Lost in Translation: How Language Changes Which Brand AI Recommends

Most of us who track our brand's presence in AI responses do so with a specific bias: we test in English. It feels like the default, the baseline. But for a global audience, English is rarely the only lens through which a product is discovered. If a potential customer in Singapore asks a chatbot for a recommendation in Simplified Chinese, or someone in Riyadh asks in Arabic, there is a common assumption that they are simply getting a translated version of the English answer.

The reality is more complex. AI doesn't just translate; it reasons and composes within the specific language of the query. This means your brand's visibility isn't a static fact across the globe, but a variable that shifts based on the language the user speaks. In markets like Saudi Arabia, Kazakhstan, and Singapore, U.S.-based models like ChatGPT, Claude, and Gemini dominate the referral traffic, yet the answers they provide can diverge wildly depending on the tongue used to ask the question.

The Hidden Influence of Language in AI

Brand visibility in AI is not a one to one transfer from English to other markets. The challenge begins with accessibility. For instance, while ChatGPT and Gemini are widely used, they aren't available in China, and Claude isn't officially offered in Russia, Iran, or mainland China. In those regions, visibility depends on homegrown models like Naver in South Korea, Mistral in France, or Baidu's ERNIE in China.

These regional models are built on entirely different foundations, including national corpora and specific cultural identities, sometimes even government mandates. Even when using Western models, the "values" of the AI, the factors that shape a response when there is no single right answer, can shift. This might manifest as a change in the level of caution, warmth, or brevity in the response.

Expert Interpretation: The tradeoff here is between global scalability and local relevance. Many brands attempt to manage their AI presence through a single, English centric strategy, but this creates a blind spot. The decision to ignore non English AI visibility is essentially a decision to ignore how the AI perceives your brand through different cultural and linguistic datasets. You cannot assume a "global" reputation exists in a vacuum; it is fragmented by the training data of the specific language being used. The same pattern also shows up in We Earned 1, where the practical question is how the signal becomes visible.

The Methodology: Testing 1,650 AI Responses

To move beyond theory, a test was conducted using six common shopping and booking scenarios. These included looking for noise canceling headphones under £400, project management software for a 50 person UK firm, and return flights from London to Tenerife. The goal was to see if the AI's recommendation changed based solely on the language of the prompt.

The initial testing focused on Claude Sonnet 5 with web search disabled. This was a critical choice because it forced the model to rely entirely on its internal training data and memory, removing the "noise" of live web results. Using the DataForSEO LLM Responses API, the model was queried across five different languages. Each scenario was run 40 times per language, resulting in a core dataset of 1,200 responses. To add further depth, two of these scenarios, hotels and project management software, were also tested with web search enabled and across other models like ChatGPT and Gemini.

Expert Interpretation: Testing with web search off is the only way to uncover the "latent bias" of a model. When the web is on, the AI is essentially performing a real time search. When it is off, you are seeing what the AI "believes" about your brand based on its training. For a brand manager, this is the difference between knowing your current SEO strength and knowing your ingrained brand authority within the model's weights.

Where the Data Diverges

The results revealed startling disparities, particularly in the travel sector. When asked for a flight from London to Tenerife, the language of the query completely flipped the recommendation. In Persian, Claude recommended British Airways 52.5% of the time, while easyJet was picked only 12.5% of the time. However, in Russian, the script flipped: easyJet was the top choice at 72.5%, and British Airways was not recommended a single time.

The same model, the same question, and the same timeframe produced opposite results simply because the language changed. In the hotel category, the results were more varied because there are more options in Tenerife, but the gaps remained. For example, Bahía del Duque was a top pick in 43.6% of Chinese responses but only 7.7% of Arabic ones. Hotel Botánico, a frequent winner in English, vanished entirely in Russian responses.

Expert Interpretation: This proves that AI recommendations are not objective truths but probabilistic outputs influenced by linguistic context. The risk here is "invisible loss." A brand might be winning in English speaking markets but be completely erased in other high value languages without ever knowing it. If you only monitor your primary language, you are operating with a partial map of your market share.

Decoding the "Why" Behind the Shifts

It is difficult to pinpoint exactly why some categories remain stable while others swing wildly. In categories with a few dominant global leaders, there is a broad consensus that transcends language. But where consensus is lacking, the AI's internal associations take over.

