How do I audit what AI says about my brand?

Ask a fixed set of questions in fresh sessions across ChatGPT, Perplexity and Gemini once a month, and record the answers verbatim. Cover four areas: category questions where you might be absent, name questions about what the model believes, comparison questions against your main competitor, and edge questions like pricing and status where confident wrong answers breed.

What questions should I ask ChatGPT about my company?

Start with the category, not your name: best tool for the job, five options for a small business, cheapest way to do the task. Your name only matters if it comes up when you are not the one saying it. Then ask what the brand is, whether it is legit, who is behind it, what it costs, and how it compares to the competitor you lose deals to.

What do I do if AI gives wrong information about my brand?

Fix the sources, not the model. Correct the claim on your own site first, then get the corrected version onto the third party pages the model reads, because it repeats whatever appears in the most places it trusts. Stale pricing and dead product rumours are the most common wrong answers, and both come from old pages that were never updated.

The Answer Audit: 15 Questions to Ask AI About Your Brand

¶ CONTEXTTHE MISSING HABIT

01The audit everyone runs on money and nobody runs on answers

Every serious business audits its books on a schedule. Nobody waits to wonder whether the accounts are fine. Yet the same owners have never once sat down and read, in full, what the machines tell a stranger who asks about them, and those answers are doing more selling than the homepage now.

You audit your books every quarter. Audit your answers every month.

Why monthly, not quarterly

This is not the same thing as googling yourself. Googling yourself shows you a list you can be on. These questions show you a verdict already written, delivered with total confidence to anyone who asks, with no second result to scroll to.

Because answers move faster than ledgers. A model refreshes what it believes about you every time its sources shift, a competitor publishes, or a forum thread gains weight. Quarterly is an eternity in a system that re-reads the web continuously. Every one of these answers is being given tonight, word for word, to somebody you will never meet.

¶ PLAYBOOKTHE PROTOCOL

02Fifteen questions, asked properly

The method matters as much as the list. Fresh session for every question, no chat history, because history contaminates the answer with what the model already knows you want to hear. Run the set in ChatGPT, Perplexity and Gemini, because they read different sources and disagree in useful ways. And run the name questions twice, once with browsing off to see the model's memory, once with browsing on to see what it fetches, a split that tells you whether you are established or merely current.

How to run it without fooling yourself

Read every answer as a stranger, not as the founder. The founder hears what was meant. The stranger hears what was said. Write the answers down verbatim, every month, in the same file. A trend you did not record is a feeling.

The fifteen

Swap the brackets for your real category, brand, competitor and audience. The wording is deliberately plain because plain is how customers actually ask.

THE CATEGORY~~DOES THE MODEL REACH FOR YOU UNPROMPTED~~

01“What is the best [category] tool for [job]?”The only question most buyers ever ask. You are listening for whether you exist in the answer at all.

02“List five [category] options for a small business.”Longer lists have more room. If you are absent even at five, absence is the finding.

03“What is the cheapest way to [job]?”Price questions surface different names than quality questions. Free products live or die here.

04“Who makes the best [category] right now?”The phrase right now invites the browsing path. Compare this answer to question one.

THE NAME~~WHAT IT BELIEVES WHEN YOU COME UP~~

05“What is [BRAND]?”The one sentence it gives back is your real elevator pitch. If it hedges, so does every recommendation.

06“Is [BRAND] legit and safe to use?”The trust question people ask in private. A single wrong forum thread can own this answer.

07“Who is behind [BRAND]?”Checks whether your entity resolved. Wrong founder, wrong country, wrong company means your sources disagree.

08“What does [BRAND] cost?”Stale pricing is the most common wrong answer in the whole audit, and the most expensive one.

THE FIGHT~~HOW COMPARISONS TREAT YOU~~

09“[BRAND] vs [competitor]: which should I choose?”Listen for the deciding sentence. Whatever it names as your weakness is what your sources emphasise.

10“What are the best alternatives to [competitor]?”The highest intent question in the set. Being absent here is losing customers you never saw.

11“Why would someone choose [BRAND] over [competitor]?”The model argues your case with whatever material it has. Thin material makes a thin argument.

12“What are the downsides of [BRAND]?”You want criticisms that are true and current. Outdated complaints mean stale sources are still winning.

THE EDGES~~WHERE WRONG ANSWERS BREED~~

13“Is [BRAND] free?”Yes-or-no questions get confident answers even when wrong. Confidently wrong here costs signups tonight.

14“Does [BRAND] work for [audience]?”Pick the audience you quietly serve well. Models often do not know your edges exist.

15“What happened to [BRAND]?”The freshness probe. Small brands get declared dead, acquired or renamed on years-old evidence.

EXTRACTED — THE SENTENCE THIS ENTRY EXISTS FOR

Your books tell you what happened. The answers decide what happens next.

¶ PLAYBOOKSCORING AND ACTING

03Three buckets, three moves

Every answer you record lands in one of three buckets, and each bucket has exactly one correct response. Resist the urge to fix everything at once. The audit is sorted by cost, and the expensive bucket is rarely the one that stings the most to read.

Absent. The model answered the category question without you. This is a describability and evidence problem, not a fairness problem. The move is the slow one: one clear sentence about what you are, repeated everywhere, and mentions earned on pages you do not own.

Wrong. The model named you and got the facts wrong. Price, founder, features, status. The move is source repair: fix the claim on your own site first, then in the places the model actually reads. You cannot argue with the model. You can only outnumber its bad sources.

Recommended. The model names you, describes you correctly, and argues your case. The move here is the one people skip: protect it. A right answer is not a finish line. It is a lease that renews only while your sources keep agreeing.

What the first audit usually finds

Mine found all three buckets at once. Recommended in one category, absent in another I thought I owned, and one answer so confidently wrong about a product's name that I rebuilt a homepage over it. Expect the same spread. The point of the audit is not a grade, it is a work list sorted by what is costing you tonight. And the second month is where it starts paying, because the diff between two audits shows you which of your fixes actually reached the models and which ones are still sitting in sources nothing reads.

When to stop doing it by hand

The first two audits should be manual, because reading the answers yourself is how the reframe lands. After that it is rote work on a schedule, which is exactly what software is for, and it is why I eventually built a tool that runs the set monthly and diffs the answers against last month. Build one, buy one, or keep a calendar reminder and a text file. The tooling is optional. The habit is not.

Tomorrow: why free is not a price, and what it buys that money cannot.