What AI says about you right now: a brand audit in 30 minutes
Millions of people decide to buy (or not) based on what ChatGPT, Perplexity, and Gemini say about your brand, and you've never read that answer. This lesson gives you the script to audit that in half an hour: asking what AI says about your brand and your category, capturing the hallucinations, and classifying each one by severity (presence, positioning, sentiment, gaps). A wrong price becomes a lost sale, a false accusation becomes a lawsuit. And you turn this audit into a recurring ritual, the RESPOND moat.
A potential customer opens ChatGPT and types: "is it worth hiring [your company]? Give me the pros and cons." You're not in the room. You don't see the question, you don't see the answer. And AI answers as naturally as can be: it cites a price you no longer charge, invents a "weakness" that was never true, and at the end suggests a competitor. The customer reads, nods, and disappears. You never knew that conversation existed, nor that you lost that sale inside it. This lesson's audit is you finally sitting in that customer's chair and reading, with your own eyes, what the machine has been saying about you.
Someone searches "is [your consultancy] reliable? Did they have any problem with clients?" on Perplexity. AI, trying to be helpful, stitches together scraps from the internet and answers with a "fine" figure that never happened and a lawsuit that belonged to another company with a similar name. In finance, that's not a minor detail: it's your reputation for handling other people's money being described wrong, with confidence, to someone about to hire you. The audit catches this before it becomes every prospect's first impression.
A client searches "which is the best firm for a labor case in my city?" and Gemini lists three names. Yours isn't there. Worse: when they ask specifically about you, AI describes the firm as "focused on real estate law", which isn't your practice. Wrong positioning pushes away exactly the right client. And if AI invents an accusation (a case you lost, a sanction that doesn't exist), that's no longer marketing, it's defamation. The audit shows you both things in the same screenshot.
You ask ChatGPT "what are the best performance agencies for e-commerce?" and yours doesn't show up. Then you ask directly about your agency, and AI answers that you're "more focused on branding", when your strength is precisely performance. Zero presence on one question, swapped positioning on the other. Each of these answers is being read right now, by people deciding who to send the brief to. The audit is your way of reading along with them.
A great candidate, the kind you really want to attract, searches "what's it like to work at [your company]?" on ChatGPT before responding to your job posting. AI stitches together scraps from the internet and answers with a vibe that isn't yours: it cites a "high turnover rate" that doesn't exist and a "slow and confusing" hiring process you fixed a while ago. The candidate reads, shrugs, and doesn't even apply. You never knew you lost that talent in a conversation that happened far off your radar. This lesson's audit is you sitting in the candidate's chair and reading, with your own eyes, what the machine has been saying about your employer brand.
Someone evaluating tools asks Perplexity "what's the best product for roadmap management?" and yours doesn't show up on the list. Then they ask directly about your product, and AI describes a core feature from your PRD as "still in beta", when it's been in production for months, and states a plan price you don't even sell anymore. Zero presence on one question, wrong data on the other, and each one being read by whoever's deciding to buy right now. The audit is you reading that answer on purpose, before your next user reads it first.
A lead who'd easily enter your pipeline asks Gemini "is it worth buying from [your company]? Pros and cons." before even accepting your proposal. AI answers naturally: it invents an after-sales "weakness" that was never a real complaint and suggests a competitor at the end. The lead reads, cools off, and your salesperson never understands why the negotiation died before it began. No CRM logs that conversation, because it happened outside it. This lesson's audit is you finally reading what the machine tells people who haven't even entered your forecast yet.
A potential customer searches "does [your carrier] meet deadlines? Any SLA problems?" on ChatGPT before signing a contract. AI, trying to help, mixes scraps and answers with a delay rate that was never yours and a "history of lost shipments" that belonged to another company with a similar name. In operations, that's not a minor detail: it's your reputation for on-time delivery being described wrong, with confidence, to someone about to hire you. The audit catches this before it becomes every prospect's first impression.
Someone searches "is [your company] reliable with data? Any LGPD problems?" on Perplexity. AI, gathering whatever it finds on the internet, answers with a "regulator fine" that never happened and a "breach" that belonged to another company. Here that's not marketing, it's risk: your reputation for taking controls seriously being stained by a fabricated accusation, read by someone about to trust you with their data. The audit shows you this, and you screenshot the red flags with a date, to have proof if you ever need it.
