Verified Alt-Text for Product Catalogs: A Five-Agent Framework

Product image alt-text is an accessibility contract and retrieval signal, yet large catalogs commonly generate it with one unverified vision--language model (VLM) call. We introduce a five-agent pipeline that grounds catalog claims, verifies them against the image, composes a 125-character description, and records a per-claim audit trace. On a bilingual benchmark of 300 live listings (150 Spanish and 150 English outputs), metadata-only generation asserts the catalog brand in 99.6\% of outputs, whereas image-grounded systems fall to 18--28\% and track whether the brand is visible. A paired analysis (95\% bootstrap intervals; exact McNemar $p<10^{-40}$) confirms this gap. The pipeline returns an auditable description for 255 products (85\%) within two attempts; rejected claims become catalog--image mismatch candidates for review. We also report a 200-trial GEO-style visibility simulation, while reserving independent cross-system faithfulness judging and online testing for future work.