Dialectal Variation in Generative AI: How Spanish Dialect Choice Shapes B2B Recommendations

This paper presents a two-phase empirical study on the impact of Spanish dialectal variation on generative AI responses in B2B commercial contexts. Phase 1 executed 50 B2B queries formulated in four Spanish dialect variants (Argentine voseo, Mexican tuteo, Colombian ustedeo, and Castilian Spanish) on Claude (Anthropic), producing 200 analyzed query-response pairs across ten enterprise software categories. Phase 2 replicated the experiment on Gemini (Google) and ChatGPT (OpenAI) for cross-platform validation, adding a neutral Spanish control variant and a direct assessment of 15 regional regulatory terms.

Results demonstrate that dialectal markers function as implicit geolocation signals that determine which companies AI recommends, which regulatory frameworks it references, and what tone it adopts. Response geo-relevance — the percentage of recommended companies with verified operations in the market implied by the query's dialect — ranges from 31% for neutral Spanish to 67% for Argentine Spanish with voseo. All seven original findings were confirmed across all three platforms, establishing dialectal geo-localization as a structural property of contemporary generative AI rather than a model-specific artifact.

The paper introduces the concept of regulatory blind spots — a distinction between LLM passive knowledge (ability to explain a regulatory term when asked directly) and active knowledge (spontaneous mention in context-relevant responses) — and documents cross-platform anchor brand convergence per market. Direct implications are provided for B2B AI visibility strategy in Spanish-speaking markets, including the finding that neutral corporate Spanish is a GEO anti-pattern for companies with specific geographic markets.

This working paper is part of the Exista.io Research Series on AI visibility in Spanish-speaking B2B markets.