On this page
Start from the goal, not the word countQuality KPIs: score errors, don't count opinionsOperational KPIs that predict troubleMarket KPIs: measure the Korean version against itselfNumbers that look good and mean littleA quarterly review that fits on one pageA company's Korean budget review usually starts with one number: what we paid per word. Then someone asks whether the Korean site is working, and nobody can answer, because the only thing anyone measured was the invoice. Translation can only support growth if you decide in advance what each piece of Korean content is supposed to achieve and pick a metric that shows whether it did, from translation quality metrics to market results.
This post is about that alignment: mapping business goals to translation KPIs, scoring quality in a way that survives a budget meeting, and avoiding numbers that look precise but tell you nothing.
Start from the goal, not the word count
Different Korean content exists for different reasons. A product page has to sell, a help article has to stop a support ticket, a contract has to say exactly what the English says, and a release note has to ship on the same day as the feature. One KPI cannot serve all four. The table below is a starting map; adjust it to your own goals.
| Business goal | Content it depends on | Primary KPI | Supporting signal |
|---|---|---|---|
| Win Korean customers | Landing pages, product pages, store listing, ads | Conversion or install rate of the Korean version | Bounce rate, search impressions in Korean |
| Keep Korean customers | Onboarding, UI, emails, in-app messages | Activation and retention of Korean-locale users | Support tickets that mention confusing wording |
| Reduce support cost | Help center, FAQs, error messages | Contacts per 1,000 active Korean users | Help article search terms with no result |
| Limit legal and regulatory risk | Terms, privacy notice, labels, contracts | Error score with zero critical errors | Number of open queries at sign-off |
| Ship globally at once | Release strings, release notes, docs | Share of releases shipped in Korean on day one | Turnaround time per batch |
Notice that only one row uses an error score as its primary KPI. For the others, translation quality is a cause, and the KPI measures the effect. That is why market KPIs need a quality floor alongside them: a page can convert well and still contain a wrong price.
Quality KPIs: score errors, don't count opinions
"The Korean reads badly" is not a metric. A reviewer's list of issues becomes one when every issue gets a category and a severity, and the total is normalized by the size of the sample. The most widely used framework for this is MQM (Multidimensional Quality Metrics). In its default scoring model, described by Lommel and colleagues, there are four severity levels with penalty multipliers of 0, 1, 5 and 25 for neutral, minor, major and critical issues. Penalties are added up, divided by the number of words evaluated, and expressed per a reference word count, commonly 1,000.
The pass threshold is your decision, not the framework's. The paper's worked example tolerates roughly 10 minor or 2 major errors per 1,000 words. A contract or a dosage instruction may tolerate no critical errors at all; a support article may accept more minor issues if it is accurate. Use the calculator below to see how severity drives the score.
Illustrative only. Weights follow the MQM default; set the tolerance per content type, and fail any batch with a critical error regardless of its score.
Two rules keep scores useful. First, review the same kind of sample each time, so this month's score can be compared with last month's. Second, keep a second linguist as the reviewer. ISO 17100 requires revision by a person other than the translator; the same logic applies to scoring. More on sample sizes and risk tiers in our post on sampling-based and risk-based translation QA, and on setting tolerance levels in defining fit-for-purpose quality.
Operational KPIs that predict trouble
Quality scores look backwards. A few operational numbers warn you before a launch slips:
- Turnaround per batch, from source freeze to translated files merged. Rising turnaround usually means unclear source text or missing context, not slow translators.
- Queries per 1,000 words. A high query rate is a source-quality signal. Questions like "is Charge a fee or a battery state?" point to strings that need context notes in English.
- Rework rate: how many delivered segments get changed after in-market review. If the same reviewer changes the same terms every month, the glossary is missing them.
- Day-one coverage: the share of a release that ships with Korean, not English placeholders.
A Korean-specific reading of rework. Reviewer changes on Korean UI often concentrate on two things: the politeness level of instructions and the treatment of English product terms (keep, transliterate or translate). Both are style decisions, not translator mistakes. If they dominate your rework rate, write the decision into the style guide once rather than fixing it batch after batch.
Market KPIs: measure the Korean version against itself
The tempting comparison is Korean conversion versus English conversion. It rarely isolates translation. Traffic sources, price, competitors and payment methods differ between markets, and W3C's own definition of localization covers far more than text: formats, currency, symbols, legal requirements and more. A gap between languages can come from any of them.
A cleaner test is Korean against Korean. Compare a page before and after a rewrite, or run two versions at once. Google Play's store listing experiments support this directly: a localized experiment can test the icon, screenshots, feature graphic and descriptions in up to five languages, with up to two variants against the current listing, measured on install or open clicks. For websites, the same idea works with any A/B testing tool. Our post on measuring localization uplift yourself walks through the setup, and app store optimisation for Korea covers what to test first in a listing.
For retention, the most practical signal is support data. Tag tickets from Korean users that mention wording, missing translations or confusing instructions. The pattern is described in more detail in how poor localization shows up in churn signals.
Numbers that look good and mean little
A quarterly review that fits on one page
- List the Korean content types and the business goal each one serves.
- For each, record one primary KPI and its value this quarter and last.
- Add the error score and critical-error count for content with legal or safety impact.
- Note turnaround and day-one coverage for release content.
- Pick one change to test next quarter, such as rewriting the onboarding emails or adding context notes to UI strings, and decide in advance how you will measure it.
The value is not the spreadsheet; it is the conversation it forces between whoever owns the budget and whoever owns the Korean market. Teams that run it consistently tend to move spend from low-impact volume to the pages that decide revenue. For how workflow maturity affects outcomes, see translation workflow maturity and business outcomes, and for a worked method on campaign returns, measuring the ROI of marketing translation.
Want Korean deliveries scored against a tolerance you set, with severity-tagged reports you can show your budget owner? See our translation services or contact us.