--- title: "How to Run a Claude AI Visibility Audit" slug: "claude-ai-visibility-audit" category: "claude" canonical_path: "/articles/claude/claude-ai-visibility-audit" meta_title: "Claude AI Visibility Audit: Prompts & Sources — Prime AI Visibility" meta_description: "How to run a Claude AI visibility audit: a repeatable test protocol for prompts, run conditions, source tracing, accuracy checks, competitor gaps, and a 30-day retest." author: "Alex Mannine" reviewer: "Bob Generale" date: "2026-08-06" last_updated: "2026-08-06" read_time: "13 min" keywords: - Claude AI visibility audit - Claude audit protocol - measure brand in Claude - AI answer accuracy - repeatable prompt testing featured_image: "/brand/articles/claude/claude-ai-visibility-audit.png" featured_image_alt: "A closed loop of five navy nodes joined by directional arrows returning to the start, with one node marked by a small citrine dot" og_image: "/brand/articles/claude/claude-ai-visibility-audit.og.png" cta_mid_headline: "Run a Claude visibility audit" cta_mid_body: "Prime AI Visibility runs your approved prompt set through Claude at recorded conditions, captures every answer and cited source, verifies the facts, and hands back a prioritised correction queue you can remeasure against." cta_mid_button: "Run a Claude audit" cta_bottom_headline: "Turn one audit into a retest habit" cta_bottom_body: "Fix the prompt set, record the conditions, and schedule the 30-day retest so your Claude visibility is a trend line, not a one-time screenshot." cta_bottom_button: "Start your audit" --- # How to Run a Claude AI Visibility Audit A Claude AI visibility audit is a repeatable test that records how Claude describes, recommends, and cites your brand under fixed, disclosed conditions. Approve a prompt set, record the product, web-search state, location, and date, run priority prompts several times, capture answers and cited sources, verify facts, compare competitors, and assign corrections. The output is a dated baseline you retest — not a permanent rank. > **Who this is for:** practitioners who need a defensible, re-runnable protocol for auditing Claude — not a one-off manual spot check that no one can reproduce next quarter. > **This is a test protocol, not a buyer's guide or a metric glossary.** For which software features make an audit auditable, and for the exact metric formulas, see the companion pages linked in the first section. ## Claude AI visibility audit: the short answer 1. **Fix the questions first.** A Claude AI visibility audit is only repeatable if the prompt set is approved, named, and versioned before the first run. 2. **Record the conditions.** Model or product, web-search state, location, date, and run number belong on every answer, or the audit cannot be reproduced. 3. **Keep the evidence.** Save the verbatim answer and every cited URL; a chart without its answers cannot be defended. 4. **Retest on a schedule.** One audit is a snapshot; the value comes from re-running the identical protocol after changes. ## Where this audit sits This page owns the *test protocol*. Two companions complete the set: the [Claude reporting tool features](https://primeaivisibility.com/articles/claude/claude-ai-visibility-reporting-tool-features) buyer's guide tells you which software will hold the evidence, and the [Claude visibility metric dictionary](https://primeaivisibility.com/articles/claude/measure-brand-visibility-in-claude) gives you the formulas and denominators you will apply to the answers this protocol captures. If you want to see a finished first measurement rather than the method, the [Claude AI visibility baseline case study](https://primeaivisibility.com/articles/ai-visibility/claude-ai-visibility-case-study) documents one anonymized audit end to end. Read them alongside this one; do not expect this page to repeat them. ## The Claude Repeatability Test: the protocol This is the original framework this page contributes — a nine-step test designed so a second analyst, months later, can reproduce your run. Each step is an instruction, not a theory. 1. **Approve the question set.** Agree the prompts with the people who own the answers (marketing, product, compliance where relevant), then freeze them. Freezing is what makes remeasurement honest. 2. **Record the conditions.** For each run, log the model or product, whether web search was on or off, the location, the account context, the exact prompt wording, and the date. 3. **Run the priority prompts repeatedly.** Do not run once. Run each priority prompt several times across separate occasions so you can measure consistency, not a single draw. 4. **Capture answers and citations.** Save the verbatim response and every cited URL. For web-search-on runs, the citations are your raw material for source tracing; Claude's web search cites the pages it draws from [[1]](#references). 5. **Verify the facts.** Check every claim Claude makes about you against an approved source of truth, and mark it correct, outdated, or wrong. Explicitly verify high-stakes answers against original sources — Anthropic's own guidance is that web results should be checked, not trusted blindly [[1]](#references). 6. **Compare competitors.** Run the identical set for named competitors so shares are comparable under the same conditions and window. 