How to Run an AI Search Visibility Audit

An AI search visibility audit should end with evidence someone else can inspect: the question asked, the answer returned, the sources cited, and the next decision. A screenshot of your company name is a useful observation. It is not a measurement plan.

The method below is a proposed working process for a small team. The question counts and review cadence are practical starting choices, not provider requirements or validated benchmarks. Adjust them to your market and document the change.

Decide what the audit needs to answer

Write one decision at the top of the audit. For example: “Which of our service pages needs clearer evidence before we commission another article?” This makes the research useful even when your brand never appears.

Define the business, service, language, and market. Pick the answer surfaces you actually want to inspect. Treat ChatGPT search, Google AI Overviews, and Google AI Mode as separate observations. Google explains that its AI features can produce different responses and links, and that an AI Overview does not appear for every query. Record a missing AI answer as its own result rather than silently dropping that question. Google's AI feature guidance

Also decide what counts as your brand. Include the company name, owned domain, and any established trading name. Exclude unrelated businesses with similar names.

Build a small question set from real decisions

Start with twelve questions across three groups. Use actual sales questions, support conversations, and existing search queries where you have permission to use them. Remove personal details.

• Understand the problem: What causes it, what does it cost, and when does it need attention?
• Compare approaches: Which options fit different constraints, and what are their tradeoffs?
• Choose a provider: What evidence should a buyer check before hiring someone?

Write questions the way a prospective buyer might ask them. Do not insert your brand into every prompt. “Is our company the best?” tests a different behavior from an unbranded comparison.

Keep branded questions in a separate group. They can reveal incorrect facts about your company, but mixing them with unbranded questions makes a visibility total hard to interpret. Assign each question a stable ID and record why it belongs in the set.

Save enough context to repeat the observation

Use a fresh conversation for each test where the product allows it. Avoid adding your positioning document or telling the system which business you want it to recommend. Record any personalization, location, account state, or search setting you cannot control.

For each observation, save:

• Question ID, exact wording, date, and time.
• Product and search mode; model name only if actually shown.
• Language and location context, including anything unknown.
• Whether an answer appeared and whether your brand was named.
• Every relevant cited URL and the nearby claim.
• A saved answer or screenshot that a reviewer can inspect.

Repeat the fixed set on two later dates as a first check on stability. Three observations still do not establish market share. They help identify a result that happened once and disappeared. Do not keep rephrasing a question until the answer includes you.

Check the sources before scoring the answer

Open the cited page. Confirm that it is accessible, that it is the page you recorded, and that its content supports the nearby claim. A citation can point to your domain while the answer describes your service incorrectly.

Use separate fields for four outcomes: mentioned, cited, claim supported, and factually accurate about us. Mark unknowns explicitly. If a source is inaccessible, record “not verified” rather than treating the presence of a link as a pass.

Here is a synthetic example, not a customer result:

The useful finding is the coverage mismatch. A single “visibility score” would hide it.

Report counts with their denominator

For a fixed batch, report something like “named in 3 of 12 recorded answers” and “owned pages cited in 2 of 12.” Keep the surface, dates, and question set attached. These are sample counts, not estimates of all searches in your category.

Show the number of questions that produced no answer and the number of sources you could not verify. Compare the same question IDs across runs. If the set changes, start a new version instead of presenting the result as a continuous improvement.

Keep a short change log alongside the results. A page edit, product rename, altered test setting, or different location can matter to interpretation even when you cannot establish its effect.

Separate answer observations from website outcomes

A citation is an observation on someone else's interface. A referral visit is a session on your website. A qualified inquiry is a business outcome. Do not add them together.

OpenAI says ChatGPT search referral links include utm_source=chatgpt.com, which can support analytics reporting. Inspect what your own site actually receives and preserves before relying on that classification. OpenAI's publisher FAQ

Search Console provides clicks, impressions, click-through rate, and average position, with views such as query and page. Use those records for search performance, with the date range and filters saved. Search Console reporting documentation

Maintain three adjacent reports: answer evidence, referral sessions, and verified inquiries. A change in one can suggest a question to investigate in the others. It does not prove the same person moved through all three.

Choose one correction you can defend

Sort findings by what a buyer could misunderstand. An incorrect service boundary, missing price condition, or unsupported claim deserves attention before a cosmetic rewrite.

For each proposed change, write the affected page, the evidence, the exact correction, its owner, and the next review date. Keep the previous copy. A good first action might be clarifying what a service includes and linking to the evidence behind that statement.

Use AEO and SEO together to connect those findings to a page plan. For location-dependent services, add the local SEO baseline. The AI Search and SEO Growth Suite provides a structured starting point for this work; browse the AI search guides for the surrounding methods.