AI Visibility Audit Workflow

What this workflow does

This workflow creates an auditable baseline of how a defined brand or entity appears across a controlled set of questions. It separates four things that are often mixed together: whether an answer was produced, whether the brand was mentioned, whether a page was cited and whether the cited page actually supports the claim.

For Google Search, use the same foundations that apply to ordinary search visibility. Google’s guidance says that pages must be indexed and eligible to appear with a snippet to participate in its generative search features; there is no separate AI markup or special file that guarantees inclusion. Treat visibility as an observed outcome, not an entitlement.

When to use it

Use this workflow when you need to:

• establish a baseline before changing content or site structure;
• compare visibility across markets, languages or prompt families;
• identify which third-party and first-party sources answer an audience’s questions;
• verify whether brand claims are represented accurately;
• retest a documented hypothesis after a meaningful source improvement.

Do not use it to manufacture a single visibility score, reverse-engineer a hidden ranking system or promise citations. A one-off answer is a sample, not a stable pattern.

Define the measurement model

Write the decision rules before collecting answers. This prevents a favorable response from changing the definition of success after the fact.

Keep commercial value outside the visibility classification. A citation for an irrelevant prompt is not more useful than no citation.

Prepare a reproducible test

1. Define entities and audiences

Create a short entity sheet containing the canonical brand name, common variants, products, locations, named experts and the first-party URLs that can verify each claim. Add disambiguation notes where another entity shares a similar name.

2. Build prompt families around real tasks

Use prompts that represent what an audience is trying to accomplish, not prompts written only to force the brand into an answer. A balanced set can include:

• category discovery and comparison;
• problem diagnosis and how-to questions;
• branded factual questions;
• alternatives and suitability questions;
• current or location-sensitive questions where freshness matters.

Give every prompt a stable ID. Preserve wording, language, market and any follow-up sequence. When testing an updated prompt, create a new version rather than overwriting the old one.

3. Freeze the test conditions

Record the product, access mode, date and time, locale, signed-in state where known, conversation state and whether browsing or live search was visible. Start a new conversation for independent tests unless the test is explicitly about follow-up behavior.

Personalization and product changes may not be fully controllable. Document what is observable and mark the rest as unknown.

Run the audit

1. Capture raw evidence first

Save the exact answer and all visible citations before summarizing anything. Take a screenshot or export where the product permits it. Resolve redirecting citation URLs to the final destination, but preserve both the displayed and final URLs.

2. Classify each result

Use mutually exclusive top-level outcomes:

1. brand cited;
2. brand mentioned without a citation;
3. competitors cited or mentioned, but not the brand;
4. category discussed without named entities;
5. no usable answer.

Then add separate fields for accuracy, sentiment, citation support and relevance. This preserves nuance without double-counting a response.

3. Verify the citations

Open every cited page and locate the passage that could support the adjacent statement. Record whether the source is first-party or independent, whether the claim is current, and whether the page contains a clear author, publication context and update history where those details matter.

A citation gap can have several causes: the site may not contain the needed evidence, the evidence may be difficult to find, the page may not be eligible for Google Search, or the system may simply select a different source. Report the gap; do not invent the cause.

4. Compare source patterns

Group cited URLs by the audience task they solve instead of grouping only by domain. Look for patterns such as original research, primary documentation, clear comparisons, current product facts or concise explanations backed by deeper evidence. Use these as hypotheses for content improvement, not universal ranking factors.

5. Check Google Search eligibility

For Google AI features, confirm that the candidate page:

• is accessible to Googlebot and returns a successful response;
• is indexable and contains useful visible content;
• is eligible to appear with a snippet;
• has a clear canonical URL and descriptive internal links;
• follows Google’s spam policies;
• offers original, people-first value rather than a thin reformulation of existing pages.

Google states that familiar SEO best practices remain the foundation for its generative features. Avoid unsupported “AI optimization” artifacts. Resolve ordinary content, crawling and indexing problems first.

Turn evidence into actions

Prioritize recommendations in this order:

1. Correct unsupported, outdated or ambiguous first-party claims when they create an accuracy risk.
2. Publish the primary facts, methods, limitations or examples users need when evidence is missing.
3. Repair crawl, indexation, canonical or internal-link problems that affect discovery or eligibility.
4. Make authorship, dates, entities and supporting evidence easy to verify.
5. Improve headings, summaries, tables or media when the change helps the reader complete the task.

Each recommendation needs an owner, affected URL, evidence from the audit, expected user benefit and a retest condition. Do not prescribe schema markup unless it accurately represents content already visible on the page.

Evidence log template

Quality checks

• Every reported citation can be reopened and tied to a specific claim.
• Mentions, citations, accuracy and business relevance are reported separately.
• Prompt wording and test conditions are versioned.
• Recommendations trace back to visible evidence.
• Google-specific recommendations start with Search eligibility and people-first content.
• The report states sample size, products tested, markets, dates and known limitations.
• No result is described as proof of a ranking factor or a guaranteed future citation.

Limitations

Answer engines change quickly and may vary responses by time, location, account, model or conversation context. Citation interfaces can also omit or transform URLs. A repeatable audit reduces ambiguity but does not eliminate it.

Use several scheduled observations before calling an outcome a pattern. When a content change and a visibility change happen together, describe correlation unless the test design rules out credible alternatives.

Google documentation used

This workflow aligns its Google-specific checks with the following first-party guidance:

• Optimizing your website for generative AI features on Google Search
• AI features and your website
• Google Search technical requirements
• Creating helpful, reliable, people-first content
• Spam policies for Google Web Search