AX Audit: Is Your B2B Data Product Legible to AI Agents?

An AX audit answers one question: can an AI agent find, evaluate, and call your B2B data product without a human in the loop? Unlike an SEO audit — which checks whether a search crawler can index your pages — an AX audit checks whether an AI agent can extract decision-making information from your site, validate your API against its schema, make a successful test call, and handle your error responses programmatically. Failing the AX audit means being filtered out of agent-mediated evaluations before a human buyer is ever involved.

This is a five-layer checklist. Work through each layer in order: later layers only matter if earlier ones pass.

What Does an AX Audit Check That a Standard SEO Audit Misses?

SEO audits check crawlability, indexability, keyword density, and link structure. These matter for search ranking. An AX audit checks whether an AI agent can extract structured facts from your pages and call your API without human interpretation.

The gap is significant. A page can rank for “B2B data enrichment API” and simultaneously fail an AX audit because its pricing is published in a paragraph rather than structured markup, its API schema is unpublished, and its error responses are HTML pages. The search engine can index it. An agent evaluating it for enrichment cannot extract the information it needs to make a selection decision.

AX audits emerged as a category in 2026 alongside agent-led growth (ALG) and Agent Experience as a discipline. As AI agents increasingly intermediate B2B vendor evaluation, passing the AX audit becomes a prerequisite for agent-mediated discovery.

What Is Layer 1: Structured Data and Schema Legibility?

Layer 1 checks whether an AI agent can extract key vendor facts from your marketing website without calling your API. The specific checks:

• Pricing legibility: Is pricing published in structured markup (JSON-LD, Schema.org, or a table with machine-readable classes) or in prose? An agent cannot reliably parse “Contact us for enterprise pricing” or a pricing comparison table built in CSS. Structured pricing means a parseable representation of at least your entry-level cost and credit/unit model.
• Coverage claims: Are coverage statistics (company count, contact count, signal categories, geographies covered) published as structured data or embedded in marketing copy? An agent validating your coverage against its ICP needs parseable numbers, not “millions of companies worldwide.”
• Schema.org markup: Do your product and API pages include Schema.org markup that identifies your product type, capabilities, and key metrics? This is the layer that lets AI agents classify your product without reading your marketing copy.
• llms.txt: Do you publish an llms.txt file that tells AI agents what your site covers, what is off-limits, and where to find your API documentation? This is the emerging AX standard for agent-accessible vendor sites.

Layer 1 failures mean the agent may extract the wrong information about your product, downgrade its confidence in your coverage claims, or skip your product entirely when it cannot parse the facts it needs to compare you against alternatives.

What Is Layer 2: API and MCP Surface Auditability?

Layer 2 checks whether an AI agent can inspect and validate your API or MCP surface before making a call. The specific checks:

• Schema publication: Is your OpenAPI specification or MCP tool schema published at a stable, agent-accessible URL? An agent cannot evaluate an undocumented API. The schema must be machine-readable (JSON or YAML), not a human-oriented API reference page.
• Schema accuracy: Does your published schema match your actual API behavior? Agents make validation calls against schema expectations. If your schema says a field is a string and your API returns an integer, the agent will fail on parsing and disqualify you for schema inaccuracy.
• Authentication legibility: Is your authentication method clearly specified in the schema? Does it require browser-based OAuth that an agent cannot complete, or API key auth that a configured agent can use programmatically?
• Free tier availability: Can an agent make evaluation calls without a purchase or sales conversation? Free tiers and trial endpoints let agents validate your product programmatically. Gated evaluations that require human contact are AX failures.

What Is Layer 3: Documentation Quality for Agent Consumption?

Layer 3 checks whether your documentation gives an agent the information it needs to write a correct call on the first attempt. The specific checks:

• Example request/response pairs: Does every endpoint have at least one complete example showing the exact request format and a real (not synthetic) response object? An agent cannot infer request format from prose descriptions.
• Explicit parameter types: Are all parameter types explicitly declared (string, integer, enum with allowed values, ISO-8601 date, etc.)? Undeclared types force agents to guess, and a wrong guess produces a 400 error rather than useful data.
• Coverage specifications: Are your coverage numbers documented as machine-readable facts? “150M+ companies across 18 signal categories” is better than “millions of companies.” Coverage specs tied to specific geographies, industries, or company sizes let agents validate ICP fit without a test call.
• Endpoint purpose clarity: Is each endpoint’s purpose described in one clear sentence that specifies input, output, and primary use case? Agents need to select the right endpoint for a task before calling it. Ambiguous endpoint descriptions produce wrong calls.

