AI Visibility Audit: A 450-Answer Sheet and a Tool Rule

An AI visibility audit tests whether selected answer engines can access, understand, cite and accurately describe your brand for the prompts your buyers use. A search engine optimization (SEO) audit asks whether your pages can be found and ranked. The AI audit adds the answer itself.

An AI visibility audit is not a generic checklist and it is not a one-time chatbot query. It joins the site conditions we can inspect with the answers buyers actually receive, then asks why a brand was absent, inaccurately described, mentioned without a citation, or displaced by another source.

A traditional SEO audit still matters because it checks whether pages can be found, crawled, indexed, and ranked. This audit adds the answer layer: prompt-specific mentions, citations, sentiment (whether the answer's tone toward your brand is positive, neutral or negative), accuracy, source gaps, and the work needed to change them.

The engines in scope are ChatGPT, Perplexity, Gemini and two Google surfaces: AI Overviews, the AI summaries Google shows on Search results pages, and AI Mode, Google's conversational search tab.

Three terms appear in the table below. The robots.txt file sits at the root of a site and tells crawlers which paths they may fetch. A WAF (web application firewall) is an edge filter that can block a crawler even when robots.txt allows it. An entity is a clearly identified brand, product or person that a machine can tell apart from similar names. Rendered HTML is the page after JavaScript has run.

The important distinction is causality. A page can be crawlable and still never appear in an answer. A brand can be mentioned but described incorrectly. A source can be cited without contributing the recommendation a buyer sees. Our audit keeps those outcomes separate, so teams do not mistake technical eligibility for earned visibility.

The stakes are measurable on the click side. Pew Research Center followed 900 US adults' browsing in March 2025 and, across 68,879 Google searches, found that users clicked a traditional result link in 8% of visits when an AI summary appeared and 15% when it did not. When the answer absorbs the click, being named in it is what remains.

Run every prompt three times on each engine: 30 prompts × 5 engines × 3 runs is 450 answers per pass. Three runs is our recommendation, not a published threshold. Its job is to separate prompts where your brand appears reliably from prompts where it appears by chance.

The reason is variance. The SparkToro study used 12 prompts and 2,961 runs from 600 volunteers on ChatGPT, Claude and Google AI Overviews. It found less than a 1-in-1,000 chance that the same list comes back in the same order, and Claude was only slightly more consistent than the others. Our own research points the same way: in Citation Research Study 008, we re-asked 5 questions on four engines (20 answers) and brand lists moved by about a quarter on the second asking. That is a small sample, but it points the same way.

Two consequences follow:

• Do not record position. Order changes on almost every run, so "we were third" describes one draw. Record the share of runs in which you were mentioned.
• Read three runs coarsely. With three runs, a prompt scores 0, 1, 2 or 3 out of 3. Treat 3 of 3 as reliable, 1 or 2 of 3 as unstable and 0 of 3 as absent. A move from 0 of 3 to 3 of 3 after a fix is readable. A move from 2 of 3 to 3 of 3 is not, and you need many more runs to detect small changes.

Different trackers given identical prompts can also return different answers, as Digiday reported in May 2026. Whatever method you use, ask how many runs sit behind each number.

Usually yes: ChatGPT search uses OAI-SearchBot, Perplexity uses PerplexityBot, and Google's AI features use normal Googlebot access, with no separate crawler. Confirm each in about ten minutes, because a blocked page cannot be cited.

Review discovery, user-requested retrieval, and training controls separately. OpenAI's crawler documentation lists OAI-SearchBot for surfacing sites in ChatGPT search, GPTBot for training generative models, and ChatGPT-User for actions a person triggers, where robots.txt rules may not apply. Perplexity's documentation describes PerplexityBot as a search crawler not used for foundation-model training, and Perplexity-User as an agent that generally ignores robots.txt because a person asked for the fetch. Google states that no additional requirements apply to appear in AI Overviews or AI Mode beyond standard Search technical requirements.

The next table refers to a CDN (content delivery network), the service that sits in front of many sites and can filter crawlers, alongside the WAF defined above.

Read robots.txt by user agent and URL path, then record the rule that actually matches. The Robots Exclusion Protocol specifies that the most-specific matching rule wins, so a blanket verdict can conceal an allow or disallow rule on the exact page that matters. The robots standard also distinguishes unavailable responses from server failures, another reason to retain the raw evidence.

To print every group that applies to these bots, plus the User-agent: * group that covers any bot without its own, run this with GNU awk (gawk):

curl -s https://yourdomain.com/robots.txt | awk 'BEGIN{IGNORECASE=1} /^user-agent:/{p=($0 ~ /(\*|oai-searchbot|gptbot|chatgpt-user|perplexitybot)/)} p'

Then test your priority URL against the printed Allow and Disallow lines. Pass: the longest matching rule is an Allow, or no Disallow matches the URL. Fail: the longest matching rule is a Disallow. When an Allow and a Disallow match with equal length, the standard says the Allow is used.

