In one sentence: a second language needs its own prompt set written natively, its own competitor set, and its own measurement series — because translation smooths away exactly the phrasing that determines whether you are found.

The mistake that wastes the whole effort

Teams translate the site, translate the question set, and then compare the numbers across markets as if they measured the same thing. They do not. A German buyer asking about robots.txt validation types robots.txt testen, not a translation of an English phrase — and we saw 134 impressions on that exact query before writing a single line of German.

Write the prompt set in the market's language from scratch. If you translate it, you are measuring how your translated site performs against translated questions, which is a closed loop that tells you nothing.

Competitor sets differ by market

The vendors an engine names in one market are frequently not the ones it names in another — local players, regional resellers and incumbents with no international presence all show up. Declare the competitor set per market. Carrying an English-market set into a new language produces a share-of-voice number computed against companies nobody there is choosing between.

The technical layer

  • hreflang must be mutual. Every version points at every other, including itself. One-directional hreflang is the most common multilingual mistake and it wastes the translation budget that produced the pages.
  • Declare alternates in the sitemap rather than scattering them across page heads — one place to audit, one place to break.
  • One URL form across all languages. Whatever your trailing-slash and www decisions are, they apply everywhere, or you multiply a duplicate-URL problem by the number of languages.
  • Self-host fonts and check reachability from the market — a page that is slow or blocked from where the buyers are is not a content problem.

Do not machine-translate your way to page count

The temptation with a large site is to run everything through translation and multiply the URL count. This is the fastest way to turn a language version into a liability: low-quality translated pages at scale look exactly like the scaled content Google's spam policy describes, and they drag on the rest of the site.

The workable shape is layered. Programmatic pages — comparisons, references, structured data-driven pages — carry their copy in data files, so a new language costs a few hundred strings and the quality stays controlled. Long-form content is translated selectively: the thirty to fifty pieces that matter, done properly, and the rest simply not published in that language. A version with fifty good pages beats one with three hundred mediocre ones, and it is also cheaper.

Measure per market, never blended

Keep a separate series per language, with its own baseline. A blended number hides the only thing you wanted to learn — whether the new market is working — behind the volume of the established one.

Which market first

Pick from evidence you already have, not from market size. Search Console shows impressions by country before you translate anything; that is a free signal of where demand already exists for your category in your current language. Then check whether that market has an established local competitor set, because entering a market where engines already name three local vendors is a different project from entering one where they name nobody.