How to audit AI visibility: five layers, in the order that explains the result

An AI visibility audit is a one day check in five layers: can the engines reach you, are your pages eligible, is there a passage worth quoting, is your business one clear entity across the web, and what do the assistants actually say when a buyer asks. Run them in that order, because each layer is pointless until the one before it passes. This page gives you the checks in each layer, a scoring sheet, and the one page summary the audit should end with.

Why the order matters

Most audits start at the last layer, because asking ChatGPT a question is the fun part. That produces a finding with no cause attached. Work upward instead and every answer you get has a reason beside it. A page that cannot be fetched explains an absence completely, and no amount of writing fixes it.

Layer	The question	Time	Fix speed
1. Access	Can the named AI crawlers fetch my pages	30 minutes	Days
2. Eligibility	Are the pages indexed and snippet eligible	30 minutes	Days to weeks
3. Page	Does any single page answer one buyer question outright	90 minutes	Weeks
4. Entity	Does the web describe one business, consistently	60 minutes	Weeks to months
5. Answer	Who gets named when a buyer asks, and what is cited	90 minutes	Follows the four above

Layer 1: access

• Open robots.txt on every host you own, including blog, docs, help and shop subdomains. Rules apply only to the host serving the file, so each one needs its own.
• Look for OAI-SearchBot, GPTBot, ChatGPT-User, PerplexityBot, ClaudeBot, Claude-SearchBot and Google-Extended by name. OpenAI publishes its tokens at platform.openai.com/docs/bots and states that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers.
• Read a week of access log lines filtered to those agents. The file says what is permitted; the log says what happened. A bot management rule at the CDN can refuse an agent the file allows.
• View source on your three most important pages and search for your key sentence. If it only appears after JavaScript runs, treat it as a sentence the fetcher may never see.
• Check for a login wall or an interstitial on pages you want quoted.

The copyable robots blocks are in how to write a robots.txt, and the per crawler decisions in which AI crawlers should you allow.

Layer 2: eligibility

• Open the Pages report in Search Console and count indexed against not indexed, with the reason for each exclusion.
• Run URL Inspection on your five highest value pages and confirm Google can fetch and index each one.
• Check the robots.txt report in Search Console for parse errors and for the host list Google found.
• Confirm the pages you care about are snippet eligible, because Google's documentation on AI features states that a page must be indexed and eligible to be shown in Search with a snippet to appear as a supporting link, with no additional technical requirements.

Layer 3: the page

Take your ten buyer questions and ask, for each one, which single URL answers it. Then score that URL out of five.

• The title is the question, or close to it.
• The answer is in the first two or three sentences, with no preamble.
• The page answers one question rather than nine.
• There is at least one checkable fact in the answer: a count, a price, a date, a named source.
• Sub headings are written as questions, each with a self contained answer underneath, because retrieval lifts a passage rather than a page.

A question with no URL at all is your highest value finding of the day. The structure to use is in how to write a page an assistant will quote.

Layer 4: entity

• Search your business name and list every page on the first three pages of results that is not yours.
• Compare the name, category description, address, phone number and founding year across all of them. Log every difference.
• Count the sources outside your own domain that describe what you do. If that count is low, closing it is worth more than another page.
• Check your Organization or LocalBusiness structured data matches the visible text, since Google asks for exactly that.
• Look for a second business with a similar name in your market. Confusion between two entities is a specific failure with a specific fix.

Layer 5: the answer

• Use ten buyer questions in your buyers' words, with your company name in none of them. Naming yourself primes the answer, as set out in the blind prompt rule.
• Temporary chats, memory off, custom instructions off, consistent locale.
• Three runs per question per engine, across at least two engines.
• After each answer, ask the assistant whether it searched the web and which pages it read. That single follow up tells you whether you are looking at a fetched answer you can change this month or a training answer that moves over longer.
• Record four things per run: named or not, ordinal position, every company named, every source cited.

The scoring sheet and the one page summary

Score each layer out of ten and write one sentence of evidence per score. Then the summary page carries five things and nothing else: the mention rate per engine with a range, the layer that is holding you back, the three cited domains you are missing from, the two buyer questions with no page behind them, and the date of the next round. Everything else belongs in the appendix.

An audit is worth little as a one off. The layers above change at different speeds, and the answers themselves move week to week, so the value is in taking the same reading again and comparing. Keep the question set fixed, keep the brand name out of it, and date every run. AI Knows Us holds the set and the method steady between runs and stores each answer with its citations.

This page is the audit: what to check, in what order, and how to score it. For ongoing tracking after the audit, see AI visibility tracking. For cadence, see how often you should measure.

The audit AI visibility questions worth an answer

How long does an AI visibility audit take?

One focused day for a single business with up to a few hundred pages. The access and eligibility layers take an hour between them, the page and entity review takes half a day, and the prompt round takes 90 minutes.

Is a GEO audit the same thing?

Yes. GEO, AEO and AI visibility describe the same work under different names, and an audit for any of them covers these five layers. The terminology is sorted out in LLMO and how it relates to GEO, AEO and SEO.

How many prompts do I need for the audit to be meaningful?

Ten questions run three times each across two engines gives you 60 readings, which is enough to see a pattern and to produce a rate with a range around it. Fewer than that and you are reading individual answers rather than a rate.

What is the most common finding?

Two, usually together. There is no single page that answers a given buyer question, and almost nothing outside the company's own domain describes what it does. The first is a week of writing, the second is a quarter of outreach.

Should I audit all five engines?

Start with the two your buyers use and add the rest as the programme matures. Each engine is its own market and the answers differ, which is covered in why one AI app is not enough to measure with.

What to do first

• Block out one day and run the layers in order, writing the evidence sentence as you go.
• Fix every access finding the same week, since those are the cheapest and fastest items on the list.
• Write the ten buyer questions before you touch layer five, and keep your company name out of all of them.
• Produce the one page summary and give it a date. The date is what makes the next round comparable.
• Diarise the next round for the same days next month.