Visibility through the lens of language and geography

The AI100 corpus already contains an article about the “answer bubble” — the effect in which a brand looks different in ChatGPT, Google, Copilot, and Perplexity. But there is another layer of divergence that companies notice less often: the same brand, the same platform, but a different query language — and the answer changes radically. A company that confidently lands in ChatGPT’s top three recommendations in English may not be mentioned at all in the answer to the same question in Russian, Arabic, or Japanese.

This is not a bug. It is a consequence of how modern AI systems are built. A language model is trained on corpora of text, and the distribution of those texts across languages is extremely uneven. English dominates the training data of most major models. That means that for English-language queries, the model has a denser and more up-to-date map of entities, properties, and relationships. For less represented languages, that map is sparse: some brands are present in it, others are not, and still others are present but with distorted attributes.

The problem stopped being marginal once answer systems moved beyond English at mass scale. By May 2025, Google AI Overviews were operating in more than 200 countries and 40 languages [1]. AI Mode, launched in March 2025 in English only, expanded to nearly 100 languages by February 2026 through three major waves: November 2025 (35 languages), February 2026 (53 languages) [2]. ChatGPT serves more than 900 million weekly active users around the world, a significant share of whom do not work in English [3].

For an international brand, this means AI visibility has to be checked not in one language, but in every language in which its markets operate. One run in English says nothing about what the brand looks like in Spanish, German, or Chinese.

The first mechanism is asymmetry in training data. A brand that is heavily described in English-language media, reviews, and directories may be only weakly represented in the Russian-language or Arabic-language segment of the internet. The model knows it in one language, but literally does not “remember” it in another. This is not a translation issue — it is an issue of presence in the sources of a specific language segment.

The second is different web retrieval behavior. When an answer system uses web search to supplement an answer — as ChatGPT Search, AI Overviews, and Perplexity do — it looks for sources in the language of the query. If the brand has no high-quality content in that language, the system will find competitors that do. The user will receive a recommendation without your brand — not because you are worse, but because you are absent in that language.

The third is category drift through language. The same product can belong to different categories in different language cultures. What is described in English as an “analytics platform” may be interpreted in Russian as a “reporting service” or a “business intelligence system” — and each of those terms pulls in a different set of competitors and comparison criteria. The model inherits these categorical differences from the training data.

Language is not the only variable. Geography changes the answer too. Google AI Mode and AI Overviews take the user’s location into account when forming an answer [1]. Perplexity retrieves web sources with regional relevance in mind. That means the same English-language query from London and from Singapore may produce different results — with different competitors, different prices, and different recommendations.

For brands with several regional versions of a site, this creates an additional layer of complexity. You need to check not only “are we visible in English,” but also “are we visible in English from the United Kingdom,” “are we visible in English from India,” and “are we visible in German from Germany and from Switzerland.”

The problem is made worse by the fact that AI-system answers are stochastic by nature. According to SparkToro, the probability that ChatGPT or Google AI, across 100 repeated queries, will produce an identical list of brands in even two answers is less than 1% [4]. Every answer is a probabilistic sample, not a deterministic result. Changing the language and geography adds another layer to that stochasticity.

For diagnostics, this means that one run is not enough even within a single language. And for an international brand, the minimum meaningful diagnostic requires a matrix: several languages × several platforms × several repeats.

For a company operating across multiple markets, language and geographic diagnostics are not a luxury, but hygiene. At a minimum, the following makes sense:

Check visibility in every language in which the brand wants to be visible. Do not assume that a strong position in English carries over into other languages.

Create content in the target language rather than relying on automatic translation. Models are trained on human texts, and the language of machine translation often does not match the language of real demand in a given culture.

Make sure the external trust contour — reviews, directories, and industry media — covers the target languages. A brand with 50 English-language reviews and zero Russian-language reviews will be invisible to the model when the query is in Russian.

Add multilingual labels in Wikidata. This is one of the fastest ways to give AI systems an entity anchor across several languages.