AI Brand Audit: How to Check What AI Says About Your Company

An AI brand audit answers one question: when a buyer asks an AI engine about your category or your company, what does it say, and is it right? It is not a ranking report. Google's own AI reporting in Search Console exposes impressions only, with no position data, so no tool can check an "AI rank" against ground truth.

What you can measure is the content of the answers. That means three things:

• Presence: does the engine name you for questions a buyer would ask?
• Accuracy: are the facts it states about you correct and current?
• Sourcing: when it cites pages, are they yours, a third party's, or none?

If you have never looked, start with a quick read of what AI says about your company. The audit below turns that first look into something you can repeat and compare month to month.

An AI answer passes through four stages. Knowing which stage produced an error tells you whether you can fix it.

1. Parametric: the model answers from what it learned in training, with no web access. That knowledge is frozen at a cutoff. OpenAI's model documentation, as verified in August 2026, lists a February 16, 2026 cutoff for the GPT-5.6 family and August 31, 2025 for the Instant path ChatGPT also uses. "ChatGPT" is not one knowledge state.
2. Access: a crawler either can or cannot fetch and store your page.
3. Retrieval: for a given question, the engine decides which pages enter its context.
4. Generation: the model picks which retrieved pages to use and cite.

Retrieval is discretionary on every major platform. ChatGPT decides per question whether to search. On Claude.ai, web search is a user toggle. Google AI Overviews often don't trigger at all. So two identical questions can produce one answer from training data and another from live pages.

The question the engine searches is also not the one you typed. OpenAI says ChatGPT search typically rewrites your query into one or more targeted queries, and saved memories can shape that rewrite. Google states that AI Overviews and AI Mode may use different models and techniques. Treat each surface as a separate product.

Citations add one more layer of noise. OpenAI's API documentation says the number of sources a model consulted is often greater than the number it cited. A missing citation does not prove your page was ignored.

Be plain with yourself and your team about this split before you start.

In your control

• Crawler permission per bot. OpenAI documents its bots as independent: GPTBot governs training use, OAI-SearchBot governs ChatGPT search. Blocking one does not block the other.
• Network reachability. OpenAI's help centre says eligibility requires allowing OAI-SearchBot and confirming your host or CDN allows its published IP addresses. A firewall can turn away a crawler that robots.txt allows.
• Rendering. Vercel's crawler study found ChatGPT and Claude crawlers fetch JavaScript files but don't execute them. A price that appears only after JavaScript runs may be invisible to them. Googlebot renders, so this mostly matters outside Google.
• Index and snippet eligibility. Google says a page must be indexed and eligible for a snippet to appear as a supporting link in AI Overviews or AI Mode. noindex, nosnippet, max-snippet and data-nosnippet keep content out.
• Consistency of your own facts. If your pricing page and your footer disagree, an engine can quote either.

Not in your control

• What a model learned before its cutoff.
• Whether an engine searches for a given question, and what query it runs.
• Which retrieved page the model chooses to cite.

The controllable items are mostly about removing blocks. No markup or writing style has credible evidence behind it as a way to get cited.

1. Write down your true facts first

You can't score an answer without a reference. List the facts an engine is likely to state: what you do in one line, pricing tiers, founding year, headquarters, product names, key integrations, ownership. A brand fact sheet template gives you a structure. Note where each fact lives on your site.

2. Fix a prompt set

Write 10 to 20 questions and keep them frozen. Mix three kinds:

• Category questions a buyer asks before they know you ("best invoicing tool for small agencies").
• Brand questions ("what does [your company] do", "how much does [your company] cost").
• Comparison questions ("[your company] vs [competitor]").

Say you run a dental clinic in Austin. Your set might include "emergency dentist open Saturday in Austin" and "does [clinic] accept [insurer]".

3. Fix the test conditions

Run logged out, or with memory and custom instructions off, so personal context doesn't change the answer. The differences are covered in logged-in vs logged-out AI testing. Record the date, the surface and the model name where the app shows it.

4. Run each question twice where you can

Once with web search on, once with it off. In the consumer apps you often can't force this, so record whether the answer showed citations. No visible citations means you don't know whether it searched. Through the developer APIs you can set search on or off explicitly, which gives a cleaner comparison.

5. Log every brand statement

For each answer, record: were you named, each fact stated about you, whether it matches your fact sheet, and which URLs were cited. One row per statement makes the next step easy.

An error log is only useful if each entry points to something to check. Use a small, fixed set of types.

Keep "unverifiable" separate from "wrong". An engine stating a number you never published is a different problem from an engine contradicting a number you did publish. The first needs a decision. The second needs a fix.

Weight errors by consequence, not just count. A wrong price on a buying question matters more than a slightly off founding year. For a fuller treatment of root causes, see AI hallucinations about brands.

Work from the gates that block everything to the details that affect one answer.

1. Access. Check robots.txt for each retrieval bot: OAI-SearchBot, PerplexityBot, Claude-SearchBot, Googlebot. Then check that your CDN or firewall actually lets them through. Robots.txt saying "allow" while the firewall returns 403 is a silent failure. 99Visibility's AI Crawler Checker requests a page as each AI crawler and shows which ones get turned away.
2. Rendering. Load your pricing, product and about pages with JavaScript disabled. Any fact that disappears needs to be in the server-rendered HTML.
3. Suppression. Search your templates for noindex, nosnippet, max-snippet:0 and data-nosnippet. These often get inherited from a template by accident.
4. Contradictions. Make each fact say the same thing everywhere on your site. Retire or redirect pages with old prices and old product names.
5. Clear statements. State each key fact once, plainly, on a page that can be fetched. A specific sentence (a price, a number, a one-line definition) gives the engine something it can quote. Your About page is the natural home for most of these.
6. Third-party pages. Where engines cite directories, review sites or comparison pages, correct your listing there. Being named accurately on the sites engines already cite is associated with appearing in answers, though nobody can promise the effect.

Then rerun the same prompt set under the same conditions. Answers given with search off won't move until a model trained after your fix is released.

• Changing prompts between runs. A new prompt set makes month-on-month comparison meaningless. Add new questions as a separate group.
• Testing logged in. Memory and custom instructions change the query the engine runs. Your result may not match what a buyer sees.
• Blending every engine into one number. Each surface fails differently. Google AI Overviews depend on Google's index, so audit them on their own, as covered in checking AI Overview accuracy. A single blended score also averages answers given with and without live search.
• Reading a missing citation as rejection. Engines retrieve more pages than they cite.
• Reaching for markup. Google says structured data isn't required for generative AI search, and Ahrefs' matched test found AI Overview citations fell 4.6% after pages added schema. llms.txt is ignored by Google Search and no engine documents reading it. Existing markup does no harm, but it won't fix a wrong answer.
• Allowing GPTBot to fix ChatGPT search. GPTBot governs training. OAI-SearchBot governs search.
• Trying to steer answers with hidden text. Prompt injection, hidden instructions and cloaking for AI crawlers are manipulation. They risk penalties and reputational damage. Audit third-party pages that mention you for this kind of content, because it can misrepresent you too.