Abstract: [Objective] Mineral exploration in deep, concealed and covered areas is increasingly confronted with weak, indirect and multi-interpretable ore-forming signals. Under such conditions, the key issue in mineral prospectivity mapping is no longer whether more heterogeneous geoscience data can be collected, but whether effective metallogenic information can be extracted and identified from geological, geochemical, geophysical, remote sensing and three-dimensional spatial data. Metallogenic information is the critical intermediate link among metallogenic theory, deposit models and quantitative prediction. It can be metaphorically regarded as the “information food” of mineral prediction, because it determines whether geological understanding can be transformed into meaningful model inputs. [Methods] This paper reviews recent progress in the intelligent extraction and identification of metallogenic information following the logic of “theory and model—information classification—information extraction—intelligent identification—geological verification”. On this basis, the paper discusses model-guided extraction of geological-structural, geochemical, geophysical, remote sensing and three-dimensional metallogenic information, and summarizes GIS-based spatial quantification, statistical learning, machine learning, deep learning, graph learning and knowledge-data collaborative identification methods. Special attention is paid to the relationship among metallogenic models, deposit models and prospecting models. Metallogenic models explain ore-forming processes and genetic mechanisms, deposit models summarize diagnostic feature associations of specific deposit types, and prospecting models translate these understandings into observable, extractable and verifiable prediction criteria. [Results] Metallogenic information should be distinguished from raw geoscience data, data anomalies, prospecting indicators, prediction variables and algorithm-derived model features. Raw data are observations, anomalies are statistical differentiations, and model features are numerical representations learned or transformed by algorithms; metallogenic information, however, refers to computable evidence that is constrained by target deposit types and metallogenic models and that can indicate ore-forming processes, controlling factors, anomaly responses and spatial associations. Effective metallogenic information should have genetic relevance, deposit-type specificity, spatial expressibility, scale compatibility, computability, interpretability and verifiability. According to mineral system elements and deposit models, metallogenic information can be classified into information related to material sources, migration pathways, ore-forming spaces, precipitation and enrichment processes, preservation conditions and known mineralization indicators. Geological-structural information should therefore be extracted from favorable strata, ore-related intrusions, faults, contact zones, alteration-mineralization features and preservation conditions rather than from undifferentiated geological layers. Geochemical information should emphasize element associations, zoning patterns, background correction, robust anomaly recognition and three-dimensional grade or mineralization voxel expression. Single-element highs are only candidate anomalies unless they are consistent with the element association, spatial zoning and geological setting of the target deposit model. Geophysical and remote sensing anomalies become metallogenic information only when they can be interpreted, under the constraint of a metallogenic model, as responses of concealed geological bodies, ore-controlling structures, alteration zones or mineralization-related interfaces. Three-dimensional geological models are carriers for metallogenic information extraction rather than prediction results themselves. Their value lies in transforming faults, folds, intrusions, contact surfaces, ore bodies, physical-property fields and drilling constraints into distance fields, structural attributes, voxel properties and three-dimensional spatial relationships. Intelligent methods can improve the recognition of nonlinear combinations, complex spatial patterns and multi-source associations, but they cannot replace deposit models or geological interpretation. Automatically extracted features must be translated back into geological semantics. Only information that satisfies statistical relevance, geological plausibility, spatial continuity and verifiability can be regarded as effective metallogenic information for prediction. [Conclusions] The reliability of mineral prospectivity mapping depends first on the quality of metallogenic information and only secondarily on the complexity of prediction algorithms. Model-guided classified extraction is the foundation for transforming geological understanding into computable evidence. Artificial intelligence should be used to support information extraction, pattern identification and relationship modeling, while its outputs must be constrained and verified by metallogenic models, deposit models, prospecting models and geological facts. [Significance] Future research should focus on model-guided multi-source classified extraction, knowledge-data collaborative identification, multi-scale and three-dimensional information recognition, uncertainty assessment and interpretable validation. These directions can provide more reliable essential information input for deep mineral prospectivity mapping and improve the geological credibility of intelligent prediction results.