How to audit AI visibility without mistaking signals for rankings

AI visibility is not one metric. For a given prompt, a system may name a company, link to one of its pages, quote it, mention a product category, or omit it entirely. Those are different observations with different meanings. A named brand is not automatically a recommendation, and a citation is not proof that the cited page determined the answer. Start by documenting precisely which of these events your audit will count.

An AEO audit should therefore preserve context: the prompt wording, market and language, model or interface, date, answer text, cited domains and the analyst's classification. This makes later comparisons possible. It also keeps a dashboard honest when a model changes its retrieval behaviour or an answer varies between runs. The goal is a reviewable record, rather than a single score presented as a search ranking.

• Mention rate: the share of selected prompts that name the brand.
• Citation rate: the share that display a link or named source from the domain.
• Recommendation context: whether the answer frames the mention as an option, example or warning.
• Answer coverage: topics where the brand is absent, ambiguous or incorrectly described.

Useful prompts resemble the questions a prospective customer could ask while researching a problem. Include broad category questions, comparison questions, local or regulatory questions where relevant, and task-focused questions. Do not only test your own brand name: branded prompts mainly show whether a model recognises an existing entity, while unbranded prompts expose whether it associates the brand with a need.

Keep the sample intentionally small enough to review. Group prompts by intent and record a stable version of the set. A monthly run can reveal a change worth investigating, but it cannot establish a causal trend from a handful of answers. If languages, countries or audiences matter, separate them; combining German and English results hides important differences in wording, available sources and model behaviour.

• Discovery: “How do I choose an SEO audit process for an agency?”
• Evaluation: “What should a white-label audit report include?”
• Alternative: “Which tools help track visibility in AI answers?”
• Verification: questions about a known feature, policy or technical requirement.

A mention without a link can be useful evidence of recall, but it gives no clear route for a reader to verify the claim. A displayed citation is more concrete, yet citations can point to third-party pages and can change after a model update. Capture the source URL, the surrounding sentence and whether the page actually supports the statement. Treat an answer as editorial output that needs fact checking, not as analytics data that explains itself.

Sentiment deserves the same restraint. Label wording as positive, neutral, mixed or negative only when the language supports that classification, and keep an “unclear” option. A model may list a brand next to a limitation without making an overall judgement. Reviewing samples by a person reduces false certainty and helps separate factual correction work from ordinary variation in a generative response.

• Save the complete answer and visible source links.
• Flag outdated claims and route them to the responsible content owner.
• Compare competitors as answer presence, not as a definitive market-share metric.
• Note answer refusals, missing citations and unstable outputs.

Accessible, well-maintained pages make it easier for crawlers and people to retrieve and understand information, but no technical file guarantees inclusion in an AI answer. Review indexability, HTTP status, canonical signals, internal links and the readable page content first. Structured data can help machines interpret eligible page properties when it follows the documented vocabulary and matches visible content; it is not a switch for AI citations.

Crawler directives are another operational check. Robots.txt describes crawl access for cooperating user agents, while a server, CDN or login wall can still block a fetch. Keep records of what your own testing shows, and avoid assuming that every model, product surface or third-party retriever uses the same crawler. A change in access may be worth fixing for its normal web benefits even when AI-answer effects remain unmeasured.

• Test important URLs for status, canonical consistency and renderable content.
• Validate structured data against the relevant schema and search-engine guidance.
• Review robots.txt and server responses for documented user agents.
• Avoid blocking assets needed to render useful public information.

llms.txt is a proposed convention for presenting a concise, machine-readable guide to a site. It can be a practical editorial index: point to canonical documentation, explain terminology and keep important URLs easy to find. It is not a web standard with universal adoption, and publishing it does not require any AI system to crawl, retrieve, quote or recommend a site.

If you publish one, keep it aligned with the public site and its governance. Include stable, useful pages rather than marketing slogans, do not expose private material, and assign someone to update it when URLs or policies change. The audit question is modest: can a visitor or compatible tool retrieve an accurate guide? It is not whether the file “unlocks” visibility in a particular assistant.

• Link only to public canonical resources.
• State page purpose plainly and avoid unsupported superlatives.
• Version and review the file with documentation changes.
• Keep robots directives and access controls as separate decisions.

Begin with a baseline run and annotate unusual conditions: model version where shown, country, language, date and prompt version. Then review the answers for factual errors, missing source pages and confusing positioning. Prioritise improvements that are useful regardless of an AI interface, such as clearer expert documentation, accurate product pages and technically accessible public resources. This avoids spending effort on tactics that cannot be verified.

AuditLumo can track selected prompts across ChatGPT, Claude, Perplexity and Gemini alongside an SEO audit, but the output should start a conversation rather than end one. Pair the observations with Search Console and analytics data for conventional search performance, and retain the raw answer evidence. A scheduled monthly or weekly check is most helpful when the team has named owners and a way to distinguish a content fix from normal response volatility.

• Set a review cadence that matches how often pages and messaging change.
• Assign ownership for factual corrections, technical issues and editorial gaps.
• Keep a change log for prompts, pages and classifications.
• Report uncertainty alongside counts so stakeholders can interpret movement responsibly.