AreaAudit

Which languages each generator says it supports, and where

Coverage table: the artefact this register is trying not to be

A coverage table is a grid of vendors against language counts. Most of them are undated, mix units, merge three different capabilities into one column and record an absence of disclosure as an absence of capability. As of 2026-09-22.

Four failures of the usual grid, all of them editorialNo dates, units mixed inside one column, interface and output merged, and blanks read as refusals. None of the four is caused by the data being hard to get, and all four persist because a single ranked column is easier to read.Most common failureNo date on any figureCannot be re-checked or corrected. Free to fix.Units mixed in one columnRanking becomes meaningless. Visible in the vendor wording.Interface and outputmergedA translated menu counts as evidence of a voice.A blank read as noA disclosure gap is reported as a product limitation.Most consequential failure
Fig. 1 Removing them leaves a narrower table that answers fewer questions than a ranked grid appears to and more than it actually does.
Four failures of the usual coverage table, and whether each is avoidable. Recorded 2026-09-22.
FailureConsequenceAvoidable
No date on any figureCannot be re-checked or correctedYes
Units mixed in one columnRanking is meaninglessYes
Interface and output mergedA translated menu counts as a voiceYes
Blank read as noA disclosure gap becomes a product claimYes

Inclusion rule. The failures observable in published coverage tables of this kind. All four are editorial rather than technical. Order. Most common failure first.

1All four failures are choices

None of these is caused by the data being hard to get. Dates are free, units are visible in the vendor wording, the three capabilities are described on different pages, and a blank can be labelled as a blank.

They persist because a single ranked column is easier to read and easier to sell. The cost is a table that answers a question nobody asked while appearing to answer the one they did.

2The merge is the worst of them

Counting a translated marketing site as evidence that a model speaks the language is the error with the largest consequence, because the interface is the cheapest of the three to build and the most visible.

Among the vendors read here, the set with the longest voice lists and the set with the longest site lists overlap on very few companies. A merged column would put the wrong vendors at the top for either question.

3What is left when the failures are removed

A narrower table. Three columns instead of one, most cells empty, every figure dated, and no ranking at all. It answers fewer questions than the usual grid appears to and more than it actually does.

That is the trade this register takes. It is worth saying out loud, because a reader arriving from a ranked comparison will find this one frustrating before they find it useful.

A definition, not a measurement: no vendor figure appears here that is not also in the register. The three columns and their separate units are described on the columns page.

Nearby terms: Unit mismatch, Spot check. The whole vocabulary is at terms; nothing on this page is a claim about a named product.