Fake reviews in online shopping are a central threat to consumer trust and platform integrity. This paper presents a novel ensemble method that combines multiple classifiers based on RoBERTa, fine-tuned individually on a fake review detection dataset and tuned independently with a distinct metaheuristic algorithm: Artificial Bee Colony, Ant Colony Optimisation, Bayesian Optimisation, Firefly Algorithm, Grey Wolf Optimiser, and Particle Swarm Optimisation. Rather than selecting the best-performing single model, we equally weight all models and employ a weighted ensemble approach. The ensemble weights were tuned using BO to achieve maximum classification accuracy. The resulting ensemble had approximately 95% accuracy. We built a prediction pipeline showing more robustness than single models that lets new reviews be classified in real time, depending on the learnt ensemble.