Can an AI agent discover and use your website?
Agentic Search Optimization asks whether a website gives an automated system enough information to discover a supported task and understand how to approach it. In Seryvon, the audit supplies readiness evidence. A separate interaction test is needed to establish whether a particular agent actually completes that task.
What is evaluated
- Machine-readable structure and identity.
- Discoverable actions and forms.
- Clear availability and constraints, reviewed against the actual interface.
- Navigation and data that an agent can interpret, with safety requiring separate validation.
The ASO rule guide describes the implemented signals in detail. The practical objective is to make an existing task understandable, not to invent an interface for a score. A public documentation search may need clear labels and a discoverable entry point; a reservation journey also needs explicit conditions and a trustworthy confirmation step.
Example: review a reservation journey
This is an illustrative workflow for a website that already supports reservations, not a claim about a Seryvon booking feature. Begin with the page where a visitor selects an option. Identify what is available, what information is required and which action submits the request. Record any authentication or human confirmation required by the real service.
Next, inspect the audit evidence. A detected ReserveAction indicates declared intent. A form classified as usable indicates that the extracted form signals met the static rule. Neither proves that the selected option remains available or that the request succeeds. Compare the declaration with the real interface before treating the result as actionable.
Finally, test the supported journey in a controlled environment with appropriate authorization. Check empty fields, invalid inputs, unavailable options and the final confirmation. Record the actual outcome. A task that stops at authentication is different from an interface with missing labels, and the remediation should reflect that difference.
Keep three kinds of evidence separate
| Evidence | What it supports | What remains unknown |
|---|---|---|
| Declared action | The page describes an action. | Whether the action works. |
| Static form signals | The parser detects required structural clues. | Whether validation and submission succeed. |
| Recorded interaction | A tested agent completed or failed a specific journey. | Whether other agents or conditions behave the same way. |
What ASO does not promise
An ASO score does not guarantee rankings or recommendations by any specific AI system. Experimental WebMCP, action-schema and discovery checks are defined in the current ASO implementation. Missing external probe data should remain not measured, and detected API documentation links should not be described as tested API connectivity.
For content-only tasks, first review crawl accessibility and answer structure. For before-and-after audits, keep scope and configuration comparable so a provider change does not masquerade as a usability improvement.
Seryvon is a public open-core release published by Powehi. The managed product is coming soon. The product audits readiness signals; these pages do not claim a universal agent-completion benchmark.