Also inspect page-level robots directives, X-Robots-Tag headers, authentication gates, challenge pages, and soft errors. A clean robots.txt file cannot override a CDN rule that blocks a legitimate crawler, and a successful browser visit cannot prove what a crawler received.

Treat llms.txt, a proposed plain-text index of a site's key pages, as implementation status only. Google's AI features documentation states that you do not need new machine-readable files, AI text files or special structured-data markup to appear in AI Overviews or AI Mode.

Each row is one prompt, on one engine, on one day. The columns record whether you were mentioned, cited and described accurately, plus who else appeared. Our recommended 30-prompt mix is 8 category, 8 comparison, 8 recommendation and 6 use-case prompts, which covers the path from learning to deciding.

Write prompts in the wording buyers use, not your internal product names. Use the same location and signed-in state for every run, and spread the three runs over at least two days.

Row columns: prompt text, cluster, engine, run number, date, location, interface, full answer text, brands named, your domain cited (yes or no), cited domains, accuracy label.

Scoring rules:

• Mentioned: pass if your brand name appears in the answer text. Mention rate for a prompt and engine is the runs mentioned divided by 3.
• Cited: pass if your domain appears in the answer's source list.
• Accurate: pass if price, product description, location and key claims match your own pages. Label the answer correct, mixed or wrong. A cited link is a source to inspect, not proof that the answer is right, so accuracy is scored separately.
• Stable: 3 of 3 is reliable, 1 or 2 of 3 is unstable, 0 of 3 is absent.

Then roll rows up by cluster, per engine, never averaged across engines:

Engines draw on different kinds of sources, so one blended score hides the cause of a gap. In Citation Research Study 007, a community platform (Reddit, YouTube or Quora) was cited in 117 of 150 Google AI Mode answers and 3 of 150 ChatGPT answers.

Our research at Citation Research asked 150 everyday decision questions on four engines in October 2026 (one run each, logged out). The India-asked results for community-platform citations were:

The same study asked 50 questions from the US: 64 of 200 US-asked answers cited a community platform, against 61 of 200 from India. Community sources are not a feature of one market; the spread by engine is what to plan around. In Study 008, Perplexity also returned identical sources for 60 of 60 questions asked from India, the UK and the US, while Google AI Mode localized the most. Run your sheet from the location your buyers search in.

Our inference from this: if a competitor beats you on AI Mode but not on ChatGPT, check whether community threads and videos name them before you rewrite your own pages. If you are absent everywhere, look at your own pages and entity facts first.

Match each pattern on the sheet to a cause, then to an owner. A finding becomes actionable only when it names the evidence, the affected page or prompt, a responsible owner, a proposed change and a validation test.

Structured data helps clarify meaning when it is accurate and visible on the page, but it is not a citation trigger. Fix conflicting facts before adding markup.

Rerun the same prompts after a material change, with the same engine, locale, and collection conditions where possible. Compare what changed in the answer, citation set, and source gap, then record whether the result supports the original hypothesis. For month-over-month comparison, keep the prompt set fixed and track AI citations over time. To see whether visits follow, track AI traffic in GA4, Google's free web analytics tool. Google reports clicks from AI Overviews and AI Mode inside the Web search type in Search Console's Performance report, so they are not broken out on their own.

Do the first pass by hand, then move to a tracker when you will not repeat the 450-answer pass at least monthly. That threshold is our recommendation. The sheet's value comes from repetition, and one pass already means opening, copying and scoring 450 answers.

Our model: at an assumed 2 minutes per answer, 450 answers is 900 minutes, or 15 hours per pass. Repeating monthly means 15 hours a month. If your time costs $50 an hour (an assumption; use your own rate), that is $750 a month, against $299 a month on our Launch plan. Time your first ten answers and rescale: at 1 minute per answer the pass is 7.5 hours.

A hand audit is free apart from time, and it teaches you which prompts matter. It breaks down on repetition, because each retest after a fix costs another pass. Trackers, including ours, automate the collection. Whichever you consider, ask four things: which engines are covered, how many runs sit behind each score, whether the location is settable, and whether you can export the raw answers and cited sources. Our AI visibility tool trial checklist turns those into a 14-day test.

How we do it at PageLens.ai. We run the prompts you choose on a schedule, store each answer with the sources it cited, and show brand sentiment, the pages models quote and human and AI crawler traffic in one dashboard. As of October 2026, in US dollars with monthly billing, our plans are:

Annual billing saves 8%, and every plan has unlimited users. Our model of the sheet on Launch: 30 prompts × 3 engines is 90 answers per weekly check. Each weekly check uses 90 of Launch's 300 weekly answers, so three runs take three weeks (270 answers in total). Gemini needs our Growth plan or above, and Google AI Overviews still needs manual collection. See our pricing for the full plans.

If you want a first look before spending a day on the sheet, our free AI visibility audit takes your domain and returns a visibility score, what ChatGPT, Gemini and Perplexity say about you, and the competitors they name instead, in about 30 seconds, with no signup or credit card. To see this sheet running on your own prompts, Book a demo.