A team evaluating vendors asks Gemini "what's the best deployment platform for small teams?" and yours doesn't show up. Then they ask directly about yours, and AI describes your architecture as "closed and hard to integrate", when you were born API-first, and invents a "security incident" that never happened. Swapped positioning pushes away the right team, and the false incident accusation is a whole other level of damage. The audit shows you both things in the same screenshot, before an architect reads it and cuts you from the shortlist.
A potential customer searches "is [your platform] easy to use?" on ChatGPT before requesting a demo. AI answers that your signup flow is "confusing and full of steps", based on an old version you redesigned months ago, and doesn't even mention the accessibility work that became your differentiator. The user reads, imagines a bad journey, and gives up without ever testing the real prototype. This lesson's audit is you sitting in that user's chair and reading what the machine says about your experience, before they read it first.
An investor doing due diligence asks ChatGPT "has this company ever had a problem with a previous investor? Is this round worth joining?" before the first meeting with you. AI, stitching together scraps from the internet, invents a "messy exit" for a co-founder who never existed and cites a previous round at a valuation that isn't yours. In the investment committee, that's not a minor detail: it's your strategic credibility being described wrong, with confidence, to someone deciding whether to bet on you. The audit catches this before it becomes every investor's first impression.
Let me start with a question that tends to give the good kind of chill. When was the last time you asked a ChatGPT, a Perplexity, a Gemini, what they say about your brand? Think with me: right now there's someone deciding to buy from you (or give up on you) inside a conversation with AI you've never read. You spend a fortune watching what Google shows, what Instagram shows, and there's this new, giant channel, talking about you all day long, and you've never opened the door to listen. This lesson fixes that in half an hour.
The core idea of this lesson. AI has become a discovery channel for your brand, the same way Google did in the 2000s. People ask it "is it worth it?", "who's the best?", "is this company reliable?", and it answers with absolute confidence, mixing what's true with what it invented. The brand audit is you sitting down and reading that answer, on purpose, with a script. You ask about your brand and your category across the three main engines, capture whatever comes up (including the hallucinations), and classify each finding along four axes: presence (do I show up?), positioning (what do I show up as?), sentiment (do they speak well or badly?), and gaps (what's missing or wrong?). And the trick: this isn't a one-time screenshot, it's a ritual you repeat. Later on this track we call it the RESPOND moat. The rule to stick in your head: what AI doesn't know how to say about you, it invents, and what it invents, someone reads as truth.
01Why this stopped being optional
Let's call it what it is. There's a data point circulating a lot in the market, from an analysis by Lily Ray on Google AI Overviews: on category prompts like "what's the best X?", around 69% of listicle answers end up recommending a competitor, not the reference brand. Not because the competitor is better. Often it's just because they show up more times, in more places, and AI learned to cite them. You could be losing the recommendation without ever having entered the fight.
And there's the other side, the one that's genuinely alarming. AI doesn't stay silent when it doesn't know. It fills the hole. If it doesn't have solid data about your price, it guesses one. If it doesn't quite know your positioning, it invents one. In regular marketing, a shallow text goes unnoticed. Not here: a wrong price in AI's mouth becomes a sale that died before it started, and a false accusation (an "this company had problems with x") turns, at the limit, into a lawsuit. There's already a real case (Walters v. OpenAI) of someone suing an AI company because the model invented a defamatory accusation, name and all. OpenAI even won on summary judgment, but the lawsuit, the cost, and the wear-and-tear were real. When the wrong thing is about your brand, the damage is yours first.
For the board. You don't control what AI says about you. But you can (and need to) at least READ what it says, before your customer reads it first. Whoever has never audited is driving with their eyes closed on a channel that already decides sales.
The good news is that reading this is cheap and fast. It doesn't need expensive tooling or an engineer. It needs half an hour, a script, and the willingness to face whatever shows up. Fair enough?
02The script: the questions you ask the machine
The audit has two families of questions, and you run both on every engine (ChatGPT, Perplexity, and Gemini, at minimum). Do it in a clean chat, without a login that already knows you and biases the answer, to see what a stranger would see.
Family 1, category questions (you don't even mention your name). Here you find out whether you exist in AI's head when the customer doesn't know you yet:
- "What are the best companies in [your category] for [your type of customer]?"
- "Who would you recommend for [the problem you solve]?"
- "I'm choosing between options in [category]. Which ones should I consider?"
Family 2, direct questions (you name yourself explicitly). Here you find out how AI describes you once someone has already heard of you:
- "What do you know about [your brand]?"
- "Is it worth hiring/buying from [your brand]? Pros and cons."
