AEO and AI visibilityHow to Run an AI Visibility Audit in One Day

An AI visibility audit is a structured test of whether AI answer engines mention, describe, and cite your brand for the questions your buyers ask. It also checks whether your site gives those engines a fair chance to, and it ends in a baseline plus a prioritized fix list.

In The Agentic AI Marketing Leader Playbook I treat it as one workflow with two halves. The tracking half asks whether engines cite you today, and the audit half, which many teams call an AEO audit, covers the AEO and GEO layers together because the fixes overlap.

Others call it an LLM visibility audit or an AI search audit, and my comparison of AEO, GEO, and SEO separates the terms.

A complete audit checks five layers, in the order they usually break. When an early layer fails, the later ones do not matter yet, so this sequence doubles as my AEO audit checklist.

1. Reachability. Can AI search crawlers fetch your pages, or does robots.txt, a firewall, or a CDN rule turn them away? Pass when priority pages are indexed and crawlers get a normal response.
2. Extractability. Does the answer exist in the raw HTML, or only after JavaScript runs? Pass when the answer sentence shows in view-source.
3. Citation presence. Do the engines name you, link you, or name a competitor instead? Pass when you appear on flagship prompts run after run.
4. Accuracy and framing. Does an engine get your category and offer right? Pass when brand prompts score 2.
5. Structure. Schema, answer-first headings, bylines, dates, and internal links. Pass when priority pages score 2 on every factor and pass the page-alone test.

Accuracy is the layer most audits skip, and a confident wrong description of your offer reaches buyers you never get to correct.

You can run a first-pass AI visibility audit in one working day with a spreadsheet, five engines, and twenty prompts. Each block of the day below maps to the nine steps that follow.

1. Hour 1. Design the prompt set.
2. Hours 2 to 4. Run the sampling passes and log every answer.
3. Hour 5. Score the sheet and build the citation gap map.
4. Hour 6. Check crawler access and JavaScript rendering.
5. Hour 7. Score page structure, run the page-alone test, and check how third-party sources describe you.
6. Hour 8. Pull first-party data and write the findings-to-fixes table.

Step 1: Design a Prompt Set Buyers Would Actually Type

Start with twenty prompts written in buyer language, never in the language of your keyword list. Pull them from sales call notes, demo request forms, and the People Also Ask boxes on your head terms.

Spread the twenty across five intents so one type of question cannot dominate the score.

• Problem prompts. The buyer has a pain and no vendor in mind, such as "How do I cut employee onboarding time?"
• Category prompts. The buyer wants options, such as "What are the best onboarding platforms for mid-market companies?"
• Comparison prompts. Your top competitor against you, by name.
• Shortlist prompts. "Which vendors should I consider for this, and why?"
• Brand prompts. "What does [your company] do, and who is it for?" These test accuracy rather than presence.

Mark three to five prompts as your flagship set, the topics the business has staked its authority on. Then pick one prompt where you already win and label it the canary, retested every cycle purely to catch regression.

Before the first run, write the scope on one page. List every name buyers use for you, the competitors you expect, the engines, and a test protocol for location, language, device, account state and engine mode, held fixed for the quarter because most apparent drops are protocol drift. My guide to measuring AI visibility adds the control prompts that separate a real change from noise.

Step 2: Choose the Engines and Run Sampling Passes

Test ChatGPT with search on, Perplexity, Gemini, Microsoft Copilot, and Claude, then add Google AI Overviews for any prompt that triggers one. Run every prompt at least three times per engine, each time in a fresh session with memory off.

Three runs is the floor because these systems rarely repeat themselves. SparkToro's January 2026 research on AI brand recommendations ran 12 prompts through ChatGPT, Claude, and Google's AI a combined 2,961 times. ChatGPT and Google's AI had less than a 1 in 100 chance of repeating the same brand list.

Turn web search on wherever offered, because default modes answer from memory and cite nothing. Keep each engine separate too, since averaging a Copilot win with a Perplexity loss hides both.

