How AI handles brand queries in different languages

TL;DR: AI assistants like ChatGPT, Gemini, and Perplexity learn from training data that skews heavily English. The same brand query in French, Japanese, or Arabic often returns different recommendations, lower brand recall, or no mention at all. English content gets cited far more than equivalent content in other languages. Brands need original, structured content in each target language to stay visible across AI engines globally.

Why does the same brand query return different answers in different languages?

The short answer is training data. Every large language model learns from a corpus, and that corpus is not spread evenly across languages. English dominates. A 2023 analysis of Common Crawl, one of the primary pretraining datasets behind models like GPT and Llama, found English accounts for roughly 46% of all web content, with German, French, and Russian each under 6%, and most languages under 1% [1]. When a model sees a brand mentioned 10,000 times in English and 300 times in Spanish, its confidence in recommending that brand, its ability to describe it accurately, and its tendency to recall it at all all shift between those two query languages.

That's a real asymmetry. A user asking ChatGPT "what's the best CRM for small businesses?" in English gets a confident, detailed answer with named brands. The same question in Thai or Vietnamese can return vague or hallucinated responses, or the model defaults to a different set of brands that happen to have stronger footing in those language segments of the training data.

It isn't purely a volume problem, either. Retrieval systems like Perplexity also pull live web sources, and those sources get ranked partly by domain authority and link signals that again favor English-language sites. The model's "knowledge" and the retrieval layer's preferences compound each other. So a brand with strong English SEO but thin German content faces two headwinds on German queries: weaker model recall and weaker retrieval relevance.

For brands thinking about AI search visibility, this is the most underestimated gap in the field right now.

Which languages get the best AI brand coverage, and which get the worst?

The gap is bigger than most marketers expect. The Common Crawl 2023 breakdown is a reasonable proxy for where AI models have strong foundational knowledge [1]. English is the clear leader. After that, German, French, Spanish, and Russian carry meaningful weight. Japanese and simplified Chinese have moderate coverage, partly because major technology companies in those markets have contributed structured data. Arabic has grown but stays underserved relative to how many people speak it.

Here's a rough picture based on language share in major pretraining corpora:

| Language | Approx. share in Common Crawl | AI brand recall quality | |---|---|---| | English | ~46% | High, consistent | | German | ~6% | Good | | French | ~5% | Good | | Spanish | ~5% | Moderate-good | | Russian | ~4% | Moderate | | Chinese (simplified) | ~3% | Moderate | | Japanese | ~2% | Moderate | | Portuguese | ~2% | Moderate | | Arabic | ~1.5% | Low-moderate | | Hindi | ~0.4% | Low | | Swahili | <0.1% | Very low |

The practical consequence: if you're a US software company with 200 English blog posts and nothing in Hindi, an Indian user asking an AI assistant for a tool recommendation in Hindi is far less likely to hear your name. That user exists. India has over 500 million internet users [2], and AI assistant adoption there is climbing fast.

The "worst" languages for AI brand visibility aren't disadvantaged forever. They're disadvantaged because less authoritative content has been indexed and trained on in those languages. That's fixable.

Does AI translate your brand information, or does it treat each language separately?

Both, and the distinction matters. Modern LLMs are multilingual by design. They translate, and they show cross-lingual transfer, meaning knowledge learned in English can partly inform responses in other languages. But "partly" is doing a lot of work in that sentence.

Researchers studying cross-lingual knowledge transfer have found that multilingual models carry factual recall across languages, yet accuracy drops sharply for lower-resource languages even when the underlying fact was well-represented in English training data [3]. The model "knows" something in English, but the path from a Japanese-language query to that English-trained fact is less reliable.

For brand queries, this plays out two ways. First, brand names get transliterated or translated differently across languages. "Apple" becomes 苹果 (Píngguǒ) colloquially in Chinese, though the company name stays Apple. A lesser-known brand with no presence in Chinese-language sources may never get connected to its English reputation by the model. Second, brand attributes get distorted. A brand known in English for "best customer support in SaaS" may not carry that attribute in French queries, because French reviews and articles haven't made that claim at scale.

So no, AI does not simply translate your English brand reputation into other languages. It rebuilds your brand from whatever evidence exists in each language. If that evidence is thin, the reconstruction is thin or wrong.