- "Is [your brand] reliable? Have there been any problems?"
- "How much does [your brand] cost?" (this one catches invented pricing the most)
- "Compare [your brand] with [a competitor]."
The difference matters. Family 1 measures presence: if you don't show up when the customer doesn't even know you, you're out of the discovery game. Family 2 measures everything else: how they describe you, whether they speak well of you, and where they lie. Back in the evals lesson (5.3, and its cousin G.6 in the Guardian track) you saw the idea of having a fixed set of questions you run the same way every time, to compare. This is exactly that, applied to your brand: a small library of prompts that becomes your recurring thermometer.
03The four axes: how to classify what comes up
Capturing the answer is half the job. The other half is reading with method, so you don't just walk away with "ah, they spoke okay". You sort each finding into one of four axes. Think of them as four lenses, and look at the same answer through each one.
- Presence. Did you show up? On category questions, did your name appear, or only competitors? There's even a market name for turning this into a number, they call it Share of Model: out of how many answers in your category your brand participates in. You don't need the fancy name, you need the count: out of ten category questions, in how many did I show up?
- Positioning. What did you show up as? Did AI describe you doing what you actually do, for the customer you actually want? Or did it file you in the wrong drawer ("focused on branding" when you're performance)? Swapped positioning very politely pushes away the right customer.
- Sentiment. Was the tone favorable, neutral, or against you? "A solid, reliable option" is different from "one option among many" which is different from "had some complaints". Sentiment is the answer's tilt, and it pushes the decision.
- Gaps. What was missing or what's wrong? Here live the two most dangerous things: wrong data (price, deadline, offer, a feature you don't even have) and false accusation (a problem, a sanction, an invented complaint). A gap isn't just what's missing; it's mainly what AI filled in with invention.
Notice these four lenses aren't equally weighted. Presence and positioning are a marketing game, you improve over time. Gaps, especially the false accusation, are a risk game, and that's where severity points to. Let's get to it.
04The severity scale: what's cosmetic and what becomes a lawsuit
Here's what separates this audit from a curiosity. Every finding you captured gets a severity score, because your time is short and not everything deserves action today. I use three levels, and the logic is simple: the closer to a lost sale or a lawsuit, the more severe.
- Green, cosmetic. AI described you in a less brilliant way than you'd like, but nothing false. "One option among several." Annoying, not urgent. Goes into the queue for improving presence and positioning, at your own pace.
- Yellow, lost sale. AI got something wrong that pushes the customer away: an outdated price, a wrong deadline, a feature you have and it says you don't, or it simply left you out of category recommendations. Here you're losing money invisibly. Real priority.
- Red, legal or reputational risk. AI invented an accusation, a sanction, a complaint, a lawsuit, something that isn't true and stains your name. This isn't marketing, it's damage. Someone already went to court (the Walters v. OpenAI case) over a defamatory sentence a model spat out about a person; the company won, but the wear and cost happened. Red demands immediate action, and sometimes demands documenting (screenshot with a date) to have proof.
For the board. The question that classifies severity isn't "does this bother me?". It's: "if a customer reads exactly this, what happens?". If the answer is "they give up on buying", it's yellow. If it's "they might tell others I'm problematic", it's red.
And here it's worth the honest stitching to what you already saw on the Guardian track. Back in the prompt injection lesson (G.2) and the audit one (G.3) you learned that AI can be pushed to say things through third-party planted content, and that checking the output isn't distrust, it's hygiene. This audit is the brand version of that: you're auditing what the model says, knowing part of it comes from sources you don't control. There's no way to stop AI from talking. There is a way to know what it says, and act on what's severe.
05From one-time scare to ritual: the RESPOND moat
Now the turn that transforms the lesson into an advantage. If you run this audit once, you get a scare and forget it. The real value is in making it a recurring ritual. AI's answers change: the model gets updated, a new piece of news enters the mix, a competitor publishes something, and suddenly what was green turns red. Auditing once is taking a photo; auditing always is having a thermometer.
The mechanism is the same as the prompt library you saw in evals: you fix your set of questions (the two families from section 02), and you run that set on a cadence (weekly if your brand is busy and exposed, monthly if it's calmer). Keep every round with a date. Then you don't just see today's state, you see the movement: did my presence go up or down? did that wrong price get fixed? did a new accusation show up? It's the difference between knowing it's raining and having a weather forecast.
