A Failure Taxonomy for Machine-Mediated Shopping

Seven classes cover every failure we have logged so far. The taxonomy is versioned and open.

Purpose

A shared vocabulary for failure types allows stores, agents, and researchers to communicate precisely about what went wrong. We publish this taxonomy so external teams can apply it to their own traces.

• F1 - Unresolvable Element: The agent could not locate or interact with a required page element (e.g., a price rendered as an image, a button requiring hover to appear).
• F2 - Variant Ambiguity: The agent could not determine the correct variant from available signals (e.g., size/color selectors that do not expose machine-readable state).
• F3 - Cart State Loss: Items were added to the cart but were not present at checkout due to session expiry or redirect behavior.
• F4 - Payment Wall: The agent reached the payment step but could not proceed because the step required a human-scale interaction (e.g., CAPTCHA, 3D Secure modal with no programmatic path).
• F5 - Dynamic Content Race: A price, availability status, or form field changed between the agent reading it and acting on it, producing an inconsistent state.
• F6 - Navigation Dead End: The agent reached a page or state from which no further progress toward task completion was possible and no error was surfaced.
• F7 - Data Inconsistency: Structured data (feed, schema.org, meta tags) disagreed with rendered page content, causing the agent to act on incorrect information.

Usage

Each failure logged in our benchmark traces is tagged with one or more taxonomy codes. Codes appear in published scorecards and in the raw trace dataset. Stores can use these codes to prioritize remediation. The taxonomy will be extended as new failure modes are observed; additions will be announced as minor version increments.