How to Run an AI Visibility Audit

Specify the audit before running it

An audit should state which business question it is testing, which answer interfaces are included and which buyer prompts represent the scope. Include brand discovery questions and comparison questions, but do not mix in navigational searches when the goal is to understand non-branded discovery.

For each run, preserve provider, interface or mode, prompt text, language, location when known, timestamp, raw answer, cited URLs, competitor mentions and description accuracy. Without those fields, a score cannot be independently reviewed and a future result cannot be compared fairly.

• Report mention separately from a clickable citation.
• Check whether the cited URL is relevant and current.
• Flag unsupported claims and outdated product descriptions.
• Retain negative results as carefully as positive ones.

Turn findings into a testable action

Do not issue a generic recommendation such as publish more content. Map the actual citation or accuracy gap to a page: a comparison gap may need a fair alternatives section, an unclear category may need a stronger opening, and a factual error may need a precise source or a public correction. The audit should state why that change addresses the observed answer.

After editing, rerun the same prompt and keep the prior answer. Review several runs before claiming the brand has gained visibility. Bing Webmaster Tools also now offers aggregated AI citation insights for supported Microsoft experiences; its own documentation warns that aggregate metrics do not establish page rank or causation.

• One finding, one page, one proposed change.
• Name a review date and the expected observable result.
• Keep an explicit 'no action' reason when evidence is insufficient.

Read the report as a decision record

A useful report preserves the negative finding, its supporting answer and why it matters to the business. It should distinguish a direct observation from an interpretation, then make a proportionate recommendation. If the evidence is mixed, recommend a bounded follow-up test rather than a confident content rewrite.

Close the loop after the review date. Record whether the action was completed, what the new answer shows, and whether the original gap remains. This produces an audit history that a founder can learn from and lets someone else inspect how a conclusion was reached.

• Attach evidence to every material finding.
• State confidence and unresolved questions.
• Close actions with a before-and-after record.

What an AI visibility audit checks

Start with four answer engines: ChatGPT, Gemini, Perplexity and Google AI Mode or AI Overviews. Run the same buyer prompts in each one, then record whether your brand appears, which competitors appear, and which URLs are cited.

Do not only test your brand name. Test category prompts, competitor alternatives, problem prompts and buyer questions. Those are the prompts that show whether the market can discover you before it already knows you.

• Mention: does the answer name your brand?
• Citation: does the answer cite your URL?
• Accuracy: does it describe your category and offer correctly?
• Competitors: who is recommended instead?
• Source gap: which page should exist or be improved?

Use Search Console to choose prompts

Google Search Console will not tell you whether ChatGPT mentions your brand. But it does show the language people use before they ask AI systems the same question. Start with queries that already have impressions: brand mentions in Gemini, track ChatGPT mentions, AI visibility audit, Perplexity brand tracking and AI SEO tools for small business.

Each query becomes a prompt to test and a page to improve. If the query has impressions but the AI answer ignores you, the next action is not another dashboard. It is a clearer source page.

Turn findings into fixes

A useful audit ends with page-level fixes: add a direct answer at the top, clarify the category, add a comparison table, publish a FAQ block, add Organization or SoftwareApplication schema, and link from related pages.

Re-run the same prompts after the pages are indexed. The metric that matters is not one lucky mention. It is whether the same buyer prompts start producing more accurate answers over time.