7. **Trace sources.** For each citation, note whether it is an owned page, a marketplace or directory, a review site, or a third-party article — and whether Anthropic's crawler can even reach your owned pages, since site owners control that access [[2]](#references). 8. **Assign corrections.** Turn each finding into a task with an owner and an evidence threshold, so the audit routes to work instead of dying in a spreadsheet. 9. **Retest under disclosed conditions.** Re-run the frozen set after changes, with the same conditions recorded, and compare against the dated baseline. ## Audit scope and the minimum viable prompt set You do not need hundreds of prompts to start; you need the *right* ones, frozen. A minimum viable set covers the questions a real buyer would ask, structured by intent rather than by keyword volume. A practical starter shape: | Prompt class | What it tests | Example shape | |---|---|---| | Branded | Does Claude describe you accurately? | "What is [brand] and what does it do?" | | Category | Are you present when the category is asked about? | "What are the best tools for [job]?" | | Comparison | How does Claude frame you against rivals? | "[brand] vs [competitor] for [use case]" | | Use-case | Are you recommended for the job to be done? | "How do I [task] for [audience]?" | | Objection | Does Claude repeat a myth or weakness? | "Is [brand] worth it / any good?" | | Recommendation | Who does Claude pick when asked to choose? | "Which [category] tool should I buy?" | Balance the set across classes rather than loading it with branded prompts, because branded prompts flatter you and category prompts tell the truth. Keep the set small enough to run repeatedly and large enough to cover the funnel. Commerce brands extend the same classes down to product and constraint questions; the [e-commerce AI visibility audit](https://primeaivisibility.com/articles/ai-visibility/ecommerce-ai-visibility-audit) shows how SKU-level and comparison prompts change the picture. ## Search state, location, follow-up context, and run count Four conditions change Claude's answers and therefore must be fixed and recorded: - **Search state.** Web search on lets Claude retrieve and cite live pages; off relies on training data. Never blend the two in one metric [[1]](#references). - **Location.** Locale can change what is retrieved and recommended; record it and hold it constant within a comparison. - **Follow-up context.** A follow-up question inside a conversation inherits context from earlier turns. Decide whether you are testing cold single-turn prompts or a defined conversation, and keep it consistent. - **Run count.** State how many times each prompt was run. Consistency across runs is a finding in itself. ## Evidence capture template Capture the same fields for every answer. Rendered as a table so it is legible and re-runnable: | Field | Example value | |---|---| | Prompt ID | CAT-03 | | Prompt class | Category | | Model / product | Claude (record exact product) | | Web-search state | On | | Location | US | | Date | 2026-08-06 | | Run number | 2 of 5 | | Mentioned? | Yes | | Recommended? | No | | Cited your page? | No | | Cited URLs | competitor.com/guide; review-site.com/list | | Factual accuracy | Correct | | Framing | Listed among alternatives | | Correction owner | Content lead | Keeping this template identical across runs is what lets a later analyst reproduce the audit — the same discipline behind the reusable [agency AI visibility audit template](https://primeaivisibility.com/articles/agencies/ai-visibility-audit-template-for-agencies), adapted to a single engine. ## Accuracy and source review Two reviews matter more than the counts. First, **accuracy**: an answer that names you but gets a fact wrong is a liability, so every checkable claim is verified against your source of truth and flagged if outdated or incorrect. In high-stakes categories a wrong answer needs a defined response path — see how healthcare brands [monitor and correct AI misinformation](https://primeaivisibility.com/articles/healthcare/healthcare-ai-misinformation-monitoring). Second, **source quality**: for web-search-on answers, examine which sources Claude cited and whether your owned pages are reachable and current. If Anthropic's crawler is blocked from your site, owned pages are less likely to surface as citations — a fixable technical issue, not a mystery [[2]](#references). Where the fix is technical or editorial and you lack internal capacity, Percepture provides [technical and source remediation](https://percepture.com/seo-insights/technical-seo-audit-service/) as managed implementation. *Disclosure: Prime AI Visibility and Percepture have a commercial relationship. Prime provides visibility intelligence and diagnosis; Percepture provides managed implementation. Recommendations and comparisons use the criteria shown on this page.