AI visibility audit: A practical 7-step checklist

For a Google-specific implementation, use the verification-first Gemini SEO workflow to separate Search eligibility from observed links and mentions across Gemini Apps, AI Overviews, and AI Mode.

An SEO audit can tell you whether a page is crawlable, indexable, internally linked, and technically sound. An AI visibility audit starts with those foundations but asks a different outcome question: when a buyer asks an answer engine for advice, does the response mention your brand, cite your domain, describe you accurately, or recommend a competitor instead?

That distinction matters. A page can be accessible without being selected as a source, and a brand can be mentioned without its own website being cited. Treating every signal as one score hides the reason a result is weak.

What an AI visibility audit should produce

A useful audit is not a screenshot of one answer and not a mysterious score. It should leave your team with five durable outputs:

1. a written scope covering platforms, market, language, entities, and observation dates;
2. a fixed set of buyer prompts that can be rerun;
3. answer-level evidence for mentions, citations, recommendations, sentiment, and competitors;
4. a separate website-readiness review covering access, indexability, structure, and identity;
5. an action backlog with owners, evidence, priority, and a remeasurement date.

The audit is broader than a technical site check and narrower than a complete marketing strategy. Its job is to identify which observable gap should be fixed next.

AI visibility audit vs. SEO audit

Run both. The website-readiness layer overlaps with SEO; the observed-answer layer does not.

Step 1: Freeze the audit scope before collecting answers

Write down the configuration first. At minimum, record:

• the AI platforms you will observe;
• country or market and answer language;
• the brand, products, and spelling variants included as entities;
• the competitors used for comparison;
• the collection window and intended remeasurement date.

Do not mix markets or languages into one average. A buyer asking in Taiwan Traditional Chinese may receive different brands and sources from a US-English buyer, even when the literal topic is similar.

Step 2: Build a balanced buyer-prompt set

Use prompts that could change a shortlist, not only branded questions that already contain your name. A practical starter set includes four groups:

• Category prompts: “What are the best tools for monitoring brand visibility in AI search?”
• Problem prompts: “How can I find out when ChatGPT recommends my competitors?”
• Comparison prompts: “Compare tools for tracking AI mentions and citations.”
• Branded prompts: “What does ExampleCo do, and who is it for?”

Category, problem, and comparison prompts test discovery. Branded prompts test accuracy. Keep the wording stable for later comparisons and document deliberate exclusions. The guide to finding buyer prompts provides a fuller sourcing and scoring method.

Step 3: Capture an observed-answer baseline

Run the same configured set across the platforms in scope. Preserve the full answer and cited sources with the platform, prompt, market, language, and timestamp. For every response, record separately:

• whether the brand is mentioned;
• whether the brand's own domain is cited;
• whether the brand is recommended or merely listed;
• the framing or sentiment;
• which competitors appear;
• which third-party pages are cited.

Do not turn one response into a “rank.” Answers vary with model updates, retrieval state, location, and wording. Repeat observations and report the sample definition beside every rate.

Step 4: Audit website readiness separately

Now test whether the pages that should support those answers are eligible and understandable. Check the exact public URLs, not only the homepage:

1. useful HTTP response and no accidental noindex;
2. one stable canonical URL with consistent redirects and sitemap references;
3. access for the relevant search or citation crawler;
4. important copy present in rendered HTML, not hidden behind a fragile interaction;
5. clear H1/H2 structure and direct answer passages;
6. metadata and structured data that match visible content;
7. explicit organization, product, author, and source context.

Keep crawler purposes distinct. For ChatGPT search discovery, review OAI-SearchBot rather than using GPTBot as a substitute. For Google search features, normal Googlebot eligibility remains relevant; a separate content-use control is not proof of search access. AEO Mantis's site audit can surface access, extractability, and identity findings, but passing those checks still does not guarantee a citation.

Step 5: Trace citations and entity evidence

For prompts where competitors appear and you do not, inspect the sources that support the answer. Classify each cited page:

• your page;
• a competitor page;
• an independent publication or directory;
• a community or user-generated source;
• an outdated or inaccurate source.

Then ask two questions. First, do you have a public page that answers the same buyer task with clearer, sourced evidence? Second, do independent sources describe your brand consistently enough to corroborate that page?

The output is not “copy the cited competitor.” It is a source gap: the answer lacks a sufficiently clear first-party page, sufficient third-party corroboration, or both.

Step 6: Compare competitors on the same evidence base

Use the same prompt set, platforms, markets, and dates for every brand. Compare:

• mention rate and citation rate;
• share of voice within the defined prompt sample;
• prompts where only competitors appear;
• repeated sources that support those competitors;
• differences in positioning, accuracy, and recommendation context.

This turns competitor analysis into a controlled comparison. Changing the questions for each brand makes the result impossible to interpret.

Step 7: Prioritize fixes and set remeasurement

Do not rank fixes only by how dramatic they sound. Work through dependencies:

1. Eligibility blockers: inaccessible pages, accidental noindex, broken rendering, or conflicting canonicals.
2. Identity errors: unclear product facts, inconsistent names, or outdated third-party descriptions.
3. Answer gaps: missing direct explanations, comparisons, evidence, or relevant public pages.
4. Citation gaps: weak sourcing, absent independent corroboration, or no authoritative page that supports the claim.
5. Measurement gaps: unstable prompts, missing answer captures, or mixed markets and languages.

Give each action an owner, evidence link, expected observable change, and review date. Rerun the unchanged audit scope after the fix. If you change the prompts, platform mix, or market, label that as a new baseline rather than a trend.

Use a two-layer evidence matrix

The matrix below prevents technical checks from being mistaken for answer performance.

Worked example: from absence to an owned test

Imagine a B2B software brand audits 20 unbranded buyer prompts in US English across three answer platforms. The brand appears in 4 of 60 captured responses; its domain is cited in 1. A competitor appears in 18 and is repeatedly supported by two comparison pages.

The site's product page is crawlable and self-canonical, so the team does not file a generic “fix robots.txt” task. Instead, it finds that the page never answers the comparison criteria used in the buyer prompts, while third-party profiles describe an old positioning. The action plan becomes specific: add a sourced comparison section, correct the public profiles, and rerun the same 20-prompt sample after the pages are recrawled.

Those numbers are illustrative, not benchmarks. The method is the point: observed answer evidence determines the problem, and the readiness review prevents the team from guessing at the cause.

Common audit mistakes

• Using only branded prompts. They test recognition, not discovery.
• Counting mentions as citations. A name can appear without a link to your domain.
• Calling one answer a ranking. Answer order is not a stable SERP position.
• Combining markets. Regional sources and recommendations can differ materially.
• Scoring llms.txt as mandatory. It is an optional discovery convention, not a universal eligibility requirement.
• Treating schema as a switch. Structured data should match visible content; it cannot guarantee selection.
• Publishing a score without evidence. Stakeholders need the prompt set, sample size, dates, and answers behind it.

How often should you run an AI visibility audit?

Run a full audit when establishing a baseline, entering a market, changing positioning, or launching an important product family. For ongoing work, monitor the stable core prompts continuously or on a defined cadence, then repeat the deeper readiness and source review quarterly or after material site changes.

Keep publication, retrieval eligibility, answer visibility, referral traffic, and conversion as separate stages. Improvement in one does not prove improvement in the next.