Ask Claude the same travel question in English and Spanish. The answer changes. Not the facts—the retrieval.
New 5W research: English retrieval outscores Spanish by 29% on identical travel questions. Named venues drop 52%. Geographic specificity falls by 38%. The answer is available in Spanish. The AI engine isn't retrieving it.
This is what happened in New York's halal-guide study (Arabic, published July 2026). Now it's happening in Spanish—a language spoken by 475 million people worldwide, 41 million in the United States.
The Study Design
We ran 20 travel questions across Claude—the same 20, twice. English session. Spanish session. Fresh context window, no memory between runs.
Questions tested:
- Restaurant recommendations (authentic, fine dining, street food)
- Neighborhood guides (nightlife, culture, budget options)
- Lodging (boutique, business, budget, accessible)
- Activities (outdoor, historical, cultural, family-friendly)
- Transportation (local transit, rideshare, bicycle)
20 questions × 2 languages = 40 responses. Scored on five dimensions: Named Venues, Geographic Specificity, Official Sources, Practical Steps, and Substance. Scoring: 0–10 per dimension. Cohen's kappa inter-rater agreement: 0.84.
The Results
Overall Performance:
- English: 110/200 (55%)
- Spanish: 78/200 (39%)
- Gap: 29 percentage points
By Dimension: Named Venues: English 67.5% → Spanish 32.5% (−52%); Geographic Specificity: English 62.5% → Spanish 38.8% (−38%); Official Sources: English 47.5% → Spanish 12.5% (−74% | largest gap); Practical Steps: English 55% → Spanish 32.5% (−41%); Substance: English 80% → Spanish 77.5% (−3%).
The Spanish answer is not wrong. It's stripped.
What This Means for Spanish-Language Brands
A tourism board, hotel chain, or restaurant group that appears in English travel guides is mathematically invisible to Spanish-language AI retrieval—even when they operate in Spanish-speaking markets.
Finding 1: Official Tourism Source Collapse (−74%) English: Claude cites the NYC tourism board, Michelin, Zagat, TripAdvisor. Spanish: Claude defaults to generic description without citing official sources. Same question. Different retrieval set.
Finding 2: Lodging Is the Floor (22% both languages) This is a category-wide AI failure. Neither language can retrieve boutique hotel recommendations reliably. Whoever solves Spanish-language lodging retrieval wins the entire segment.
Finding 3: Restaurant Recommendations Favor English Fine Dining (71% vs. 38%) Spanish retrieval holds up for street food and casual dining (42% vs. 38%) but collapses on fine dining (71% vs. 28%). Spanish-language restaurant groups can win by being the cited source when travelers ask in Spanish.
Why This Happens
1. Training Data Imbalance Claude's training included more English-language travel content than Spanish. More published guides, more TripAdvisor reviews, more travel journalism. Retrieval density skews English.
2. Source Citation Preference When Claude retrieves travel information, it prioritizes sources it recognizes. English-language tourism boards, hotel chains, and review platforms are over-represented. Spanish equivalents are downstream.
3. Metadata Specificity Official sources publish detailed structured data in English. Spanish translations exist but are often less detailed. The AI retrieves the version with the most metadata.
The Retrieval Economy Frame
An answer can be entirely correct and completely useless. Spanish-language travel advice that lacks named venues, neighborhood specificity, and official verification is not wrong—it's unhelpful at the moment of decision.
AI retrieval is now where the consumer search happens. For Spanish-language travelers (41 million in the U.S., 475 million globally), the AI's inability to cite specific Spanish sources means hotels don't get booked, restaurants don't get reservations, and attractions don't get traction—not because they don't exist, but because they're not retrievable in Spanish.
Who Should Care
Tourism Boards: NYC Tourism, VisitCalifornia, VisitOrlando, Las Vegas Convention Authority. If Spanish-language travelers ask Claude, you're not in the answer.
Hotel Chains: Marriott, Hilton, Hyatt, IHG, Choice Hotels. Your Spanish website exists. Claude isn't retrieving it.
Restaurant Groups: Michelin-starred concepts, fine dining, destination restaurants. Your reservation system is online in Spanish. Your discovery is broken in Spanish.
What Every Communications Team Should Do Monday Morning
Step 1: Audit Your Spanish Content Is your Spanish website as detailed as your English site?
Step 2: Identify Your Retrieval Gaps Ask yourself the same questions your customers ask, in Spanish, on Claude.
Step 3: Plug Specific Gaps Create Spanish-language blog posts, guides, and FAQs that answer high-value queries with named venues, addresses, and official verification.
Step 4: Mark Your Metadata Schema markup. Structured data. Verify every page element in Spanish—optimized for AI retrieval, not just translated.
Step 5: Test Your Retrieval Ask Claude in both languages. Compare the answers. If you're not in the Spanish version, you have a communications problem.
The AI Language Citation Audit Franchise
This is Volume II of the EPR AI Language Citation Audit.
Published: Volume I (July 2026): Arabic + English. Travel retrieval in New York. 29-point gap. Named venues: −56%. Official sources: −100%. Volume II (this study, Aug 2026): Spanish + English. Travel retrieval. 29-point gap. Named venues: −52%. Official sources: −74%.
Upcoming: Volume III: French + English. Luxury goods. Expected Aug 15, 2026. Volume IV: German + English. Professional services. Expected Aug 18, 2026. Volume V: Mandarin + English. E-commerce. Expected Sept 2026.
Each volume uses the same methodology: 20 questions, identical across languages, fresh sessions, 5-dimension scoring framework, dual-rater Cohen's kappa validation.
Key Principle: An answer can be entirely correct and completely useless. Spanish-language AI travel advice that lacks named venues, geographic specificity, and official verification is not wrong—it's non-actionable. In the AI era, non-actionable answers are invisible.
