<p>Multimodal Recommender Systems (MMRS) improve recommendation quality by incorporating content modalities, such as text, images, audio, and video. This survey synthesizes over 127 recent publications (2016–2025) and organizes the MMRS landscape through a novel pipeline-based taxonomy comprising six stages: modality extraction, filtration, bridging, fusion, augmentation, and optimization. Each stage is analyzed in terms of its design choices, representative methods, and practical trade-offs. Different from prior surveys that focus on model architectures or application domains, we emphasize the intermediate processing stages, i.e., filtration, bridging, and augmentation, that are critical for robustness yet under-analyzed in the literature. We further examine how generative models, large language models (LLMs), and multimodal knowledge graphs are reshaping key stages of the MMRS pipeline, and discuss open challenges including cross-domain adaptation, personalized modality preference modeling, and LLM integration. By linking design choices to evaluation practices and identifying concrete research gaps, this survey provides a structured technical roadmap for advancing MMRS.</p>