An AI visibility audit for B2B websites: useful checks, realistic limits

An AI visibility audit should answer a practical question: can people and relevant discovery systems understand what the business does and find reliable evidence for its claims? It should not begin with the assumption that every company needs a special file or a new platform to appear in an AI-generated answer.

For a B2B website, the most useful work often overlaps with established technical SEO and content quality. Clear services, accessible pages, accurate company information and credible project evidence remain important. Integrations with assistants are a separate opportunity when there is a defined task to support.

Start with what can actually be accessed

Check that important pages return the expected status, can be reached through normal navigation and contain the information visitors need. Review canonical URLs, indexing directives and any accidental restrictions that prevent relevant crawlers from accessing public content.

JavaScript rendering deserves attention when essential content depends on it, but a headless architecture is not automatically required. The audit should identify the actual delivery problem and the simplest reliable fix, rather than prescribing a fashionable stack.

Private information belongs behind proper access controls. Making a page easier for discovery systems to read must never mean exposing internal documents or customer records.

Make the business understandable in the visible content

A visitor should be able to identify the service, intended customer, delivery process and next step without decoding vague slogans. Use headings that describe the subject and explain specialist terms where they matter.

For B2B services, useful detail may include scope boundaries, integration requirements, responsibilities and examples of the problems the team solves. A case study should distinguish the company's contribution from the work of other parties and support any measured results.

Concise summaries and well-structured answers can help readers navigate longer material. They should reflect the actual page rather than introduce claims that the rest of the content cannot substantiate.

Use structured data to describe visible facts

Structured data can help supported search features interpret entities and content. It needs to match the information on the page and follow the relevant rules. Correct syntax alone does not make a claim true or guarantee a rich result.

It also does not guarantee that an AI service will quote the page. Google's guidance states that there are no additional schema requirements for its AI features. Treat markup as a useful descriptive layer, not a mechanism for forcing citations.

Review organisation details, breadcrumbs and other appropriate types for accuracy and consistency. Avoid invented ratings, unsupported awards or hidden service claims inserted only into JSON-LD.

Separate public discovery from an assistant integration

A public page being found by a search system is different from an assistant retrieving data through an API or a retrieval-augmented generation system. The latter needs a defined data source, access policy and operational purpose.

If a sales assistant needs approved product information, the audit can examine how that information is selected, updated and attributed. If it can perform actions, the review must also cover authentication, authorisation, input validation and user confirmation where appropriate.

An API or MCP server is useful only when there is a consumer and a task that justify it. Publishing one does not automatically improve public search visibility.

Review crawler instructions without confusing them with security

robots.txt can communicate crawling preferences to supporting bots. Providers may distinguish search crawling from other uses, so decisions should be based on their documented user agents and purposes.

The llms.txt project is an optional proposal, not a universal requirement. An agents.json file may describe a custom integration, but in our context it is a specifically agreed format rather than an established requirement of Google or OpenAI.

None of these files guarantees that a model has current knowledge, prevents every incorrect answer or replaces access controls. The audit should record what is known to consume each file and why maintaining it is worthwhile.

Ten checks that make an audit actionable

• Verify page status, canonical URLs and intended indexability.
• Check that important content is reachable and readable.
• Review mobile usability and measured performance.
• Protect private data and administrative functions.
• Use meaningful headings and clear service descriptions.
• Examine APIs only where a real integration needs them.
• Add context that helps readers interpret technical claims.
• State scope, conditions and terminology precisely.
• Support claims with genuine evidence and attributable sources.
• Label optional manifests honestly and keep them maintained.

Each finding should include the affected page or component, its practical consequence and a testable correction. A list of fashionable terms is not an implementation plan.

Measure with an understanding of the limits

AI-generated answers can vary by system, date, location, prompt and session context. A documented sample of relevant questions can show what appeared during that test, but it is not a stable ranking position for every user.

Record the prompts, dates and systems used. Combine that observation with measurable business signals such as qualified referrals, useful engagement and enquiries. Where analytics cannot reliably identify an AI source, do not label unknown traffic as proof of success or failure.

Search Console can help assess search performance, but its reporting does not provide a universal view of citations across all assistants. The audit should be explicit about what the available data can and cannot show.

Prioritise useful fixes over speculative additions

A broken canonical, an unclear service page or a missing project explanation can be a more concrete problem than the absence of an optional AI file. Address the issues that affect real users and supported discovery mechanisms first.