Online reviews play a key role in influencing consumer purchasing decisions and store ratings in the e-commerce world. However, the prevalence of review manipulation is a significant concern. Businesses regularly use spam reviewers or bots to either improve their reputation or harm that of their competitors. While many detection systems rely on English-language data, local languages like Bangla are largely underdeveloped. This gap is especially relevant as e-commerce grows rapidly in Bangladesh, necessitating a greater need for dependable review verification to protect consumers and businesses alike. In our study, we use multilingual transformer models on the Bengali Fake Review Detection (BFRD) dataset, which includes over 9,000 annotated reviews. We compare three models: MuRIL, selected for its pre-training on the related Indian Bengali dialect; the generative mGPT; and the lightweight Multilingual-E5. Our findings revealed that the mGPT model demonstrated exceptional capability, achieving a state-of-the-art 98% accuracy and a 95% F1-score. Following closely, MuRIL achieved an F1-score of 91%. Coming in last place, the compact Multilingual-E5 also proved its value, delivering a competitive F1-score of 88%. This result shows that mGPT works well in identifying fake reviews in Bangla. Multilingual-E5 is also a useful and scalable solution for real-world use. It also shows how important it is to have a range of transformer architectures in the digital marketplace of Bangladesh.