How agent-ready is OpenAI?

OpenAI's public website scored 40/100 on agentability for its homepage, with lower scores on pricing (12/100) and documentation (32/100). These scores indicate that autonomous AI agents will face moderate to significant friction when attempting to parse information, navigate flows, or programmatically interact with the site.

The audit identified three principles as critical weaknesses: transparency (0/100), shadow UI avoidance (15/100), and defaults (33/100). Machine readability (69) and status signaling (65) performed relatively well, while chunking (41), control (48), and clean handoffs (50) fell in the middle range. The transparency score of zero reflects a complete absence of confidence indicators, source citations, audit logs, and explanatory affordances that agents require to assess reliability and trace decisions.

These gaps mean that while agents can technically read page content, they lack the metadata, control signals, and structured outputs needed to operate confidently or recover gracefully from errors. Improvements in the lowest-scoring principles would yield the most significant gains in agent-readiness.

The following improvements, drawn directly from the audit findings, would address OpenAI's weakest agentability principles:

• Add structured data markup (JSON-LD using schema.org vocabularies) to describe primary entities on each page, improving machine readability and enabling agents to identify key resources programmatically.
• Include confidence or certainty indicators on API responses and UI outputs, such as a 'verified vs best-guess' label or a confidence field, so agents can assess result reliability.
• Expose an activity log or 'why this result' feature that records system actions in machine-readable format, providing both a visible feed and a JSON event log for auditability.
• Attach source references or input citations to generated outputs, using links or identifiers that allow agents to trace provenance and verify information.
• Pre-fill form inputs with sensible defaults wherever reasonable values exist, reducing the number of decisions agents must make during automated flows.
• Provide clearly labeled pause, cancel, or stop controls for long-running operations, using accessible names like 'Cancel import' rather than unlabeled icons.