An AI visibility audit examines how a defined set of AI experiences represents your brand for a defined set of questions. Its output should be an evidence register and a prioritized action list. A single screenshot showing a competitor above your company is an observation; it becomes useful audit evidence when the question, surface, date, context, and classification are preserved.
This guide proposes an audit protocol. It is not a provider-endorsed ranking test, and it cannot reveal every answer a prospective customer might see. Use it as the diagnosis stage of an AI search optimization program.
Define the decision and the boundary
Choose an audience and one decision journey. A regional accountancy practice could examine questions about selecting a bookkeeping partner for a small retail business. A software provider could examine evaluation, migration, and implementation questions. Avoid mixing every audience and product into one score before you understand the segments.
Write down the included brands, aliases, locales, languages, surfaces, and time window. Freeze the competitor set for the baseline. Record account conditions and whether search was enabled or visibly used. Preserve consumer-interface and API observations as different groups. Differences between them are part of the study design, not missing data to smooth away.
Select questions from actual customer language where possible. Include unbranded discovery, comparisons, specific use cases, and branded verification. Keep branded questions separate because naming the brand in the prompt changes what the test is asking. Build and version the sample with the AI monitoring prompt-set guide.
Capture the answer before interpreting it
For each scheduled observation, save the exact prompt, complete visible answer, visible citations, timestamp, and collection context. Where your tooling supports it, retain a stable evidence reference. Capture the original before retrying, because a retry can produce a different result and should be identifiable as another attempt.
Use explicit statuses: successful answer, failed request, incomplete capture, or unclassifiable result. A successful answer containing no tracked brands is a valid zero-mention observation. An unavailable response is not. Keep the original scheduled total and the successful total in the report, with a separate count of unknown cases.
Do not treat every source link as proof. OpenAI's ChatGPT search guidance warns that results and citations can be incomplete, outdated, or incorrect and advises checking cited sources. Apply that check to consequential claims before marking an answer accurate.
Code four separate outcomes
First, record brand mentions. Count a brand once per answer even when it appears in the heading, body, and conclusion. Maintain an alias dictionary and flag ambiguous names for review rather than guessing.
Second, record recommendations using a written rule. For example, classify “Brand A is suitable for a small team needing external approval” as a recommendation, while a neutral historical reference is a mention only. Preserve the quoted context so another reviewer can challenge the classification.
Third, record citations. Distinguish a link to the brand's own domain from a third-party page discussing it. The absence of a citation does not mean the brand was absent. A citation also does not automatically mean the linked page supports the associated statement.
Fourth, review factual accuracy for the facts that affect customer decisions. Record the statement, the verified source, the discrepancy, and its practical significance. Separate factual errors from preferences and judgments you disagree with.
Inspect the likely source problem
Group issues into owned-page gaps, technical access problems, identity confusion, and external-source discrepancies. Do not assert causation merely because an outdated source was cited nearby. Describe the evidence: “The cited directory contains the retired plan name” is defensible; “the directory caused every wrong answer” requires evidence the audit probably does not have.
Use an AI citation source audit for deeper source inspection. Escalate indexing or delivery issues to the technical owner. Google specifies that supporting links in AI Overviews and AI Mode must be indexed and eligible for a snippet in Search; see its technical requirements for AI features. That eligibility condition should not be presented as a promise of inclusion.
Build an actionable issue register
The following worksheet contains hypothetical observations, not research findings. Replace the examples with saved evidence from your sample.
| Issue | Evidence | Impact | Proposed action | Owner and acceptance check |
|---|---|---|---|---|
| Retired plan price repeated | Answer A12 and current pricing page disagree | Buyer may budget incorrectly | Clarify current terms and inspect cited old source | Product marketing; verified price and date |
| Brand name confused with another business | Answer A19 describes a different country | Wrong shortlist | Clarify legal identity and official profiles | Brand owner; entity details consistent |
| No answer to migration question | Sample lacks useful migration details; owned docs are vague | Implementation uncertainty | Publish prerequisites and limitations | Documentation; tested workflow |
| Product page cannot be accessed | Delivery test returns an error | Evidence unavailable | Diagnose hosting or access controls | Engineering; verified page response |
Prioritize by customer impact, evidence confidence, and feasibility. Do not use an arbitrary composite visibility score to hide a high-impact factual error. Give each issue one owner and a checkable definition of done.
Report the baseline and its limits
Mention rate is the number of successful sampled answers mentioning the brand divided by successful sampled answers. Share of voice counts each tracked brand once per answer and divides your brand's appearances by all appearances across the fixed competitor set, including your brand. State the denominator and describe this as sample-specific presence, not market share.
Report own-domain citation rate separately from third-party brand-reference citation rate. Show recommendation rules and counts, failed attempts, unknown classifications, and the prompt mix. A change in question mix can change the headline even when the underlying answers do not improve.
Finish with the three most consequential issues, their evidence, and the next collection window. For support turning the audit into a recurring review, book a demo with the issue register and a representative set of prompts.