This study uses large language models (LLMs) GPT-4o and Gemini 2.5 Pro to simulate how geopolitical narratives diffuse and transform across synthetic social networks with different topologies: random, small-world, and scale-free. We measure narrative reach, semantic drift, and structural efficiency in each case. Results show that network topology significantly affects diffusion outcomes: scale-free networks enable wider but more stable narrative spread, while small-world networks amplify reinterpretation. Gemini promotes greater transformation of content, whereas GPT-4o favors stability. Our findings highlight the joint role of network structure and generative model behavior in shaping digital information flows and offer a framework for studying AI-mediated communication and disinformation.