What should an AI visibility audit include?

An AI visibility audit should show what happened when relevant buyer questions were asked, identify information your business can improve, and give each recommended change an owner and a completion check. You should leave knowing what to do next, not just whether a score went up or down.

Start with the business decision

Before anyone collects answers, write down the decision the audit needs to support. You might need to understand whether people can discover a new service, whether your coverage area is being described correctly, or which existing pages deserve attention. These are different questions and need different evidence.

Evidence, action, recheckFictional weekend-cleaning example. Later AI answers are checked separately.
  1. EvidenceOld listing advertises weekend work.
  2. ActionOffice manager requests a listing correction.
  3. RecheckSave a corrected listing or an unresolved request.

Describe the intended customer, the service they need and the places you actually serve. Include important limits: a commercial cleaning company may serve offices but not homes, for example. An audit that treats every cleaning question as equally relevant could recommend work that attracts the wrong inquiries.

Our recommended starting brief is one short paragraph followed by a list of decisions you want to make. Say which tools and markets are in scope, whether implementation is included, and who can confirm business facts. That gives the reviewer a useful boundary and gives you a way to judge the finished report.

Choose questions that resemble customer decisions

Build the question list from sales conversations, inquiry forms, support questions and the language customers use. Include discovery questions, comparisons and practical concerns. For a fictional office cleaner, examples could cover finding a local provider, arranging evening access and understanding what an initial quote includes.

Separate questions that name your business from those that do not. Asking what your company does tests a different situation from asking which companies serve a neighborhood. Keep both if they matter, but report them separately so an easy branded answer does not hide weak discovery evidence.

Agree the question list before reviewing the results. Keep questions that produce disappointing or unhelpful answers. If the reviewer changes wording during the audit, retain the original and explain the reason. You need a record of what was tested, not a collection selected afterward because it makes the service look effective.

Require a method another person can follow

Ask for the exact questions, tool names, dates, relevant settings, account state and location assumptions. Record whether the test started a new conversation or continued an existing one. If several people collect answers, give them the same instructions and record any differences rather than assuming their sessions were equivalent.

Set a repeat plan that fits the budget. More observations mean more material to review, so a small question set checked carefully can be more manageable than a huge export nobody reads. There is no universal number of questions or repetitions that turns a business audit into a complete view of all customer experiences.

Keep unavailable responses and failed attempts in the log with their reason. Do not silently remove them to improve a percentage. A reader should be able to distinguish a completed answer with no company mention from a question the reviewer could not test at all.

Save complete answers and define what counts

A screenshot of a company name is useful context, but it is not the whole record. Save the complete response, its links and the question that produced it. Preserve the original capture so someone can check the reviewer's interpretation without relying on a rewritten summary.

Define a mention, a recommendation and a link to your site separately. A response can name a company while warning that it may not fit the request. A linked article might explain a topic without recommending its publisher. Decide how those cases will be recorded before calculating a score.

Add a brief note about accuracy and relevance beside each observation. An incorrect phone number matters even when the business was mentioned. A company with a similar name is not your company. These notes make the record useful to the person who will investigate the issue later, rather than leaving them to reconstruct the original conversation.

Check what the cited pages actually support

Open the links connected to important claims. Check whether the page describes your business, a competitor or the category generally. Note the publisher, the relevant passage and whether the information is still current. A visible link does not remove the need to read what it says.

Compare the answer with the factsFictional cleaning example. A link alone does not prove what caused the answer.
Captured answer
Names the company for weekend work.
Old directory
Still advertises weekend availability.
Current service
Weekday evening office cleaning.

Separate pages you control from pages owned by other people. You can edit your service page, but correcting an old directory entry may involve a publisher's process. Record that difference in the action plan so nobody promises a change they cannot make directly.

Avoid turning an observed link into a claim about how the answer was produced. If an answer cites a directory with an old service description, the audit can establish that the cited page is stale. It cannot establish, from that observation alone, that updating the directory will cause a particular recommendation. Keep the confirmed problem and the proposed remedy distinct.

Explain website access checks in plain language

Ask the reviewer to check whether your important public pages can be found, opened and understood by the search systems in scope. The report should name each affected page and explain the issue in ordinary language. A page that requires a login and a page excluded from search are different problems.

For Google's generative AI search features, the current official optimization guide says a page must be indexed and eligible for a search snippet, and the site must be included through Search Console's generative AI control. Meeting those conditions does not guarantee inclusion in an answer.

Ask the reviewer to inspect that inclusion setting with the owner, including any inherited choice. Record what was checked without changing a deliberate exclusion automatically. For other answer tools, request their own current documentation. One provider's settings do not establish another provider's requirements.

Verify the business facts people rely on

Compare your website and important public listings against a fact sheet approved by someone in the business. Include the trading name, contact route, services, coverage, opening hours and meaningful exclusions. Where prices or qualifications are mentioned, ask the responsible person to confirm exactly what can be published.

