AI search engines are delivering dramatically different quality answers depending on the language you ask in, according to a new audit of Claude's responses to travel questions. When travelers ask the same question in Arabic instead of English, they receive answers that are 45% less actionable, with 56% fewer named restaurants, mosques, and streets, and zero official sources surfaced in any Arabic response.

The study, conducted by Everything-PR, tested Claude (Anthropic's Opus-class model) on 20 travel questions a Muslim traveler would actually search for, each run as a matched pair in both English and Modern Standard Arabic. The findings expose a critical gap in how AI engines serve multilingual users, even when the underlying information exists in both languages.

Why Does Language Matter More Than Accuracy?

The research uncovered a counterintuitive finding: the Arabic responses were not less accurate than the English ones. Both languages scored identically at 82.5% on substance and factual correctness. What disappeared was not truth, but usability. An English speaker asking where to eat in New York gets specific street names like Steinway Street and Coney Island Avenue. An Arabic speaker asking the same question gets told to look in "Queens and Brooklyn." Both answers are correct, but only one helps you actually find food.

The AI Actionability Score, which measures whether a traveler can actually act on an answer, isolates four critical dimensions: whether an official source is cited, whether specific venues are named, whether geographic detail reaches the street or neighborhood level, and whether practical verification steps are provided. Substance, which measures only correctness, is excluded by design.

What Specific Gaps Did the Study Find?

The audit revealed stark differences across multiple dimensions of travel planning:

• Official Sources: New York City publishes a halal travel guide in both English and Arabic, distributed through travel channels in the Middle East and Gulf states. When asked directly in Arabic whether an official halal guide exists, Claude returned zero results. The English version identified the resource without naming the URL.
• Named Venues: English responses averaged five named locations per answer. Arabic responses averaged one. Borough names replaced street names, making navigation impossible for someone unfamiliar with New York's geography.
• Sourcing: When both languages cited the same statistic (approximately 285 mosques in New York City), the English answer attributed it to a 2015 census and noted that the city's tourism materials cite the figure. The Arabic answer provided the number with no source attribution.
• Lodging: Both languages performed worst on hotel and accommodation questions, scoring around 25% in English and 22.5% in Arabic. This is not a translation problem; it reflects an entire category with almost no retrievable structured data about Muslim-friendly hotels or halal room service.

Across all 20 question pairs, 17 favored English, three tied (all in lodging), and none favored Arabic.

How to Ensure Your Brand Reaches Multilingual AI Users

For organizations publishing travel guides, tourism content, or destination information in multiple languages, the study offers critical implications:

• Translation Is Not Distribution: Publishing content in a language does not guarantee it will be retrievable in that language through AI search engines. Tourism boards and convention bureaus running translated campaigns are buying reach into channels that may not be reading the translation.
• Structured Data Matters More Than Volume: Named venues, street addresses, and official source attribution are what AI engines extract and surface. Unstructured prose, even if accurate, gets summarized into vague borough-level guidance.
• English-Language Bias Is Systematic: The study notes that this is not unique to Arabic. The same English-language bias was documented in a parallel audit of higher education institutions, where Tsinghua University's citation share was estimated to be depressed by 20 or more points due to Western AI engine bias.
• Category Gaps Require Direct Action: Lodging, prayer facilities, and halal dining options are underrepresented in retrievable form across both languages. Hotels, airlines, and hospitality brands cannot rely on AI engines to surface their offerings without structured, schema-compliant data.

The research tested Claude specifically, but the methodology generalizes to any language pair and any destination. The implications extend far beyond travel: governments running multilingual investment campaigns, museums publishing in multiple languages, and restaurant associations listing halal-certified establishments all face the same retrieval gap.

What Does This Mean for the Global AI Search Market?

The audit arrives as AI search engines like Perplexity, Google AI Overviews, and Claude increasingly compete for user attention by surfacing answers directly rather than linking to websites. If an AI engine cannot retrieve and surface your information in the language a user asks in, your content effectively does not exist in that language's search layer, regardless of how thoroughly you have translated it.

The study estimates Muslim international travel at roughly 186 million arrivals in 2025, rising toward 245 million by 2030. That is the addressable market for travel content that is currently being underserved by AI search engines in non-English languages. The same principle applies to any segment of users asking questions in languages other than English: their answers are systematically less actionable, less sourced, and less useful, even when the underlying information exists.

For Perplexity, Claude, and other AI search engines, the finding raises a question about equity and market coverage. For brands and organizations publishing multilingual content, it raises an urgent question about whether translation alone is sufficient, or whether structured data optimization for non-English retrieval is now a competitive necessity.