Step 3: Score Each Answer on a Simple Sheet

Log every run as one row on the same columns every time, because the trend across cycles is the real output. The rows below illustrate the format.

From the sheet, compute three numbers per engine. Mention rate is mentions divided by runs, citation rate is linked citations divided by runs, and accuracy is the average score on brand prompts.

Never count a mention without a link as a citation, since the two point to different fixes. The "Sources cited" column becomes your citation gap map, and every third-party page that keeps appearing is one you need to be on or outwrite.

Step 4: Check Robots.txt for Every AI Crawler That Matters

Open yoursite.com/robots.txt and look for each user agent in the table. A frequent mistake is one blanket rule that blocks a training crawler and a search crawler together, when the two control different things.

Each row comes from the owner's own documentation. See OpenAI's crawler overview, Perplexity's bot guide, Anthropic's crawler article, and Google's common crawlers list.

OpenAI also notes that a robots.txt change can take about 24 hours to reach ChatGPT search, so do not retest the same afternoon. Since July 2025, Cloudflare asks every new domain whether to allow AI crawlers at sign-up, so have the CDN owner confirm in writing that the CDN does not block bots your robots.txt invites.

An engine cannot cite a page that was never indexed. Google lists nosnippet, data-nosnippet, max-snippet and noindex as the controls that limit what Search shows from your pages, so hunt for stray directives, then check server logs to confirm the search crawlers actually visit.

At a seed-stage agentic AI company where I am fractional CMO, only 19 of 172 URLs were indexed, mostly because of 114 dead legacy URLs. Rebuilding the sitemap, restoring the robots.txt sitemap line, wiring IndexNow and redirecting retired pages took Ahrefs site health from 26 to 100 in a day.

Step 5: See What a Crawler Sees Without JavaScript

Fetch your top pages the way a non-rendering crawler would. Run curl -A "OAI-SearchBot" https://yoursite.com/page/ or open view-source, then search the raw HTML for your main answer sentence.

Vercel's December 2024 analysis of AI crawler traffic found that none of the major AI crawlers rendered JavaScript, including GPTBot, ClaudeBot, and PerplexityBot. Gemini was the exception, since it uses Googlebot's rendering, so a script-loaded answer likely stays invisible to the others.

Step 6: Score Page Structure and Schema From 0 to 2

Take your ten most important pages and score each on five factors, with 0 for missing, 1 for partial, and 2 for done. Mark a factor N/A when it does not apply and leave it out of the denominator, then average across pages for a GEO readiness score you track over time.

• Schema. Valid Organization, Article, and FAQ or Product markup where it fits, tested in Google's Rich Results Test.
• Direct claims. Major claims stated as plain sentences a model can lift whole.
• Question headings. Headings shaped like questions, with the answer immediately beneath.
• Byline and dates. A visible author and a real published and updated date.
• Internal links. Links that make the topic cluster legible to a machine.

Keep schema in proportion. Google's guidance on AI features and your website says pages need to be indexed and snippet-eligible, with no special schema or AI text files required, though markup should match the visible text. I treat markup as disambiguation rather than a lever that earns citations on its own.

Step 7: Run the Page-Alone Test

This is the test I trust most. Write down the required facts a correct answer must contain, then paste a page into an LLM with one flagship question and ask it to answer using only that page.

Using only the page below, answer the question that follows. If the page does not contain the answer, reply NOT STATED. Quote the exact sentence you relied on.

Where the model cannot answer, or invents something the page never said, log the failure as missing content, ambiguous content or a model error. Missing and ambiguous content get fixed first, because a page can score 2 on every factor and still only imply its answer.

Step 8: Check Entity Consistency and the Sources Engines Trust

A 2025 arXiv study of how AI search engines choose sources found a systematic bias toward earned media, meaning third-party, authoritative sources, over brand-owned and social content.

