Assistants answer the same question differently in different languages and regions, and not only because of translation. They search in the language of the question and read a different set of sources.
Measure per market
A share of voice number averaged across markets hides the thing you need to see. Prompts should be written in the target language by someone who speaks it, not translated from English, because the way a buyer asks differs more than the words do.
Localise the evidence, not just the page
Translating a page does little if the sources corroborating you are all English and all domestic. The directories, review sites and trade press that models read are market specific, which makes entity work the larger half of international AI visibility.
The classic plumbing still applies
One URL pattern for all locales, reciprocal hreflang with correct BCP 47 codes, no automatic redirects based on IP, and translated metadata including structured data. None of that is new, and all of it still decides which version gets shown.
Table stakes
- Prompts written natively per market.
- Per-market reporting, never a blended average.
- A single consistent URL pattern across locales.
- Reciprocal hreflang, including a self-reference.
- Local corroboration sources identified per market.