Study: AI search citations shift depending on query language

Computational researcher Mohamed Soufan asked ChatGPT Search, Gemini, Perplexity and Copilot 40 equivalent questions about Lebanon — across topics like economy, banking and health — in both English and Arabic, collecting 320 responses, per his own HackerNoon write-up. Lebanese sources made up 56.7% of citations in Arabic-language answers versus 31.6% in English for the same questions, a 25.1-point gap; Gemini showed the widest split between the two languages (34.8 points), ahead of Perplexity (26.8), Copilot (20.4) and ChatGPT Search (18.5). International organizations accounted for 21% of English citations but only 12% of Arabic ones, while news and media sources were 24% of Arabic citations versus 10% of English. Soufan notes the test captured each system at a single point in time with one response per combination, and did not assess whether the language-based differences affected factual accuracy.