Translating Blind and Low-Vision Users' Feedback into Domain-Specific Criteria for AI-Generated Alt Text: Evidence from E-Commerce

Abstract

Alternative text (alt text) is increasingly produced by generative models, and the guidance steering that generation comes largely from accessibility standards written to hold across domains. Prior work has established that blind and low-vision users (BLVUs) have information needs that vary considerably by domain, yet the step of turning those documented needs into generation-time guidance for such models has rarely been examined. We examine this translation step in e-commerce. In Study 1 (N=15), adding Section 508 guidance to a context-aware generation baseline did not raise perceived alt text quality for six product images, and shopping relevance trended lower. Thematic analysis surfaced three requirements that such domain-general guidance does not address: alignment with product-category knowledge, prioritization of purchase-relevant information, and sentence structures suited to auditory processing. We composed established methods into a four-stage, semi-automated workflow that converts these requirements into prompt-level generation criteria while retaining WCAG 2.1 as guardrails. In Study 2, an independent group of BLVUs (N=11) rated criteria-based alt text higher on the three targeted dimensions and preferred it in 44 of 66 forced choices, with no significant decrease on retained general-quality dimensions, though the positively keyed instrument and reused stimuli make these ratings exploratory. We contribute a documented procedure for turning situated BLVU feedback into generation-time criteria in a domain-specific setting.