Estimated AI factual accuracy by language (GPT-4)

| | | |---|---| | English | 80% | | French | 73% | | Spanish | 72% | | Arabic | 60% | | Hindi | 53% | | Swahili | 51% |

Source: Multilingual factual accuracy study, ETH Zurich, 2024

How do retrieval-augmented AI engines like Perplexity handle multilingual brand queries differently from pure LLMs?

Retrieval-augmented generation (RAG) systems, which Perplexity uses and which Google's AI Overviews also rely on, add a live web search layer on top of the base model. In theory this should shrink the language gap: ask in French, and the system retrieves French sources to ground its answer.

In practice it partly helps and partly creates new problems. The help is real. Brands with strong French-language content on authoritative domains do get surfaced in French Perplexity queries in ways they wouldn't from pure model recall. The new problem: retrieval systems still apply ranking signals that favor high-authority English domains. A brand's French microsite on a weak domain can rank below a passing mention in an English TechCrunch article once the retrieval layer runs its relevance scoring.

Perplexity has said its system fine-tunes on top of open-source base models and re-ranks web search results before passing them to the generation step [4]. The re-ranking algorithm hasn't been published in detail, but the effect is that topical authority, freshness, and structured data (like schema markup) all shape which sources build the answer. That's directly relevant to generative engine optimization strategy.

Google's AI Overviews are coupled even more tightly to Search quality signals. Google's documentation says AI Overviews draw from content that ranks well in standard Search [5]. So your multilingual Search performance is a direct input into your multilingual AI Overview presence. Fix the organic rankings in a language, and you improve the AI visibility in that language.

The bottom line: RAG systems respond to content strategy faster than pure LLMs do, which makes them the higher-priority target for multilingual brand work.

What happens to brand accuracy when AI answers in a non-English language?

Accuracy degrades, and it degrades unevenly across brands. A 2024 study from researchers at ETH Zurich examined GPT-4 factual accuracy across 12 languages using the same set of questions [6]. English accuracy landed around 80%. Spanish and French dropped to roughly 70-75%. Arabic fell to around 60%. Hindi and Swahili slid further, to around 50-55%.

Brand-specific queries weren't the focus, but the mechanism is the same. A model less reliable on general facts in Hindi will be less reliable on brand facts in Hindi. And brand facts tend to be more specific and niche than the general questions researchers benchmark on.

The errors cluster into a few types. The model may confuse your brand with a similar-sounding competitor that has more presence in the target language. It may name your brand correctly but attach the wrong features, pricing, or markets. Or it may drop your brand entirely and name competitors who did better multilingual work.

The omission case is the most commercially damaging, because you never see it unless you test. A brand could have 30% of its target customers in Germany and never know that German-language AI queries almost never surface it, because nobody on the marketing team runs German AI brand audits. Tools like AI visibility tools are starting to close this gap, but systematic multilingual auditing is still rare.

Does schema markup and structured data help AI find your brand in other languages?

Yes, more than most people realize. Schema markup, particularly Organization, Product, and FAQPage schemas, hands AI retrieval systems explicit facts that don't require inference from prose. If your German product page carries a Product schema with a clear name, description, and brand entity, a RAG system retrieving that page has clean structured data to work from instead of parsing and translating German prose.

Google's guidance on structured data says plainly that it helps their systems understand page content [5]. Since AI Overviews pull from Search-indexed content, and schema markup improves Search indexing quality, the line to AI visibility is direct. The same logic applies to Bing-indexed content feeding Copilot.

For multilingual sites, hreflang tags matter here too. They tell search engines which language version of a page to serve to which audience. If your hreflang setup is broken, Google may index your Spanish content as a duplicate of your English content, suppressing it from Spanish queries entirely. That's a technical SEO issue with direct AI visibility fallout, and it's surprisingly common on enterprise sites with large multilingual footprints.

Entity disambiguation is another structured data opportunity. Using schema's sameAs property to link your brand entity to your Wikidata entry helps AI systems tie your multilingual pages to a single authoritative brand. Wikidata has entries in hundreds of languages, and models are known to use it as a reference [7]. A brand with a complete, accurate Wikidata entry in 10 languages has a real edge over one with a sparse English-only entry.

This sits at the core of any solid AI SEO strategy for global brands.

How should brands audit their AI visibility across languages?

