How to Run an AI Search Visibility Audit

List the organization, products, people, locations, and categories the site should represent. Record official names, aliases, relationships, and canonical source pages. Then group real user questions into discovery, comparison, validation, implementation, and troubleshooting tasks.

Choose a small set of high-value prompts for ongoing observation. Include branded and unbranded questions, but avoid treating one synthetic prompt as a market-share measurement.

Request robots.txt, sitemap.xml, llms.txt if present, the homepage, About page, documentation, and representative articles. Inspect status codes, redirects, canonical URLs, meta robots, X-Robots-Tag, Content-Type, WAF challenges, and visible text.

Trace internal links from the main navigation and content hubs. Important source pages should not be orphaned or discoverable only through a form. Check whether essential text survives without a fragile client interaction.

Compare names, descriptions, relationships, and URLs across visible pages and structured data. Confirm that a reader can tell who operates the site, what the product does, who it serves, and how to contact the organization. Resolve outdated brands and competing descriptions.

Map each important question to a canonical page. Look for gaps where the organization has expertise but no public, citable explanation. Evaluate whether the page offers original evidence and is more useful than a generic summary.

Read each page at passage level. Definitions, numbers, steps, and comparisons should retain their subject, scope, and date when copied out of context. Tables need real headers. Images need nearby text. Claims need accessible primary sources or an explained first-party method.

Check author bylines, biographies, publisher identity, publication and modification dates, About and contact pages, and policies. The goal is not decorative trust badges; it is enough public evidence to assess who is responsible for the information.

Run the fixed prompt set under documented conditions. Save the prompt, answer, system and version when available, date, account state, geography, linked sources, brand description, factual errors, and whether the cited page actually supports the claim.

Repeat on a reasonable cadence and look for patterns. Answers are non-deterministic and indexes change, so one missing mention is not automatically a website defect. Repeated factual confusion or citations to weak third-party pages can reveal a missing first-party source.

Fix inaccessible canonical source pages and factual contradictions first. Then address missing topic coverage, weak entity relationships, uncited claims, and unclear passage structure. Experimental additions such as llms.txt belong after these fundamentals.

Assign every action an affected page, evidence, owner, expected result, and verification method. Re-audit the source and observe prompt behavior over time without promising a guaranteed citation outcome.

An AI visibility audit becomes actionable when each symptom maps to an owner and system layer. A missing claim may be editorial; absent server-rendered text may belong to the frontend platform; a stale page may be a CloudFront policy; crawler blocks may be WAF; inconsistent dates may originate in the CMS; and weak citations may reflect the research workflow. Record the observable evidence before assigning the solution.

For AWS-hosted systems, draw the delivery and data path alongside the audit. Use a generative AI RAG architecture when the organization also operates a private answer system, and a production web architecture for public delivery. The two views prevent teams from confusing internal retrieval readiness with public discoverability.

Prioritize fixes by reach, confidence, and reversibility. Restore access and canonical consistency first, then improve cornerstone entity pages, evidence, structure, and contextual links. Re-run the same query set and public fetch checks after release. The Website X-Ray audit supplies a repeatable outside-in snapshot, while referral and citation monitoring show whether visibility changes over time.