What an AI visibility audit should contain

Someone on your leadership team ran your product category through ChatGPT, noticed your company was omitted, and sent you a screenshot. Within hours, a vendor pitched you an AI search audit, or a consultant delivered a PDF full of scores. Before paying that invoice or changing your marketing strategy, you need to know how to separate legitimate technical measurement from automated template output.

Most AI audits sold today are traditional search engine optimization audits with new branding. They run a domain through a scraper, export a list of keywords, and slap an arbitrary visibility score on the cover.

A real AI visibility audit is an engineering assessment of how generative models retrieve, process, and cite information about your category. If an audit report lacks specific technical checks, raw prompt lists, and statistical error margins, it is not a measurement. It is an opinion piece formatted as a spreadsheet.

Technical Access Checked Agent by Agent

The most common technical error in commercial AI audits is treating all web crawlers as the same entity. An audit that tells you whether you are blocking AI without naming specific user agents is useless.

Web crawlers operated by AI companies split into two distinct functional categories: training crawlers and answer crawlers.

Training crawlers collect data to build future model weights. Blocking a training crawler stops an AI company from using your content to train next year's model. Answer crawlers, also known as search or retrieval crawlers, fetch live web pages to answer a prompt right now. Blocking an answer crawler immediately prevents an engine from citing your site in live retrieval responses.

In our crawler index evaluated on 2026-08-08, we analyzed a panel of 54,082 domains. Out of 33,670 domains that returned a readable robots.txt file, 5,497 blocked at least one AI agent. However, only 2,576 of those 33,670 readable domains blocked at least one answer crawler.

If an auditor looks at your robots.txt file and tells you that you are blocking OpenAI, they must specify which agent you are blocking. Blocking GPTBot affects model training for future foundation models. Blocking ChatGPT-User or OAI-SearchBot breaks live search citations today.

Our panel data of 33,670 readable domains demonstrates how widely these configuration rules vary across the web:

• GPTBot (training): blocked by 5,080 of 33,670 domains.
• ClaudeBot (training): blocked by 4,603 of 33,670 domains.
• Google-Extended (training): blocked by 4,275 of 33,670 domains.
• PerplexityBot (search): blocked by 2,147 of 33,670 domains.
• ChatGPT-User (search): blocked by 2,074 of 33,670 domains.
• OAI-SearchBot (search): blocked by 1,579 of 33,670 domains.
• Perplexity-User (search): blocked by 1,402 of 33,670 domains.
• Claude-SearchBot (search): blocked by 1,400 of 33,670 domains.
• Claude-User (search): blocked by 1,384 of 33,670 domains.

This configuration confusion is present even among technology companies. In a census of 4,226 active Y Combinator companies conducted on 2026-08-07, 3,755 had readable robots.txt files. Among those 3,755 readable domains, 253 were blocking at least one AI crawler, frequently stopping answer crawlers by accident when trying to restrict training bots.

A useful audit must explicitly audit your server headers, edge configuration, and robots.txt file against each specific agent name.

Disclosure of the Prompt Basket

An audit cannot evaluate your visibility without disclosing every prompt used to test the models.

Traditional search tools track single keywords like "best CRM software." AI models do not work on single keywords. Buyers ask conversational, multi-step questions with explicit constraints, such as "What software should a 20-person B2B sales team use if they need native HubSpot integration and strict HIPAA compliance?"

Audit vendors frequently hide their prompt lists behind proprietary score formulas. They present a summary figure, like an AI Visibility Index of 65 out of 100, without revealing what queries produced that calculation.

A standard audit must disclose its prompt basket in full. The prompt basket must mirror actual buyer intent across three specific stages:

1. Category discovery queries where buyers ask for options without naming vendors.
2. Direct comparison queries where your brand is evaluated head-to-head against named competitors.
3. Feature constraint queries where specific integrations, compliance standards, or pricing models are evaluated.

If an audit report does not provide the exact text of every prompt run, you cannot reproduce the test, verify the claims, or monitor changes over time.

Repeat Runs and Statistical Confidence Intervals

Large language models are non-deterministic. If you run the exact same prompt through ChatGPT twice, you will often receive two different answers with different cited sources.

Most audit tools run each prompt once. They query an engine, record whether your brand appeared, and move on. Competitors run each prompt once because running queries repeatedly increases inference costs and compute time. Running a prompt once yields a random snapshot, not a baseline.

A single point score is meaningless. Reporting that your brand visibility is 34 provides no actionable signal because that number could be an artifact of model variance on a single run. A proper measurement must be expressed with an explicit confidence interval, such as 34, plus or minus 6, calculated within a 95% confidence interval.

To measure AI visibility accurately, an engine must query the same prompt multiple times across independent sessions. At Standing, our methodology uses a fixed prompt basket run five times per question per engine across five distinct platforms: ChatGPT, Perplexity, Gemini, Claude, and Google AI Overviews. We then apply a Wilson score interval to calculate the true upper and lower bounds of your visibility rate.

If an auditor presents a single definitive percentage or integer score without an error margin or sample count, they are presenting noise as statistical fact.

Source Citation Mapping Across Platforms

When an engine retrieves web results to construct an answer, it pulls from external sources. A complete audit must trace those retrieval sources line by line across all five major engines: ChatGPT, Perplexity, Gemini, Claude, and Google AI Overviews.

Models do not rely solely on your main website to evaluate your company. They retrieve content from third-party review platforms, industry press, community discussions, software directories, and technical documentation.

A comprehensive audit must map every external domain cited when your category is queried. This source analysis reveals why an engine favors a competitor. The model may not be ignoring your site because of poor technical setup; it may simply be retrieving data from three industry review sites where your competitor has 200 reviews and you have four.

Knowing which domains supply context to answer crawlers allows marketing teams to prioritize external channels that directly influence live answers.

Separating Measurement From Judgement

An audit should clearly separate hard operational measurements from editorial advice.

Measurement consists of reproducible facts:

• Which answer crawlers are blocked by your robots.txt or web application firewall.
• Your brand presence interval across a fixed prompt basket, such as 42, plus or minus 5.
• The specific third-party URLs cited by models during live retrieval.

Judgement consists of human analysis:

• How your brand positioning compares to market rivals.
• What messaging adjustments might make your product features easier for models to categorize.
• Which external publications or review platforms require PR outreach.

Be clear about what not to buy. Do not pay for audits that guarantee specific engine rankings, mentions, or citations. Generative models are probabilistic systems. No vendor can guarantee that a specific model output will contain your company name.

Furthermore, do not buy audit services that claim adding schema markup or creating an llms.txt file will fix your visibility issues. We have no evidence that schema markup or llms.txt files improve AI citations. Recommending structural markup as a primary cure for missing AI citations is a sign of an agency applying old web tactics to a completely different distribution system.

A clean audit exposes crawler restrictions, tests prompts repeatedly to establish statistical confidence, names every cited retrieval source, and separates verifiable data from tactical advice.