Retrieval, source assessment and answer generation are useful lenses for investigating AI visibility. They are an explanatory model, not a published sequence shared by every product. Search-enabled assistants can retrieve web sources, while answers without search may rely on model knowledge. The exact selection process is not fully observable. We inspect access, the captured answer and its links, then distinguish what those observations establish from what remains a hypothesis.
Check one: access to relevant information
For web-supported answers, a product must be able to discover and use relevant sources. Search crawlers, training crawlers and user-triggered fetchers have different roles. We check the controls relevant to the product being tested, rather than assume that allowing a training crawler establishes search access. An answer can also describe a business through third-party sources even when its own website is not cited.
One issue to investigate is an access restriction. robots.txt and a CDN or firewall can impose different controls. An HTTP error on a diagnostic request is evidence about that request; it does not prove that the real crawler was blocked or that the entire product cannot mention the business. We seek corroboration from verified crawler traffic or the platform’s own inspection tools where available.
A second issue is content accessibility. Important information may be difficult to obtain if it appears only in images or depends on unsupported interactions. We inspect the available page content and check that relevant structured data agrees with it. Structured data can describe content; its absence alone does not make a page unreadable or ineligible for Google’s AI features.
Check two: support for the published claims
Source quality is another area to investigate. We compare claims with supporting sources, check that identities and dates are accurate, and distinguish independent evidence from a business’s own assertions. These are useful editorial checks. They do not reveal a product’s hidden weighting or establish why a particular source was selected.
Our review considers consistency of material business facts, corroboration from relevant independent sources, and authorship that lets readers understand who is responsible for the information. The appropriate evidence depends on the question and sector. We record what we can verify and avoid assigning unsupported weights to these factors.
Check three: the answer and its source links
Clear writing helps readers understand an answer: explain the main point early, support material facts, use useful comparisons and show genuine review dates. When testing an assistant, we inspect the answer it produced and which sources it actually linked. We cannot infer that wording alone caused a citation, or that a clear page must be selected over another source.
Why do answers and recommendations behave differently?
An explanatory question and a request to recommend a provider ask for different outputs, so both belong in a well-scoped test. We record whether a business is mentioned, whether it is recommended, which links are shown and whether the information is accurate. A recommendation without an owned-site link is a different result from a source citation. No retained evidence establishes a universal rule that schema causes citations while independent authority causes recommendations.
What should you investigate next?
Three practical investigations
- Absent from the tested answers, including brand questions? Check the exact name and intended business first. Inspect relevant indexing and access evidence, then repeat under recorded conditions. Absence alone does not identify the cause.
- Named or cited, but described inaccurately? Compare the answer with current facts and the sources actually shown. Record the error and any conflicting public information; a source link does not establish what version the product used or why.
- Present for some questions but absent for a priority service? Compare the questions, repeated answers and cited pages. Identify useful information the service page lacks, without treating a content difference as proof of the cause.
These checks narrow the investigation; they do not reveal every selection decision. An evidence-led audit identifies observed issues, unresolved questions and proportionate next steps. The Answari methodology explains how the evidence is recorded and compared.
Apply these checks to a law firm
Our AI visibility service for solicitors applies the method to practice areas, fee pages and professional profiles. The conveyancing page review shows how published information can answer a buyer’s next question, using real pages and a downloadable record.