Beyond self-report surveys: Leveraging multimodal large language models (MLLMs) for farmers market data harvesting from public digital resources

Traditional farmers market research using self-reported surveys has been constrained by high costs, extended timelines, recall bias, and frequently outdated findings. To address these limitations, this study introduced multimodal large language models (MLLMs) as a scalable, cost-efficient approach to extracting farmers market data through automated processing of diverse public digital sources, includ­ing websites, social media, photographs, and gov­ernment documents. This study adopted a two-step framework to extract relevant information and transform unstructured multimodal data into an analysis-ready format. Benchmarked against the Michigan Farmers Market Census and Directory (MIFMA, 2024), our framework covered 76% of their topics. The MLLMs demonstrated robust per­formance, achieving near-zero hallucination rates, 98% accuracy of key variables extractions, and the ability to support real-time updates. While this approach cannot capture confidential or subjective data, it paves the way for a future hybrid frame­work that integrates the comparative advantage of two methods: MLLMs for efficient, factual data collection and human researchers for conducting targeted surveys to capture subjective insights. This efficient, reliable, and scalable approach empow­ered policymakers, market managers, and research­ers to dynamically monitor trends and obtain accu­rate, detailed, and timely data, fostering resilient and inclusive food systems. Beyond farmers mar­kets, the applications of this adaptive framework could extend to other domains, such as public health, urban planning, and economic policy, high­lighting artificial intelligence (AI)’s transformative potential for streamlining data-centric decision-making.