Machine Translation Quality in 2026: What to Trust, What to Review

“Is machine translation good enough?” is the wrong question. The right one: good enough for which content? A practical confidence framework, honed on real sites.

What changed by 2026

Neural MT (and LLM-assisted translation) now passes the “unremarkable test” on most informational prose — readers don’t notice. DeepL leads fluency on European pairs; Google leads coverage. Where engines still stumble is intent: humor, persuasion rhythm, culture-bound references (untranslatables).

The review workflow that scales

Machine-translate everything; route by stakes. Your top-20 pages per language get a human pass in the visual editor — edits persist permanently over machine output, building a de-facto term base. Everything else ships and earns review only if traffic arrives.

Measuring instead of guessing

Watch per-language bounce and conversion. A locale converting at site average needs no further polish; one lagging badly earns an MTPE pass. Quality budgets follow data, not anxiety.