Everything you’re able to check about a translation is the part that carries no risk. Everything that carries risk is in a language you don’t read.
See the risk before the content goes live
Translation quality report
Release readiness
The problem
When a translated file comes back, you can verify a short list of things without knowing the language. That is useful, but it is not translation quality.
None of that has anything to do with whether the translation is correct.
What you can't verify is the rest. Whether a screw came back as a bolt — both real terms, both plausible, one of them wrong, and no way to tell which without knowing the language and the product. Whether a qualifying condition survived into the target sentence. Whether a should became a must. Whether a claim you're permitted to make in one market got stronger in another.
And when the visible layer does fail (broken tables, hanging elements on a file that was supposedly reviewed) that is a finding about the invisible layer too.
Why it persists
You receive a file and an assurance — not the findings, the severity, or the location, so there is nothing to audit. Teams describe the same pattern anyway: the file comes back, someone internally corrects the terminology, and nobody counts those hours. The consistency is also rented rather than owned. Change agency, or lose the translator who learned your product, and it starts over.
You can instruct a model to follow your terminology and it will — for a while. As the term list grows, instruction degrades; a glossary pasted into a prompt is a suggestion, and the model drifts back toward the general language it was trained on. Whether AI was involved is not really the question. Whether anything enforced your terminology while it ran, is.
A translation can be accurate and still be a claim you're not allowed to make in that market. No checker has an opinion about that, and it isn't what an agency was hired to look for.
Get a free quality report
Send one live URL in a language you are not confident about. We return findings with severity and location.
What we build instead
AI translates everything. Every segment is then scored against its source before a human sees any of it.
Thresholds can vary by content type. A safety instruction and an internal memo are not the same risk.
Below threshold, the segment is routed to a specialist. Above threshold, it moves forward.
Delivery
A gate only works if it sits in the path the content travels. That is what a supplier structurally cannot give you. An API call can.
The score becomes a condition in your release. Below threshold, it does not go out.
Agents retrieve approved terminology at generation time, so content is produced inside the rules.
Compose the gate into a workflow in a tool your team already runs.
For teams who just need a project done, with no integration.
How it works
The decision path is visible, repeatable, and kept with you after the file leaves.
Upload files or connect a CMS, PIM, documentation platform, or product feed.
Generate against enforced terminology and everything already approved, rather than from a blank slate.
Above threshold it moves. Below it, a specialist who reads that language takes it.
The file goes out; the terminology call, correction, and score stay with you.
Business outcomes
Every decision becomes evidence your team can use, export, and keep.
Specialist hours go to segments that scored badly, not spread evenly across a mostly fine document.
Terminology is enforced during generation, so the file does not come back for someone in-house to quietly fix.
Findings with severity and location, before publication.
Terminology and memory export in TBX, CSV, and TMX. Changing supplier is a decision, not a restart.
Built for complex content
Put attention where the risk is highest, without slowing every translation down.
The same procedure rendered consistently across every manual it appears in.
Full human review by design, not by whoever had capacity.
Catch an impermissible claim in that market before publish.
Check policy details against the source as facts, not just as sentences.
Read by someone in that language, not approved by someone trusting the fluency.
Keep the spec as the spec in every market you sell in.
Prove it before you believe it
Every segment scored, review concentrated where it is earned, every decision kept on your side.
FAQ
Usually not the things a checker finds. What survives is domain-level: a plausible but wrong technical term, an omitted qualifying clause, a shifted obligation, or a strengthened claim.
Yes, under a condition most setups do not meet: terminology must be enforced during generation from a maintained database, and every segment must be scored against its source.
It stops sentences your team already approved from being re-translated and re-argued each time they appear. Its value depends on whether what is stored was actually approved.
Without it, terminology is decided file by file by whoever translated it. Done properly, approved and rejected terms stay recorded and enforced when suppliers change.
