Google Gemini Knows Me in Russian — But Recommends My Competitors in English

In early August 2026, the author of five Russian-language articles on Generative Engine Optimization (GEO) and AI Visibility for SEO conducted an experiment to test how Google's Gemini AI would describe them when queried by name. The author ran the same name in both Russian and English, expecting consistent results from the same model. However, the responses were vastly different, revealing potential issues in how AI handles professional identity across languages.

In Russian, Gemini produced a confident, well-structured biography with six citations, including the author's own agency site, industry publication, and major Russian business magazine. Four of the six links were accurate, while two were fabricated, attributing articles to unrelated parties. Despite the fabricated citations, the response maintained a high level of confidence and proper formatting.

Switching to English, Gemini opened by stating there was no prominent SEO expert with that name, listing four unrelated individuals instead. The model then offered recommendations for well-known Russian SEO experts, stating these were the people the author probably wanted to find. Notably, no sources were provided to support these recommendations.

This experiment highlights two failure modes in AI-generated content. The first, similar to classic hallucination, involves the generation of false claims with citations, making it harder to spot than classic hallucination. The second failure mode, less discussed, involves AI incorrectly identifying a real person as nonexistent and redirecting users to competitors. Both failures stem from the same underlying issue: RAG asymmetry across corpora and poor cross-lingual entity resolution.