Start with manual spot-checking before you buy anything. Pick your five most important non-English markets. For each, write the five most common queries your target customer would type into ChatGPT, Gemini, or Perplexity in that language. Run them. Note whether your brand appears, where it appears, and whether the attributes the AI assigns are accurate.

Do this in the actual language, not a translation of your English queries. "Best accounting software for freelancers" in German is more than a word-for-word translation. A German freelancer has different reference points, different regulatory concerns, and phrases the query differently. Use a native speaker to write the test queries if you can.

What you're looking for in the audit:

• Brand mention rate: Does your brand appear at all in AI responses to these queries? A useful baseline. Track it over time.
• Position in response: Is your brand the first recommendation, third, buried in a list, or a parenthetical?
• Attribute accuracy: Does the AI correctly describe what you do, who you serve, and what sets you apart?
• Competitor comparison: Which competitors show up that don't appear in English responses? These may be locally strong brands you barely know.

For more systematic tracking, AI search visibility metrics and KPIs covers how to define and measure brand mention rates at scale.

Spawned's AI visibility audit covers multilingual brand queries across major AI engines, which helps if you want a baseline across languages without building the testing setup yourself. But manual spot-checking gets you most of the insight for free, at least on the first pass.

After auditing, prioritize by revenue exposure. If 40% of your pipeline comes from DACH (Germany, Austria, Switzerland) and your German AI visibility is weak, that's the fix with the clearest return. Don't try to solve 15 languages at once.

What content strategy actually improves AI brand visibility in non-English markets?

Translated content is necessary but not enough. AI models and retrieval systems reward original, authoritative content in the target language, not text that reads like it was machine-translated from English and dropped into a subfolder. Part of that is quality signals (engagement, backlinks from local sources, dwell time), and part is that native content matches native query patterns better.

Here's what actually moves the needle, in order of impact:

1. Local-language press and third-party mentions. The highest-leverage action. AI models learn brand associations partly from how brands appear in third-party sources: news articles, review sites, industry publications. A mention in a high-authority German tech publication like t3n or Heise is worth more to your German AI visibility than 50 translated blog posts on your own domain. Earned media in target-language publications is the single best investment for multilingual brand recall.

2. Original native-language content, not translations. Blog posts, help docs, and product pages written by or with native speakers, addressing local questions and concerns, beat translations. A piece on "Buchhaltungssoftware für Selbstständige" (accounting software for the self-employed) written for the German market with German regulatory context will rank in German Search and get retrieved in German AI queries. A translated American article probably won't.

3. Structured data on every page. As covered above, schema markup gives retrieval systems clean facts. Put it on your multilingual pages and keep it current.

4. Wikidata and Wikipedia presence. A Wikipedia article in the target language dramatically improves model recall accuracy in that language. Wikipedia content is heavily represented in LLM training data and in RAG retrieval. A legitimate Wikipedia article in French or Japanese requires meeting Wikipedia's notability standards, but for established brands it's worth the effort. Wikidata entries are easier to create and maintain.

5. Consistent entity signals across the web. Your NAP (name, address, phone) should match across Google Business, local directories, and your site for local markets. This helps entity disambiguation, which helps AI systems tie your multilingual content to one brand entity.

None of this is fast. Third-party press takes relationships and time. Original content in five languages is expensive. But the brands investing now are building a structural advantage in AI visibility that will compound over the next few years as AI assistant usage grows globally.

Does ChatGPT handle brand queries differently from Gemini or Perplexity across languages?

Yes, meaningfully, and the differences come from both architecture and training choices.

ChatGPT (GPT-4 and later) has been fine-tuned with multilingual RLHF (reinforcement learning from human feedback), which improves coherent, culturally appropriate answers in non-English languages. OpenAI hasn't published a language-by-language accuracy breakdown, but user testing and third-party benchmarks consistently show it performs best in English, French, German, Spanish, and Chinese, with more degradation in lower-resource languages [8].

Gemini (Google's model) has a notable edge in certain Asian languages, especially Japanese and Korean, because Google's search index and translation infrastructure give it richer training signal in those markets. Google's Gemini technical report named multilingual performance as a priority [9]. In practice, Gemini tends to perform better on Japanese and Korean brand queries than GPT-4 does.

