IoT-based soil sensing combined with machine learning (ML) can already recommend a suitable crop from soil and weather readings, with reported classification accuracies from about 91 % to over 99 %, but none of the five re-viewed systems recommends a seed variety. Soil health governs productivity, yet farmers still rely on infrequent laboratory tests and general advice, which delay decisions and ignore field-level variation. This paper presents a focused narrative review of five studies from IEEE, Elsevier and Springer venues (2022–2025) that combine low-cost sensors, wireless or cloud transmission and ML-based crop or fertilizer recommendation. The review covers sensing hardware (NPK, pH, moisture, temperature and humidity sensors, and visible-infrared spectroscopy), ML methods (random forest, classifier comparisons, neural networks), and IoT architectures built on Arduino, ESP32, MQTT and cloud platforms. Common weaknesses are small or open (non-field) datasets, limited validation across seasons and regions, recommendations that stop at crop level, and little reporting of sensor calibration or drift. The paper contributes a comparative analysis, a list of research gaps, and research directions toward variety-level recommendation that uses soil data, climate and variety traits.