This study examines how AI search assistants select and cite sources in election-related responses, revealing significant language-based differences in exposure to hostile-state information. Researchers analyzed 18,263 responses from four AI assistants (Claude, Gemini, OpenAI Search, and Google AI Overviews) during the 2026 Latvian parliamentary election. Russian-language answers cited hostile-state-owned or influenced sources in 34.6% of cases, compared with 9.8% in English and 4.0% in Latvian, with substantial variation between models. Exposure did not increase before the election, suggesting that differences resulted mainly from AI retrieval practices rather than coordinated interference. The covert Pravda network accounted for only 1% of hostile citations, while overt Russian state media and institutions dominated. These sources appeared most often on well-documented topics and were rarely identified as state-affiliated, despite generally factual answers. A comparison with the 2026 Swedish election confirmed similar Russian-language exposure (20.9%), indicating a broader language-specific retrieval bias. Latvia's Russian-language media environment further amplified this effect, reflecting Russia's longstanding influence efforts among Russian-speaking communities abroad. The findings demonstrate how AI assistants' source-selection practices shape voters' exposure to state-linked information, highlighting the need for greater transparency and safeguards.