What an AI visibility audit should actually tell you

Start with the decision the organisation needs to improve

Define the audiences, markets, products and journeys that matter. An enterprise software company may care about shortlist inclusion and technical trust. A visitor attraction may care about itinerary recommendations. A retailer may need accurate product comparisons and availability information.

These choices determine the prompt sample, competitors, platforms and evidence sources. Without them, the audit becomes an arbitrary collection of questions that cannot support a business decision.

Build a prompt set that represents real journeys

Organise prompts by discovery, comparison, validation and action. Include the broad questions that create a consideration set as well as the specific questions that test important claims or objections. Record why every group exists.

Keep a stable core sample so the baseline can be repeated. Use a separate exploratory set for new products, emerging terminology and questions uncovered during the research. This prevents the audit from improving its own result by quietly changing the test.

• Audience, country and language
• Journey stage and intended decision
• Prompt wording and meaningful variations
• Platform, model or interface tested
• Date, settings and account conditions where relevant

Check technical eligibility before recommending content

Important owned pages need to be accessible, indexable and understandable. Review robots controls, noindex rules, canonicalisation, rendering, internal discovery, status codes, structured information and the availability of important claims in visible text.

Google states that pages must be indexed and eligible to appear with a snippet before they can be considered as supporting links in its AI features. OpenAI states that publishers should allow OAI-SearchBot if they want content to be included in ChatGPT search summaries and snippets. An audit should verify those foundations rather than assume them.

Measure presence, recommendation and representation separately

A brand mention can be incidental. A recommendation places it in a consideration set. A citation connects an answer to an identifiable source. Accurate representation tests whether the claims, products and limitations in the answer match reality.

Record each outcome independently. Add qualitative review where context changes the meaning: being mentioned as an unsuitable option is not equivalent to being recommended, and a linked page may not support the sentence beside it.

Map the sources that shape the answer

Review owned pages and the wider evidence environment: publications, reviews, communities, reference sources, listings, partners and product feeds. Look for repeated claims, missing facts and sources that appear across several relevant answers.

The aim is not to manufacture mentions everywhere. It is to understand where people and systems encounter credible information about the organisation, and whether those sources describe it accurately enough to support a recommendation.

Compare competitors by evidence, not only share of voice

If a competitor is recommended more often, inspect the situations, attributes and sources associated with that advantage. They may have clearer comparison content, stronger review coverage, better product detail or a more established entity across the web.

A good audit distinguishes observable differences from hypotheses. It should show what evidence would confirm the likely cause and avoid prescribing copied content simply because a competitor appeared in one sample.

Turn findings into an owned roadmap

Group actions by the team that can deliver them: technical, content, product, digital PR, brand, analytics or local-market ownership. Describe the evidence, affected journey, expected change and method of validation for each recommendation.

Prioritise eligibility and accuracy before cosmetic gains in a score. The final plan should make clear what can be fixed now, what needs a test and what remains outside the organisation’s control.

Know what the audit cannot prove

An audit is a sample taken under defined conditions. Model outputs change, interfaces vary and not every platform exposes how an answer was assembled. Report those limitations prominently.

The audit becomes valuable when it creates a repeatable baseline. Later measurement can revisit the same journeys and show whether presence, accuracy, citations, referrals or commercial outcomes changed after the work.