What should an AI visibility audit actually answer?
An AI visibility audit should explain when assistants omit your company, what they say instead, and which evidence could change the answer. A useful audit is not a single visibility score. It is a diagnosis connecting a buyer question to an observed response, a competing answer, and a practical next action.
Start with the questions a prospective customer would ask before choosing a provider. Record whether ChatGPT, Perplexity, Gemini, Claude, Grok, and Google AI Overviews mention your company, describe it accurately, recommend it, or cite a page from your site. These are different outcomes. A company can be mentioned without being recommended, recommended without being cited, or cited for an irrelevant claim.
The audit should also identify whether the omission repeats across engines and prompts. Repeated omission suggests a broader discoverability or evidence problem. Omission from one engine may reflect different retrieval, training, browsing, or answer-generation behavior. The final output should therefore rank findings by buyer impact and confidence, rather than treating every absent mention as the same defect.
For more context, read Why Is My Brand Missing From Ai Answers.
Which buyer questions belong in the audit?
The audit should use real buyer questions, not only branded searches or generic prompts about your category. Assistants often omit a company when a buyer asks for a shortlist, comparison, recommendation, or fit judgment, so those situations deserve priority.
Build prompts around the decisions customers make. Include category questions, questions that add location or company size, comparisons with known alternatives, questions about a specific use case, and questions that reveal objections such as implementation effort or switching risk. Keep the wording natural enough to resemble an actual conversation, while preserving a stable version for later retesting.
Separate prompts where the company is eligible from prompts where it is not. A local provider should not be expected to appear in every global recommendation, and a specialist product may reasonably be absent from a broad consumer answer. Record the intended audience, category, geography, and buying stage for each prompt. That context prevents a misleading conclusion that the brand has poor visibility when the test simply asked the wrong question.
Prompt design also determines whether an audit can guide content work. A vague prompt produces a vague action. A decision-specific prompt reveals which qualification, comparison, proof point, or source is missing.
For more context, read How AI Engines Decide Which Brands to Mention.
How do you separate a knowledge gap from a selection gap?
A knowledge gap means the assistant lacks reliable information about your company, while a selection gap means the assistant knows the company but does not choose it for the buyer's situation. The distinction determines whether you need clearer facts or stronger reasons to be included.
Look for signs of a knowledge gap first. The response may confuse your category, use an outdated description, omit a documented service, or fail to connect your company with a location or audience you serve. Missing or weak sources can reinforce the diagnosis. Correcting the company description, publishing clear service and audience pages, and making important facts easy to verify may help.
A selection gap appears when the assistant describes your company accurately but recommends alternatives. In that case, adding more general brand copy may not help. Compare the reasons given for competitors with the evidence available for your own offer. The missing element may be a use-case explanation, a credible comparison, a limitation statement, or independent discussion that supports the recommendation.
The same prompt can expose both gaps. An assistant may know your name but misunderstand your best-fit customer. Treating every omission as a crawl problem wastes time and can produce more pages without improving the buyer's decision.
What evidence should you capture from each response?
Capture the complete answer, sources, competitors, claims, and prompt context before deciding what to change. A screenshot or copied response alone is insufficient because the useful evidence includes how the answer was produced and what it relied on.
For every test, record the engine, date, prompt, location or account context where relevant, whether browsing was active, and the exact response. Mark your company's status as absent, mentioned, recommended, cited, or misrepresented. Record the alternatives named and the reasons attached to each one. A citation to your domain is not automatically a success if it supports a minor or inaccurate claim.
Classify each source as an owned page, third-party publication, directory, review, documentation, or another source type. Check whether the cited page actually supports the statement made. Unsupported citations and stale descriptions deserve separate treatment from ordinary non-mention. Do not infer that an uncited answer proves the engine never saw your site, because answer generation and source display do not reveal every retrieval step.
Preserve the original wording. Paraphrasing can hide a qualification or make a weak recommendation appear stronger than it was. Consistent evidence capture turns an anecdotal complaint into a repeatable record that a content, search, or product team can review.
When is an omission a source problem rather than a content problem?
An omission is more likely a source problem when important facts exist on your site but assistants cannot reliably retrieve, interpret, or verify them. Publishing more content will not solve every failure if the underlying pages are inaccessible, contradictory, poorly connected, or unsupported elsewhere.
