Your customers have started asking ChatGPT instead of Google. This prompt shows you — in about fifteen minutes — exactly what it says back when they ask about your category, and whether your name ever comes up. It’s the first thing we run on every GEO engagement.
There’s a question every business owner should have asked by now and almost none have: when a customer asks an AI assistant to recommend someone in my category, does it say my name?
It matters more than a ranking does. A Google results page gives ten options and lets the customer choose. An AI assistant gives two or three and moves on. There is no page two, no “below the fold”, and usually no second place. You’re either in the answer or you don’t exist in that conversation.
The reason nobody checks is that checking properly is harder than it looks. Ask ChatGPT “do you know [your business]?” and it will cheerfully say yes and describe you — sometimes accurately, sometimes inventing an entire company. That’s not an audit; that’s a leading question with a confident answer attached. The only honest test is to ask what your customers ask, in the way they ask it, without ever mentioning your name.
That’s what this prompt does. It’s the same one we run in the first hour of a GEO engagement, before we’ve promised anything, because it tells us — and the client — where they genuinely stand.
Most businesses that run this discover they’re invisible. That’s the normal result, not a catastrophic one — the models were trained on the open internet, and if you haven’t published anything that explains what you do and who you’re right for, there was nothing for them to learn. Invisible is a fixable starting point. Believing you’re visible when you aren’t is the expensive part.
Replace anything in [BRACKETS] with your own details. Everything else should stay exactly as written — the constraints at the bottom are doing more work than they look like they are.
You are a generative search analyst. Your job is to show me, honestly, how a business appears when a potential customer asks an AI assistant for a recommendation — and to diagnose why it does or doesn't get named. THE BUSINESS - Name: BUSINESS NAME - Website: URL - What we sell: ONE PLAIN SENTENCE — no marketing language - Where we sell it: CITY / REGION / COUNTRY - Who buys it: CUSTOMER TYPE, e.g. "operations managers at mid-size builders" - Our three biggest competitors: COMPETITOR A, COMPETITOR B, COMPETITOR C METHOD — follow these steps in order. Do not skip ahead or summarise. STEP 1 — Build the question set. Write 12 questions a real buyer would type into an AI assistant when they are ready to spend money but do not yet know BUSINESS NAME exists. Rules: - Never include our brand name in a question. That defeats the entire test. - Mix them: 4 "best / who should I use" questions, 4 problem-first questions (the customer describes a symptom, not a solution), 2 head-to-head comparisons, 2 with a real constraint (budget, deadline, location). - Use the words a customer would use, not the words our industry uses. Show me all 12 and stop. Wait for me to approve them. STEP 2 — Answer each question cold. Answer all 12 exactly as you would for a stranger who asked with no other context. Name specific real businesses where you would name them. Critical: if you do not actually know enough to name anyone for a question, say so plainly. Do NOT invent plausible-sounding businesses to fill out a list. I would much rather see "I don't know of anyone" than a confident guess. STEP 3 — Score it. Build a table with one row per question: | # | The question | Did you name us? | Who you named instead | What earned them the mention | STEP 4 — Diagnose, bluntly. For every question where we were not named, tell me which of these is true: (a) You have no knowledge of this business at all. (b) You know of us, but not in connection with this particular need. (c) We are genuinely the wrong fit and should not be named here. (d) A competitor has published something on this topic that we haven't. Do not soften it. STEP 5 — The fix list. Give me the five things most likely to get us named for these questions. Order by leverage, not by how easy they are. For each one state: - what specifically to publish or earn - where it needs to live - which of the 12 questions it targets - what would count as proof it worked CONSTRAINTS - Do not browse the web for this run unless I explicitly say so. I want your baseline knowledge first, because that is what most users see. - Do not flatter me. If we are invisible, say we are invisible in the first line. - Never assume a fact about our business that I have not given you. Ask instead. - If you are uncertain whether something is true, mark it uncertain.
The prompt is the easy half. Most bad audits we see aren't caused by a bad prompt — they're caused by running a good one in a contaminated chat.
