The Number That Should Change Your Roadmap
Research in 2026 puts the share of non English AI citations won by localised pages at between 85 and 96%. Read that as the practical rule it implies: if somebody asks a question in German, Arabic or Portuguese, the answer is built almost entirely from pages published in that language, not from the best English page translated on the fly.
You are competing with everyone in English and with almost nobody everywhere else. Every serious competitor in your category has an English blog. Very few have a Portuguese one.
Why the Gap Is So Wide
Three things compound, and none of them are about model capability.
Retrieval happens in the query language
The pass that selects candidate documents runs before the answer is written. A page that does not exist in the query language mostly is not in the candidate set to begin with.
Translation is lossy in the quotable direction
A model can translate your English page to answer in Spanish, but then it is paraphrasing rather than quoting, and paraphrase attribution is weaker than a direct lift.
Almost nobody publishes
The competitive field in most languages is thin enough that a single competent page can own a category question, which is simply not true in English any more.
Machine Translation, Honestly Assessed
Machine translation is good enough to be indexed. It is frequently not good enough to be quoted well, and the difference matters because being quoted is the objective.
A model lifting a sentence from a machine translated page repeats whatever is wrong with it, including a mistranslated product name, a claim that reads as hedged when it was not, or a register that sounds like a manual when it should sound like a person. You do not see this happen and you cannot correct it after the fact.
The economical compromise
Machine translation for breadth across the catalogue, then a human pass on the small number of pages that carry your definitions, your prices and your limits. Those are the pages that get quoted, and they are usually short.
Which Languages to Publish In First
Start where you already have users, not where the market is largest. This is the opposite of how localisation is usually budgeted and it is the reason most localisation budgets produce nothing.
Read your own analytics by language, not by country
Country is a poor proxy. Language is what the query is written in and it is what the retrieval pass matches against.
Read your support inbox
People writing to you in a language are people trying to use the product in that language. That is demand you can see today rather than demand you are forecasting.
Pick two, not eight
Two languages done properly beat eight done through an unedited translation pipeline. You can always add the third once the first two are earning citations.
Translate the answers, not the marketing
Definitions, comparisons, limits and pricing. The homepage translated into six languages changes almost nothing, because nobody asks an assistant to describe your homepage.
What to Translate, and What to Leave Alone
| Page type | Translate? | Why |
|---|---|---|
| Definitions and explainers | Yes, first | This is what gets quoted, in every language |
| Comparisons and alternatives | Yes | High intent and almost no competition outside English |
| Pricing and limits | Yes | Most requested facts, and short enough to do properly |
| Homepage | Eventually | Persuasion does not translate into citations |
| Company news | No | Nobody asks an assistant about your announcements |
One Domain, Not Several
Do not spread languages across separate country domains unless you have a legal reason to. Separate domains split your authority, multiply your maintenance and give each new site the cold start problem you already solved once.
Subdirectories under one domain keep everything in one place. What actually has to be right is that each language version is a real page at its own URL rather than a client side switch, because a language that only appears after JavaScript runs does not exist to most of the agents that matter here.
What We See Running Tools in a Hundred Languages
We run text to speech tooling that covers more than one hundred languages, and the pattern in the logs is consistent enough to plan around. Usage is heavily non English. The pages that earn citations are the ones that exist in the user's language, and they earn them at volumes that the equivalent English page never reaches, because in English the same page is one of four hundred.
The part that makes this urgent
This is the least contested surface in AI visibility right now, and it is contested precisely in proportion to how many companies read a statistic like the one at the top of this page and act on it. That number is currently very small. It will not stay that way.