Counterfactual explanations offer a powerful mechanism for post hoc interpretability for machine learning models, enabling insight into how minimal changes to input features can alter unfavorable model predictions. Despite their growing adoption, most existing counterfactual generation approaches neglect the underlying causal structure of the data, often resulting in semantically invalid or physically implausible instances. This limitation undermines both the interpretability and real-world applicability of the generated explanations. In this work, we present a causality-aware methodology for counterfactual generation that explicitly models and leverages the causal dependencies among features. Given an observational dataset, we first construct a causal graph and learn feature-level Structural Equation Models (SEMs) to capture directional dependencies. For a given query instance, we generate perturbations guided by these learned causal mechanisms, thus constraining the counterfactual search space to causally valid regions. We evaluated our method on the Adult Income benchmark dataset, demonstrating that our approach produces counterfactuals that are not only effective in achieving prediction flips but also maintain high causal fidelity. We compared the counterfactuals generated with traditional metrics such as proximity, diversity, sparsity, and validity. Our findings highlight the necessity of integrating causal reasoning into counterfactual frameworks for robust, trustworthy, and semantically sound model explanations.