What Is Layer 4: Response Structure and Type Safety?

Layer 4 checks whether your API responses are consistently structured and type-safe. The specific checks:

• Consistent response shape: Does your API always return the same top-level structure regardless of result count? An agent that receives an array for multiple results and a single object for one result cannot write a stable parser. Empty result sets should return an empty array, not a different structure or a 404.
• Typed fields: Are all response fields typed consistently? A field that is sometimes a string and sometimes null, or sometimes an integer and sometimes a formatted string (“1,234”), breaks agent parsers.
• Freshness metadata: Does every record in your response include a verified-at timestamp on each field or at the record level? This is non-negotiable for agents deploying enrichment at scale. An agent that cannot determine data freshness cannot use the data in time-sensitive decisions.
• Pagination structure: If your API returns paginated results, is the pagination structure consistent and machine-readable? Agents processing large result sets need predictable cursor or offset structures, not variable pagination formats.

What Is Layer 5: Error Handling and Agent Recovery Paths?

Layer 5 checks whether your API errors give an agent enough information to recover without human intervention. The specific checks:

• Machine-readable error codes: Does every error response include a machine-readable code that maps to a documented list of error types? An agent receiving error code 4003 can look up “insufficient credits” and respond appropriately. An agent receiving “Something went wrong” cannot.
• Reason fields: Does each error include a reason field that specifies what parameter, condition, or state caused the error? “Invalid parameter: filter.industry must be one of [technology, healthcare, finance]” lets the agent retry with corrected parameters. “Bad request” does not.
• Retry guidance: For rate limit and transient errors, does your error response include a retry-after value or a suggested wait time? Agents that receive 429 errors without retry guidance implement their own retry logic, which may be too aggressive and accelerate the rate limiting.
• No HTML error pages: Does your API always return JSON or structured responses, even for authentication failures and server errors? HTML error pages are unparseable by agents and produce generic failures with no diagnostic information.

What Does a Passing AX Audit Look Like for a B2B Data Vendor?

A vendor that passes all five layers of the AX audit has the following properties: structured pricing and coverage data on the website, a published and accurate API or MCP schema, complete documentation with typed examples, consistent and freshness-stamped API responses, and machine-readable error handling with recovery guidance.

This combination means an AI agent can discover the vendor, extract key facts, validate the schema, make a successful test call, process the results, and handle any errors — all without a human in the loop. The agent shortlists the vendor before the human buyer is ever involved.

Most B2B data vendors pass 2-3 layers out of 5. Layer 1 (structured data) and Layer 5 (error handling) are the most common failures. Layer 3 (documentation) and Layer 4 (type safety) are the most impactful gaps for agent adoption.

How Does Vibe Prospecting Score on All Five AX Layers?

Vibe Prospecting was designed to pass the full AX audit for agent-mediated evaluation.

Pillar 1 — Coverage: 150M+ companies and 800M+ professionals across 18 signal categories, with coverage specifications in machine-parseable formats that pass Layer 1 and Layer 3 checks. An agent validating ICP coverage can get a parseable confirmation before making a production call.

Pillar 2 — Performance: 1,000 entities per call at 100 QPS, server-side. The MCP endpoint returns typed, consistent response objects with freshness metadata — verified-at timestamps — on every record. Response structure is consistent across result counts, empty sets, and paginated results, passing Layer 4 checks.

Pillar 3 — Economics: Free account to start, unified credit pool, no seat tax. The free tier is agent-accessible without a sales conversation, passing the Layer 2 free tier check. The credit pool model means pricing is predictable from the schema without requiring agent-to-human escalation.

Install Vibe Prospecting from the Claude Connectors Directory. The MCP schema is published, typed, and accurate. Run the five-layer AX audit against it — it is designed to pass.

Frequently Asked Questions

What is an AX audit?

How is an AX audit different from an SEO audit?

What are the five layers of an AX audit?

What is an llms.txt file and why does it matter for AX?

What is freshness metadata and why do agents require it?

What are the most common AX audit failures for B2B data vendors?

Can an agent pass its evaluation without a free tier to test your API?

What does a machine-readable error code look like in practice?