Perplexity, leaning on live web retrieval, is more language-agnostic in theory, but retrieval quality varies by language. For languages with strong local web ecosystems (Japanese, Korean, German, French), Perplexity finds good local sources. For languages with thinner local web presence, retrieval quality drops and the model falls back on its base model's (often English-biased) training.

Claude (Anthropic) performs well in major European languages but has less documented multilingual tuning than Google or OpenAI. Anthropic has published less on this than competitors.

The takeaway: don't assume your AI visibility strategy is uniform across engines. If you're targeting Japan, Gemini deserves more attention than it might for Western Europe. If you're targeting Germany, Perplexity's retrieval layer means strong German content on authoritative domains counts for a lot. See AI-powered search features for a more detailed breakdown of how each engine works.

For brands serious about this, testing each major engine separately per language is the only way to find where the gaps actually are.

What do we actually know from research about multilingual AI performance and brand queries?

Honest answer: published research on AI performance mostly targets general factual accuracy, not brand queries. Nobody has published a large-scale peer-reviewed study specifically on brand mention rates in multilingual AI responses. So we work from adjacent evidence.

The strongest adjacent evidence:

The ETH Zurich 2024 study showed factual accuracy drops of 20-30 percentage points from English to lower-resource languages for GPT-4 [6]. Brand facts are a subset of factual recall, so this is a reasonable lower-bound estimate.

Research on cross-lingual knowledge transfer in LLMs found that models consistently underperform in non-English languages even when the underlying knowledge was learned from English text [3]. That's the mechanism, more than the symptom.

Common Crawl's published data distributions confirm the language imbalance in training data [1]. That's the root cause.

Google's Search Quality Rater Guidelines, which influence how AI Overviews are calibrated, emphasize E-E-A-T (Experience, Expertise, Authoritativeness, Trustworthiness) and evaluate those signals per-page, meaning a French page needs to establish E-E-A-T in French [11]. A French translation of a strong English page doesn't automatically inherit those signals.

Nobody has good data on exactly how much multilingual content investment converts to AI brand mention rates. The closest we have is the general relationship between organic search rankings and AI citation rates, which BrandRank.ai visibility insights analysis has started to document for English queries. Extending that to multilingual contexts systematically is still an open question.

Are there any quick wins for improving multilingual AI brand visibility?

A few things are genuinely fast and free, and worth doing today.

Wikidata first. If your brand isn't on Wikidata, add it. If it is, make the entry complete and accurate, and add labels and descriptions in your priority languages. This takes a few hours and directly improves entity disambiguation across AI systems.

Schema markup on existing multilingual pages. If you already have translated pages but no Organization or Product schema, add it. That's a developer-hours task, not a content production task.

Hreflang audit. Run your site through a crawler (Screaming Frog is common) and check for hreflang errors. Fixing them can quickly restore correct language-market matching in Search, which flows through to AI Overviews. This is one of those things that's often broken on large sites and never gets attention until someone looks.

Claim and complete local business profiles. Google Business Profile, Bing Places, and major local directories in your target markets give AI retrieval systems clean, structured brand data. For regional markets, local directories matter more than many marketers expect.

Pitch one local-language publication. Find the highest-authority tech or industry publication in your most important non-English market and land one genuine editorial mention. One article in Heise Online or Les Echos is worth more to your German or French AI visibility than months of translated blog content.

The slower, pricier work (original content and earned media programs) pays off over 6 to 18 months. The quick wins above can show up in AI responses within weeks, especially for retrieval-augmented systems that index fresh web content regularly.

For brands managing this at scale, AI SEO tools help track which content changes actually move brand mention rates across languages, rather than guessing.

Sources

1. Common Crawl, Language distribution statistics 2023
2. TRAI, Annual Report on internet subscribers in India
3. Cross-lingual transfer research in large language models (arXiv preprint)
4. Perplexity AI, How Perplexity Works (product documentation)
5. Google Search Central, Search Essentials and Structured Data documentation
6. Multilingual factual accuracy study, ETH Zurich, 2024 (arXiv preprint)
7. Wikidata, About Wikidata
8. OpenAI, GPT-4 Technical Report
9. Google DeepMind, Gemini Technical Report 2023
10. Google Search Central, Introduction to hreflang
11. Google Search, Quality Rater Guidelines