Open the third-party pages that recur in your Sources cited column, and check whether each one names you and describes your category the way your About page does. Then confirm your Organization markup, which Google says helps disambiguate your organization in search results, with sameAs links to those profiles.

Inconsistency can start at home. At the seed-stage company, recorded demo calls showed even the product category changing from call to call, so agree on one category sentence first. My AEO strategy guide covers the earned mention work that follows.

Step 9: Pull First-Party Citation Data

Close the day with the data the engines publish about you. In February 2026, Bing Webmaster Tools added an AI Performance report showing citations across Copilot, Bing's AI summaries, and select partners, plus the grounding queries I feed straight back into the prompt set.

Google followed in June 2026 with generative AI performance reports in Search Console, covering AI Overviews and AI Mode, and they reached all sites on August 31, 2026. Neither report covers ChatGPT or Perplexity, which is why manual sampling stays in every AI visibility audit.

Read AI visibility audit results as patterns across layers, never as one score, and fix in layer order, because a rewrite cannot help a page a crawler never reaches.

The output of an AI visibility audit is a findings-to-fixes table where every row names the evidence, the fix, the owner, and the effort. A score without that table is a report nobody acts on.

The rows below are illustrative, built from the failure patterns this method is designed to catch.

Rank fixes cheapest first, and pair each with the CRITICAL prompts it would help, where a competitor is cited and you are not. A prompt you lose that also ranks worse than tenth in Google is a dual loss and goes to the top, and the AI citation tracker I built flags those pages for me.

Restructure a thin page before you write a new one, and hand each fix to its owner the same day. For the content rows, my guide on how to get cited by ChatGPT covers the page-level rewrites.

You know the AI visibility audit is working when mention and citation rates on the flagship set trend up across cycles while the canary holds. The second sign is structure score gains lining up with citation gains.

Expect the absolute numbers to start small. At the data and AI services company where I am CMO, 677 people arrived straight from ChatGPT, Perplexity, Gemini and Copilot answers in eighteen months, 0.11% of traffic. Small is the point, because cited sources get retrieved again and trust compounds.

At an HR tech SaaS company, we built titles, FAQs, glossary terms, and schema so answer engines could lift them, with schema on more than 116 blog posts. AI-sourced sessions grew from 335 to 3,200 a month, up 854%, and the HR tech case study walks through the sequence.

How Often Should You Repeat an AI Search Audit?

Run the citation check weekly on the flagship set and the full structural audit every quarter. That cadence comes straight from the workflow in my book, and it stops the audit from becoming a one-time screenshot.

Retest after any major site change too, such as a migration, a new CDN, or a redesign. Those are the moments a blanket bot rule or a script-rendered template slips in without anyone noticing.

Run it yourself if you have a single site, fewer than fifty priority pages, and someone who can read robots.txt and a spreadsheet. Bring in help when the fixes cross web, engineering, and content teams and nobody owns the outcome.

A JavaScript-heavy site, several brands or languages, or no owner for robots.txt, CDN and schema are the usual reasons to get help.

A free AI visibility checker covers the reachability layer and a quick page grade, and my guide to AI visibility tools compares what paid platforms track. No tool can tell you which buyer prompts matter, and if you are weighing outside firms, start with my questions for choosing a GEO agency.

Who Should Own the AI Visibility Audit?

The SEO function should own it, next to search visibility reporting, rather than a new AI team nobody has built yet. Assign it this quarter, even informally.

An AI visibility audit starts decaying the week after you run it, so turn it into a ritual. I call it the Monday Citation Muster, after the daily muster I run with my team.

Every Monday, run the flagship prompts three times on each engine and retest the canary. Log mentions, citations, and new sources in the same sheet, then write one line each for Achieved, Pending, and Blocking.

Fix the cheapest blocker before Friday. Do that for one quarter and you will hold a trend line instead of a screenshot. If you want the first audit run with you and the fixes built, see my AEO and GEO services.