Translation workload

Read generated words, events, tokens, and repository changes as distinct measures of recorded localization work.

Workload reports show where recorded localization activity occurs. Words generated on Radar → Overview follows Lingo.dev's generated volume this month. Events in Explore counts recorded activity over time, by language, engine, or provider.

Choose the unit before comparing. A request can contain many words, token counts vary by language and model, and one repository commit can change many keys in many locales.

Words generated #

Translation volume sums words and requests from successful recorded translation requests. The underlying data is grouped by organization, UTC hour, localization engine, source locale, and target locale. Sum words or requests; counting rows would count hourly groups rather than requests.

The overview tile shows the month's words and compares the month so far with the same dates of the previous month, which it names, such as vs Oct 1 – 2. Reports use the organization's timezone, but the underlying rollup has UTC-hour resolution. Arbitrary minute boundaries and exact local-day boundaries in fractional-hour timezones cannot always be reconstructed from that data.

The population is successful recorded traffic. Failed requests, missing records, and external translations absent from this source do not contribute. Repeated requests and samples can increase the count without increasing unique content. Daily word generation also differs from completed multilingual jobs or published text.

The translation-volume read, POST /v2/organizations/:orgId/analytics/translation-volume, requires organization membership. The overview tile and other analytics reads also check the reports entitlement. If a view is unavailable, inspect Feature access and Roles and permissions rather than assuming that access to one report applies to all.

For period growth, compare complete aligned intervals and calculate:

text
(current volume − previous volume) / previous volume

A zero previous value leaves that growth ratio undefined. Report the newly observed volume instead. A spike can come from a release, broader locale scope, or repeated work; it does not diagnose quality on its own.

Event volume #

Under Data, choose Events to count each recorded event once.

Filter by Event type before counting: a localization's start and its completion are two events. Use Words generated on the overview for word volume.

Break events down by Event type, Target locale, Source locale, Engine, or Provider. For example, Events by target locale in Explore → Examples shows how the recorded work is distributed across languages.

Detailed token and cost analysis belongs on the Usage page. An event count, a word count, and token usage describe different units of work.

Repository changes #

A connected GitHub repository records changes to supported localization keys on its default branch, using the .lingo/config.json at each commit to recognize source and target files. Source files produce source.* events and target files target.* events, dated by the commit. Merge commits and file renames produce no key-change events. Live collection skips the App's own pushes; a history import can include its earlier translation commits.

In Radar → Logs, select Source → GitHub, then use Repository to narrow the feed. Open an event and inspect its repository, commit, file, locale, and key alongside the actual repository diff. One commit can produce many events, so an event count describes changed keys rather than a count of releases or commits.

A Commit history import can add earlier events. Collect GitHub activity covers the supported formats and history coverage.

Human edit measurements compare draft and target text. Repository events count changed keys. Neither a reduction in changed keys nor a reduction in edits demonstrates improved translations or saved review time without additional evidence.

Keep repeated comparisons consistent #

Use the same filters, units, and complete time buckets when revisiting a result. Overview summaries and Explore charts can aggregate differently, so preserve the formula rather than assuming every headline uses one ratio convention.

Rollup refresh behavior does not provide a guaranteed live-data watermark. Save the retrieval date with the result when the comparison needs to be repeated. Preserve a recurring analysis describes the calculation and population record to retain.