In our track, this habit has a name: the RESPOND moat. Moat is the trench, the ditch that protects the castle, what gives you an advantage the competitor can't easily copy. While your competitor has never read what AI says about them, you read it every week, fix what's severe, and stay one step ahead of that invisible conversation. It's not one more task; it's a system that keeps watching your name while you run the rest of the business. Fair enough?
Do it now
Your mission is to run the full brand audit and come out with ONE artifact: your AI Brand Audit Sheet, first dated round. Half an hour, on the clock.
STEP 1 · BUILD THE SHEET (5 min) Open a simple spreadsheet with these columns: ENGINE (ChatGPT/Perplexity/Gemini) | QUESTION | WHAT AI ANSWERED (paste the excerpt) | AXIS (presence/position/sentiment/gap) | SEVERITY (green/yellow/red) | ACTION.
STEP 2 · RUN THE CATEGORY QUESTIONS (8 min) In a clean chat, no login that knows you, across the three engines, ask without citing your name:
- "What are the best [your category] for [your customer]?"
- "Who would you recommend for [the problem you solve]?"
Note: did you show up? in what position? which competitors came ahead of you?
STEP 3 · RUN THE DIRECT QUESTIONS (10 min) Across the three engines, now naming yourself explicitly:
- "What do you know about [your brand]?"
- "Is it worth buying/hiring [your brand]? Pros and cons."
- "How much does [your brand] cost?"
- "Is [your brand] reliable? Have there been any problems?"
Paste each answer. Hunt especially for: wrong price, a feature you have that it denies, and any accusation or problem it cites.
STEP 4 · CLASSIFY (5 min) For each row, mark the AXIS and the SEVERITY. Remember the key question: "if a customer reads this, what happens?". Gives up on buying = yellow. Could stain my name = red. Screenshot the red ones with a date, to have proof.
STEP 5 · SCHEDULE THE NEXT ONE (2 min) At the end, write at the top of the sheet: "Next round: [date]". Weekly if your brand is exposed, monthly if it's calm. That's what turns the scare into a ritual, the RESPOND moat.
Closing test question: looking at your sheet, what's the MOST RED FINDING? If one exists, it just became your most urgent marketing task of the week, and you only found out because you sat down to read. Whoever hasn't audited doesn't know it exists.
Where "Share of Model" comes from and why the GEO jargon matters to you
The language the market uses for this has a technical name, and it's worth recognizing so you don't get caught off guard in a sales pitch. When people talk about measuring your PRESENCE in AI, the trending term is Share of Model (SoM): the percentage of your category's answers in which AI cites or recommends your brand. It's cited as the KPI replacing the old advertising "share of voice", now that discovery is migrating from Google to AI. The whole discipline of getting AI to cite you is called GEO (Generative Engine Optimization), and the summary the market repeats is honest: "SEO gets you clicked, GEO gets you cited". There are paid tools that automate this audit (Semrush, Profound, Brandlight, and others), running hundreds of prompts for you and building a visibility index. Don't rush to hire any of that yet. What those tools do at scale, you do by hand in half an hour, and doing it by hand is how you learn to READ, which is the part no tool does for you. Recognize the jargon when it shows up in a proposal, and demand that it delivers exactly the four axes from this lesson: presence, positioning, sentiment, and gaps.
Practice
1. You run the audit and ChatGPT answers that your company charges a price you haven't practiced in a year, and the customer will likely walk away thinking it's expensive. What severity does this fall under?
2. What's the difference between a category question and a direct question in the audit, and what does each measure?
3. Why is running the audit just once weak, and what turns it into the RESPOND moat?
Let's close the point together. There's a conversation happening right now, about you, inside ChatGPT, Perplexity, and Gemini, and until this lesson you'd never read a single line of it. The brand audit is simply you sitting in the customer's chair and reading: run the category questions to measure whether you show up, run the direct ones to see how you're described, classify each finding along the four axes (presence, positioning, sentiment, gaps), and mark severity by the right question, "if the customer reads this, do they buy, give up, or grow suspicious?". The wrong price that costs you a sale is yellow; the false accusation that stains your name is red, and sometimes you screenshot it with a date to have proof. And what separates whoever gets a one-time scare from whoever gains a real advantage is the ritual: fixed questions, run always, dated, the RESPOND moat. While your competitor keeps flying blind, you read every week what the machine has been saying, and fix it before it becomes a lost sale or a lawsuit. What AI doesn't know how to say about you, it invents. Starting today, you're the first one to read it.
Thanks for the feedback. It helps sharpen the next lesson.