For an eligible local business, check its Business Profile against Google's representation guidelines. Those guidelines call for accurate real-world representation, including the business name and address or service area. Do not add an invented location or unsupported claim to make a listing look more appealing.

Keep a list of disagreements rather than deciding by majority vote across websites. Three directories repeating an old telephone number do not make it current. Confirm the fact with the business, correct the pages you control, and record requests sent to other publishers. Keep sensitive customer information out of the evidence you circulate.

Review whether a buyer can make a useful decision

An audit should read the pages that a potential customer reaches. Can the person tell what you do, who it is for, where you operate and what happens next? Clear answers are useful even when the audit cannot show how any particular answer system weighs the page.

Look for missing decision details rather than prescribing more articles immediately. A service page may need a clearer description of scope, a genuine example or an explanation of the quoting process. Another generic introduction to the industry may leave the buyer's actual uncertainty unresolved.

Our recommendation is to connect each proposed page change to a customer question and a business fact you can verify. Ask a colleague who handles inquiries to read it before publication. Do not invent customer stories, credentials, delivery times or results to fill a gap. When something varies by engagement, explain the factors instead of presenting a universal promise.

Turn findings into a prioritized action register

For every issue, request the evidence, the affected page or account, the recommended action, an owner and a completion check. Include effort, dependencies and any uncertainty. A useful entry should let someone begin work without holding another meeting to discover what the recommendation means.

Choose work the evidence supportsFictional audit examples. Evidence and the buyer's needs determine priority.
Confirmed error
Correct a service-area mistake before commissioning broad new content.
Uncertain cause
Investigate it; keep the uncertainty visible.
Missing buyer detail
Add information that helps someone judge the service.

Open the blank action register in a spreadsheet and add one row per finding. Its eight columns keep the evidence, issue, priority reason, action, owner, dependency, completion check and status together. The filled CSV uses the fictional weekend-cleaning example below; replace its teaching descriptions with your own saved evidence.

Prioritize according to the decision you set at the start. Correct a confirmed service-area mistake before commissioning a broad new content program. If a recommendation is only a hypothesis, label it as something to investigate. Do not give a speculative explanation the same certainty as a directly observed broken page.

Define completion around work that can be checked. An owner-approved service description published at the agreed address is a checkable deliverable. Being named in every AI answer is not something the website editor controls. Record costs and approvals separately so the audit does not become an open-ended commitment to every task it discovers.

Recognize a useful finding in a fictional example

Imagine a fictional office cleaning company that stopped offering weekend work. Its current service page explains weekday evening cleaning, but an older directory entry still advertises weekend availability. An audit captures an answer that names the company for a weekend request and links to that directory.

A useful finding saves the exact question, complete response and directory passage. It confirms the current policy with the owner. The action is to request correction of the directory, check other known listings and make the existing website explanation easy to find. The office manager owns the request and records whether the publisher accepts it.

The completion evidence is the corrected listing or a documented unresolved request. A later repeat check records what the tool actually says then. The audit does not promise that the correction will immediately change every answer, and it does not call the original mention a success merely because the company appeared. This example illustrates a review method; it is not a customer result.

Keep visibility, website visits and inquiries separate

Agree which measurements answer which questions. Saved answer checks describe the sessions you tested. Search reporting describes activity covered by that reporting system. Inquiries describe people who contacted the business. Keep those records connected where possible, but do not present them as interchangeable measures.

Google's Generative AI performance report reports impressions for AI Overviews and AI Mode. It may not appear until the site has enough impressions. Its absence is not a complete measurement of your business across all answer tools. Record the report and date range actually available to you.

Google's ordinary Search performance report also has its own grouping and aggregation rules. If totals differ, ask the reviewer to explain the filters and definitions. A movement in one report, or in your small answer sample, does not establish that an audit caused more sales. Note other business or website changes during the comparison period.

Accept the evidence and decide the next step

Before accepting the audit, check that you have the agreed question list, method, saved observations, source review and prioritized actions. Ask the reviewer to walk through one finding from the original answer to the proposed fix. You should understand why the work is recommended and how someone will confirm it is finished.

Arrange handover in a format your team can keep using. Request editable findings, evidence links and a clear list of checks that were unavailable or outside scope. If someone needs account access for the initial review, Google's hiring guidance recommends read access to Search Console at the audit stage. Agree any later editing access separately.

Then choose a manageable next step: complete a small set of confirmed corrections, investigate an important unknown, or pause because the likely work does not fit your priorities. Set a review date and retain the original baseline. The audit has earned its place when it supports that decision with evidence you can inspect and actions your team can own.

Sources

Reviewed

  1. Google Search Central · Optimizing for generative AI features
  2. Google Search Console · Search generative AI control
  3. Google Search Console · Generative AI performance report
  4. Google Business Profile · Representation guidelines
  5. Google Search Console · Search performance report
  6. Google Search Central · Hiring SEO help

Let’s find your next step

Leave your email and tell us what you want to understand or improve. Start with analytics help, an app idea or both.

We’ll reply by email about your request. By submitting, you agree to our Terms and acknowledge our Privacy Policy.