Check whether the relevant page is available to users and search systems, presents one clear description, and links related facts together. Review titles, headings, structured information, canonical choices, redirects, and outdated pages. Confirm that key claims are not buried only in images, interactive elements, or language that requires substantial inference. Search documentation from Google, OpenAI, Perplexity, and other relevant platforms for current guidance because access and display rules change.
A source problem can also come from conflicting external information. An old directory entry, incomplete profile, or third-party description may compete with your preferred positioning. Create a source map showing which page supports each important claim and where an assistant could encounter a different version.
Do not assume technical accessibility guarantees a mention. Retrieval makes information available; recommendation still depends on relevance and evidence. The audit should label source repair as a prerequisite when the facts are missing or unreliable, not as proof that improved access will produce a particular answer.
Which change should you make first?
Make the first change where buyer impact, evidence strength, and implementation effort overlap. A technically easy fix is not automatically the best fix, and a dramatic content project may be unnecessary when one inaccurate source is causing repeated confusion.
Prioritize a finding when it affects a high-value buyer question, appears across more than one engine or prompt, and has a clear correction. An inaccurate category description usually deserves attention before a low-priority omission. A missing comparison page may matter more than a general company overview when customers repeatedly ask which option fits a particular use case.
Use a simple decision sequence. Correct factual errors first. Repair access or source conflicts next when they prevent important facts from being verified. Then address selection gaps with evidence that explains who the company suits, what problem it solves, and where it is not the right choice. Finally, improve wording and coverage for lower-impact prompts.
Every proposed change should state the expected observation, not an assumed result. For example, the goal may be that an assistant accurately connects a service with a defined customer type and cites a supporting page. Retest the original prompt after publication, rather than judging success from page traffic or rankings alone.
How should results be compared across ChatGPT and other engines?
Compare engines by the same buyer task, not by forcing identical visibility expectations across ChatGPT, Perplexity, Gemini, Claude, Grok, and Google AI Overviews. Each product can differ in retrieval, browsing, source presentation, context, and answer format, so a result is meaningful only when its test conditions are recorded.
Run a matched prompt set where wording, audience, geography, and requested outcome remain stable. Then compare whether each engine identifies the company, represents it accurately, recommends it when appropriate, and provides supporting sources. Keep engine-specific observations separate from the shared diagnosis. For example, a missing citation in one answer is not equivalent to a missing mention across every answer.
Look for patterns in the reason for omission. One engine may rely heavily on accessible web sources, while another may produce a response without visible citations. A company that appears in a source-rich answer but not a concise answer may have a presentation problem rather than a basic discoverability problem.
Avoid averaging unlike outcomes into one score. A single number can conceal that the company is strong for one use case and absent for another. A decision-ready comparison preserves the prompt, response, source evidence, and recommended action for each engine.
How do you turn an audit into a repeatable test?
Turn the audit into a repeatable test by freezing a prompt set, recording response conditions, and linking every change to a later retest. Repeating the same question without preserving the original context makes apparent progress difficult to interpret.
Create a test register with the prompt, target audience, intended company fit, engine, date, response, sources, and diagnosis. Add a version label whenever the wording changes. Keep a separate exploratory set for new customer questions so the core trend remains comparable. Do not treat one response as conclusive evidence, because assistant outputs can vary with context and time.
Retest after meaningful changes such as a corrected company description, a new comparison page, a resolved source conflict, or a material product change. Review both positive and negative movement. An answer may begin mentioning the company while introducing an inaccurate qualification, or a new page may improve one prompt while creating contradictory signals elsewhere.
The audit becomes operational when every result ends with an owner, a proposed change, and a reason to retest. Marketing can then distinguish content work from technical repair, while founders can see which buyer questions remain unanswered. The purpose is not to chase every output. It is to learn which evidence changes the answers that matter.
Related reading
Sources consulted
- Google Search Central (developers.google.com)
- OpenAI Platform Documentation (platform.openai.com)
- Perplexity API Documentation (docs.perplexity.ai)
- Google Search Help (support.google.com)
Drafted with AI assistance from our own research and Search Console data, and reviewed by Rahul A before publishing. Rules and prices change; check the linked official source before you act.