* ## Competitor gap map Run the frozen set for two or three named competitors and lay the results side by side by prompt class. The gap map answers the question executives actually ask: not "are we visible," but "where does Claude pick someone else, and why." A gap concentrated in comparison and recommendation prompts points to positioning and third-party evidence; a gap in category prompts points to missing or unreachable owned content. Route each gap to the correction step with its own owner. For turning those routed corrections into governed workflow, see how to [operationalize AI visibility data into CRM and content workflows](https://primeaivisibility.com/articles/automation/ai-visibility-crm-content-workflows). ## Limitations and change log State the limits plainly and keep a change log. Claude does not return the same answer every time, does not expose an internal ranking, and does not reveal which uncited pages shaped an answer. The audit therefore measures observed outputs under disclosed conditions — nothing more. Every change to the prompt set, the conditions, or the source of truth goes in a dated change log, so a later trend reflects real movement and not a quiet edit to the method. ## The 30-day audit-to-retest sequence - **Days 1–3:** approve and freeze the prompt set; define conditions and the source of truth. - **Days 4–7:** run the priority prompts repeatedly; capture answers, citations, and conditions. - **Days 8–12:** verify accuracy, trace sources, and build the competitor gap map. - **Days 13–15:** assign corrections with owners and evidence thresholds. - **Days 16–28:** implement fixes (owned content, entity clarity, crawler access, third-party sources). - **Days 29–30:** retest the identical frozen set under the same recorded conditions; compare to the dated baseline and update the change log. ## A note on variability AI answers can vary by platform, model or product, search state, location, prompt wording, time, and repeated run. Results describe a defined observation method, not a permanent universal rank. ## Methodology and sources This article was authored by Alex Mannine; the methodology was reviewed by Bob Generale. It describes a repeatable audit protocol for observing Claude's outputs and is grounded in Anthropic's published web-search and crawler documentation. It explicitly advises verifying high-stakes answers against original sources, makes no ranking or citation promise, and does not claim access to Claude's internals. Where Percepture is recommended, the commercial relationship is disclosed above. > **Run a Claude visibility audit** > > Prime AI Visibility runs your approved prompt set through Claude at recorded conditions, captures every answer and cited source, verifies the facts, and hands back a prioritised correction queue you can remeasure against. > > **[Run a Claude audit](https://app.primeaivisibility.com/sign-up)** ## References 1. Anthropic / Claude Support, *Enable and use web search* (2026). 2. Anthropic / Claude Support, *Does Anthropic crawl data from the web, and how can site owners block the crawler?* (2026). 3. Google Search Central, *AI features optimization guide* (2026). ## Next steps 1. **[Define the metrics you will apply to each answer](https://primeaivisibility.com/articles/claude/measure-brand-visibility-in-claude)** so the audit produces numbers with denominators, not impressions. 2. **[Pick software that keeps the audit auditable](https://primeaivisibility.com/articles/claude/claude-ai-visibility-reporting-tool-features)** by retaining raw answers and traced sources. 3. When you are ready, **[create a Prime AI Visibility workspace](https://app.primeaivisibility.com/sign-up)** and bring 10 buyer prompts. ## Frequently asked questions **How many prompts does a Claude AI visibility audit need?** Fewer than most people expect, but frozen. A minimum viable set covers branded, category, comparison, use-case, objection, and recommendation classes — small enough to run repeatedly and large enough to represent how real buyers ask. **Why run each prompt more than once?** Because Claude's answers vary between runs. A single answer is one draw; running each priority prompt several times lets you report consistency and separate a real finding from run-to-run noise. **Do I need web search on for the audit?** Run and record both states, but keep them separate. Citation and source-tracing steps only apply to web-search-on answers, while search-off answers show what Claude says from training data alone. **What should I verify by hand?** Every high-stakes factual claim Claude makes about you, checked against your approved source of truth and against the original sources — Anthropic's own guidance treats web results as material to verify, not to trust blindly. **How is this audit different from a baseline case study?** This page is the reusable method; the baseline case study is one completed measurement written up end to end. Use the protocol here to produce your own dated baseline, then retest against it. **What happens after the audit finds problems?** Each finding becomes a correction with an owner and an evidence threshold, then the frozen prompt set is retested after the fix. Prime supplies the diagnosis and priorities; managed remediation is a separate discipline handled under the disclosed Percepture relationship. > **Turn one audit into a retest habit** > > Fix the prompt set, record the conditions, and schedule the 30-day retest so your Claude visibility is a trend line, not a one-time screenshot. > > **[Start your audit](https://app.primeaivisibility.com/sign-up)**