AI Visibility Audit: A Step-by-Step Checklist You Can Run Yourself

Before you commission an AI visibility programme, spend a day establishing where you actually stand. Most organisations discover that two or three concrete blockers explain the bulk of the problem, and that at least one of them is a configuration issue nobody knew about. This checklist is designed to be run internally, in order, without specialist tooling.

What does an AI visibility audit cover?

An AI visibility audit answers three questions in sequence: can retrieval systems reach your content, can they extract usable answers from it, and are they currently citing you. The order is deliberate. There is no point analysing content quality on pages that a crawler cannot fetch, and no point measuring citations before you know whether the access layer is open.

Work through five sections. Each produces a specific finding, not an impression.

Section 1: Access — can they reach you?

Every item here is binary and fixable within days. Start here.

1. Read your robots.txt in full. Fetch it directly at your domain root. Identify every user-agent group and note precisely which AI agents are disallowed. Confirm the block matches an actual decision someone made, rather than an inherited default.
2. Check for retrieval-agent blocks specifically. Separate the training crawlers from the search and retrieval agents. Blocking the latter makes you ineligible for citation in those systems. Confirm this is intentional.
3. Test server-rendered content. View the raw HTML source of three important pages, with JavaScript disabled or by inspecting the initial response. If the substantive body text is absent, some retrieval systems will never see it.
4. Check for firewall or WAF-level blocking. Bot-mitigation rules frequently block AI user-agents independently of robots.txt, and marketing teams are rarely told. Ask your hosting or security owner directly.
5. Review your access logs for AI user-agents over the last 30 days. Which agents are hitting you, how often, and which pages? Zero AI crawler traffic is a finding in itself.
6. Confirm response speed and status codes on your key pages. Timeouts, redirect chains and soft 404s all reduce retrievability.

Section 2: Extraction — can they use what they find?

Sample eight to ten of your most commercially important pages and assess each against the same criteria.

1. Does the page have one clear topic? Pages covering several unrelated things dilute the match to any specific question.
2. Do headings ask questions? Headings phrased as buyer questions map far more directly onto prompts than headings phrased as internal labels.
3. Run the self-containment test. Take five paragraphs at random. Read each in isolation. Does it name its own subject, and make a complete claim without depending on what came before? Count how many pass.
4. Is the answer at the top of each section? Check whether conclusions are stated first and elaborated afterwards, or built towards over several paragraphs.
5. Are there lists and, where genuinely comparative, tables? Note pages that are wall-to-wall prose.
6. Is anything load-bearing trapped in an image? Diagrams without text equivalents are invisible for extraction.
7. Are claims specific? Flag vague quantifiers — “significantly”, “many”, “much faster” — which cannot be extracted as checkable statements.

Score each page out of seven and rank them. Our AI citability scorer runs this assessment at site scale if the manual sample takes too long.

Section 3: Attribution — will a quote carry your name?

1. Check entity consistency. Compare how your organisation is named and described on your homepage, your about page, your structured data, and your three most prominent external profiles. Inconsistent naming or category language is the most common cause of assistants describing a firm inaccurately.
2. Confirm organisation-level structured data exists and matches the on-page facts. Contradiction between markup and visible content is worse than absence.
3. Look for named methods and frameworks. Content that describes a generic approach can be quoted without attribution. Content that names a specific method carries your name with it.
4. Check author and organisation signals on substantive content. Who is stating this claim, and is that stated in the text rather than only in a byline graphic?

Section 4: Current state — are you being cited?

This is the measurement section, and it needs to be run under controlled conditions or it’s not evidence.

1. Assemble 30 to 40 buyer-realistic prompts from sales conversations, support questions and comparison scenarios. Write them as full questions. Version and freeze the list.
2. Test in clean sessions. Logged out or in a fresh profile, with no chat history, memory or custom instructions. Personalised results are not measurement.
3. Run each prompt three times across the assistants your buyers use, and record every run.
4. For each answer, record separately: was your brand named, was one of your URLs cited, which of your URLs, which other brands appeared, and which other domains were cited.
5. Compute the two headline rates. Citation rate = prompts where your URL appears / prompts tested. Mention rate = prompts where your brand appears in the text / prompts tested. Exclude prompts naming your brand from both.
6. Assess accuracy. For every answer mentioning you, mark the description accurate, partially accurate, or wrong. Misrepresentation at scale is a distinct problem from absence.

Section 5: Competitive and corroboration landscape

1. Tally the domains cited across all your test answers. The most frequent sources are the ones the assistants trust for your topics. That tally is your corroboration target list.
2. Tally the brands named. This is your real competitive set in AI answers, which often differs from the one your sales team assumes.
3. Identify which prompt types you lose. Compare your appearance rate on informational prompts against evaluative and comparison prompts. Different gaps require different fixes.
4. Check third-party accuracy. Where directories, review sites or professional listings describe you, confirm the information is current. Assistants draw on these, and stale entries propagate.

Turning the audit into a plan

Sort every finding into three buckets, and resist the temptation to start with the interesting work rather than the blocking work.

• Blockers — anything in Section 1 that prevents retrieval. These come first regardless of effort, because nothing downstream can work until they’re cleared.
• Structural fixes — extraction and attribution problems on your highest-value pages. These are within your direct control and typically show movement within one or two measurement cycles.
• Long-horizon work — corroboration, third-party presence, and new content for question areas where you have nothing. Necessary, but measured in quarters.

Whatever you fix, keep the prompt set and protocol from Section 4 unchanged and re-run monthly. The audit only has value if it becomes a baseline rather than a one-off document. Our ARIA citation tracker maintains that loop, and our AI visibility practice covers the programme that follows. If you would rather have the audit run for you, start a project.

Frequently asked questions

How long does an AI visibility audit take?

The access and extraction sections take a few hours. The measurement section is the bulk of the effort — around a day for 30 to 40 prompts run three times across two or three assistants, depending on how much you automate the recording.

What is the most common blocker an audit finds?

Unintentional crawler blocking, usually inherited from a default configuration or a firewall rule set by a different team. It’s also the fastest thing to fix once identified.

Do I need special tools to run this audit?

No. A browser in a clean session, access to your server logs and robots.txt, and a spreadsheet cover the whole checklist. Tooling saves time at scale but does not change the method.

How often should the audit be repeated?

Run the full audit annually or after a significant site change. Run the Section 4 measurement loop monthly, using the identical prompt set, so you have a trend rather than isolated snapshots.