Measuring AI-Agent Accessibility of Web Applications: A Construct-Development Framework

Abstract

Agents built on large language models (LLMs) now browse, fill in forms and complete transactions on websites that were designed for people looking at a screen. Existing standards do not describe what such an agent needs from a page. WCAG 2.2 sets requirements for human users of assistive technology, and interoperability standards assume two systems that already share a structured contract. Neither says when a web interface can be reliably perceived, acted on and recovered from by an autonomous agent. This paper proposes the Agent Accessibility Measurement Framework (AAMF) to fill that gap. AAMF divides the construct into Core In-Page Accessibility, made up of Structural Interpretability (SI), Semantic Richness (SR), Action Executability (AE) and Resilience and Error Recoverability (RR), and a separate Machine Discoverability and Protocol Support (MDP) pillar. We define 25 indicators with explicit formulas, an equal-weight baseline score, a failure-criticality-weighted composite score, and a five-layer Agent Interaction Stack (AIS) that keeps in-page interaction apart from out-of-band discovery. We place the construct within ISO/IEC 25010:2023, work through one hypothetical scoring example, and derive 25 engineering guidelines from the indicator definitions. The paper specifies the framework only. It reports no audit of production websites and no agent task study; Section VII describes the study that is needed to validate it.