The Russian flight data is a telling example. British Airways was completely excluded in Russian queries. This could be linked to the February 2022 ban on UK registered aircraft in Russian airspace and the subsequent suspension of routes. However, this doesn't fully explain why easyJet and Jet2, also British airlines, dominated the Russian recommendations. The question isn't just why one brand was excluded, but why another was elevated so aggressively in that specific linguistic context.

Ultimately, choosing a language is not just about choosing a vocabulary. It is about providing the model with a layer of context that shapes the final output. The language acts as a filter that can either amplify or mute certain brand associations.

Expert Interpretation: This suggests that AI models may be associating certain languages with specific geopolitical or cultural events stored in their training data. The tradeoff is that the AI is trying to be "helpful" by providing what it thinks is a relevant answer for a speaker of that language, but in doing so, it may apply outdated or skewed logic. Brands need to inspect whether their "absence" in a language is due to a lack of data or a specific negative association within the model.

The Impact of Live Web Access

While the internal memory tests are revealing, most users interact with AI via web enabled search. When the test was expanded to include live web searching for hotels and project management software, the results shifted again. In seven out of ten language and category combinations, the winning brand changed once the AI could browse the open web.

In the hotel category, the top recommendation changed across all five languages. Even more surprising was that several hotels that were frequently recommended from the model's internal memory disappeared entirely once the AI had access to live web data. This indicates a disconnect between what the model "knows" (training data) and what the model "finds" (current web index).

Expert Interpretation: This creates a dual front battle for brand visibility. You have to manage your "training data" presence (which is long term and harder to change) and your "live web" presence (which is more akin to traditional SEO). The fact that live search changes the winner suggests that real time optimization can override the model's internal biases, providing a potential lever for brands to reclaim visibility in specific languages. This connects with Category Framing Changes Which Brands AI Recommends when the same signal needs a clearer operating decision. A useful companion note is Which Brands Are Vanishing from AI Search?, because it looks at a nearby part of the same system.

Comparing Claude, ChatGPT, and Gemini

To determine if this was a Claude specific quirk, the same scenarios were tested in ChatGPT and Gemini with web search disabled. The results showed that while the phenomenon is widespread, it is not uniform across models.

Gemini matched Claude's recommendations in five out of ten combinations. ChatGPT, however, only matched Claude in one. For project management software, Monday.com was a consistent winner across most languages in both Claude and Gemini. But in ChatGPT, Monday.com didn't win in a single language. Instead, other brands that were mere also rans in the Claude tests took the lead.

Expert Interpretation: This highlights the danger of "model dependency." If your brand is highly recommended by Gemini but ignored by ChatGPT, your visibility is fragile. The decision here is whether to optimize for a single "dominant" model or to diversify your content strategy to appeal to the different training sets of the major LLMs. Since each model weights data differently, a one size fits all approach to AI optimization is likely to fail.

Understanding the Scope and Limits

It is important to acknowledge the constraints of this data. The findings are based on a snapshot of 1,650 responses gathered on a single afternoon. While this is a significant sample for a manual test, it is a tiny fraction of the billions of queries processed daily. The volatility in the hotel category, for instance, is likely due to the sheer number of available properties in Tenerife, which naturally spreads the recommendations thinner.

Despite these limits, the core hypothesis was confirmed: brand recommendations vary based on the language of the user. The numbers may fluctuate, but the trend of linguistic divergence is clear.

Expert Interpretation: The limitation of this research is actually a call to action for brands. Because the data is volatile and model dependent, you cannot rely on a third party report to tell you where you stand. The only way to get an accurate picture is to conduct your own linguistic audits. The tradeoff is the time and resource investment required to test across multiple languages and models, but the cost of remaining invisible in a key market is far higher.

Moving Forward with Multilingual AI Visibility

The central question, whether a customer in Singapore and a customer in Kazakhstan get the same recommendation, has a nuanced answer. Sometimes they do, but often they don't. Language is a powerful variable, but it interacts with other factors like model architecture and live web access.

The most significant takeaway is that you cannot know your true AI visibility until you look beyond English. The fact that live web access can change the winning brand is a glimmer of hope for marketers. It suggests that while we cannot easily rewrite the internal weights of a model, we can influence the information the AI retrieves in real time.

The next step for any global brand is to move from passive observation to active testing. This means identifying the primary languages of your target demographics and auditing how the major AI models perceive and recommend your brand in those specific tongues. The goal is not just to be present, but to ensure that the "reasoning" the AI uses in Arabic or Chinese is as favorable as the reasoning it uses in English.