runs locally. If a contract forbids any online tool, Xbench is the honest answer.
• Years of checklists. Many teams have built up checklists over a decade. They still work, in Xbench and, if you import them, in What's Wrong too, where each check becomes a rule you can edit.

Where What's Wrong goes further#

• It reads for meaning. After the rule checks, a language model reads each segment next to its source and the segments around it, looking for mistranslations, omissions, additions and the wrong term or register. That is where the 16 errors came from.
• It argues with itself. A second model reads each critical or major rule finding that needs a reading, such as a term, a number or a checklist hit, and says whether it agrees, disputes it or cannot tell. A disputed finding is never hidden; it stays in the report with the reason and counts at half weight until the reviewer decides. This is our answer to the 2,000-line report.
• It tells you why. Each finding comes with a one-line reason and a suggested fix. Accept the fixes you want and download a corrected copy, or make the change in your own CAT tool. Your original is never touched.
• Your rules are one list. Import an Xbench checklist and every check becomes a rule of its own, next to the rules you write as a sentence in plain words. Test a rule on a file before it runs and switch off the one that keeps firing on nothing.
• It works where the translators already are. Inside the editor in Phrase, XTM Cloud and Smartling, in Crowdin as an app, from a Dropbox folder, by email to a profile's own address, from an AI assistant through the MCP server or by uploading a bilingual file.
• The report can go to the client. A reviewer signs it off as pass or fail; the signature records their name, their role and the score at that moment. For a formal LQA there is a scorecard with penalty points per 1,000 words and the arithmetic shown.

If you already use Xbench#

Keep it. Keep it for the terminology lookups and the checklists you built over the years. Then take one file you have already delivered, run it through What's Wrong and put the two reports next to each other. That will tell you more than this page.

There is a free plan, no card needed. See the plans. If you want the longer version of how a model reads a segment, it is in What is AI translation QA.