New conversation, memory and personalisation switched off, and not logged into an account that has discussed your business before. If the assistant already knows you, it will name you — and you'll have proved nothing except that you told it about yourself earlier.
"What we sell" is where most people cheat. If you write "we deliver end-to-end integrated solutions", the assistant has learned nothing and the twelve questions will be useless. Write what a customer would say you do: "we fix commercial fridges in Adelaide, usually same-day."
This is the step everyone skips, and it's the one that decides the audit's quality. If a question contains jargon only your industry uses, replace it with the phrase a customer would type. You're testing reality, and your customers don't know your vocabulary.
Even instructed not to, models sometimes invent companies. Search each name before you react to it. Discovering that your "competitor" doesn't exist is not a failed audit — it's the most valuable finding available, and we explain why below.
These systems are probabilistic. One run is an anecdote. If you're absent across ChatGPT and Claude in separate sessions, that's a real finding. If you appear in three of four, you have partial presence, which is a different and much better problem.
The gap between the two runs is the whole story. Named only with browsing means your website is fine but your reputation hasn't spread beyond it — the model can find you when it looks, but doesn't know you when it doesn't. Absent in both means nothing on the open internet connects your name to what you sell.
Every constraint in it is there because a simpler version failed on real client work. Understanding why matters more than copying it — because then you can build the next one yourself.
A model has no reliable access to its own reasoning. Ask why, and it generates a plausible story — fluent, specific, and unrelated to what actually happened. So the prompt only ever measures behaviour: what did you say when asked what a customer asks. Behaviour is evidence. Self-explanation is fiction.
"Is [brand] good for X?" is a leading question, and these systems are trained to be agreeable. Mention yourself and you'll get a warm paragraph about how great you are — assembled from nothing. The customer asking about you already knows you exist. The one who matters doesn't.
Left alone, a model will fill a list of recommendations because a list is the expected shape of the answer — inventing businesses to complete the pattern. Explicitly authorising "I don't know of anyone" removes the pressure to perform, and turns a fabrication risk into a genuine signal about your category.
"Act as an SEO expert" changes almost nothing. "You are a generative search analyst, I am making budget decisions off this, a confident wrong answer costs me money" changes the output — because it sets a standard for what counts as a good answer, and that standard shapes everything that follows.
Most people run this with browsing on, see their name, and relax. But the assistant only found you because it went looking — which most users never trigger. Baseline knowledge is what the majority of your customers actually get, so that's what you must measure first. The browsing run is the control, not the test.
Option (c) — "we're genuinely the wrong fit here" — is the one clients hate and need. Without it, the model will invent a path to winning every question, and you'll spend a year of content budget chasing customers you don't want and can't serve. A finding you don't like is still a finding.
The brackets are yours. The constraints aren't — every one of them is load-bearing, and removing them is how this turns back into the flattering, useless version.
| Variable | What good looks like | What breaks it |
|---|---|---|
[WHAT WE SELL] | The sentence a happy customer would use. “We fix commercial fridges in Adelaide, usually same-day.” | Anything with “solutions”, “end-to-end”, or “bespoke”. The model learns nothing and generates twelve useless questions. |
[WHO BUYS IT] | A specific role with a specific problem. “Venue managers who’ve just had a fridge fail on a Friday.” | “Businesses” or “everyone”. Vague buyers produce vague questions, and vague questions can’t be won. |
[COMPETITORS] | Who you actually lose deals to. Check their names come back in step 2 — if the model doesn’t know them either, your whole category is invisible. | The market leader you’ve never competed with. You’ll measure a race you’re not in. |
[WHERE] | How a customer describes it. “Adelaide” or “northern suburbs” — the words they’d type. | Service-area jargon like “SA metro region 4”. Nobody types that. |
| The 12-question rule | Keep 12. Enough to see a pattern, few enough to read properly in one sitting. | Asking for 50. You get generic filler questions and you won’t read them, so the audit gets worse as it gets bigger. |
| The “don’t invent names” line | Leave it exactly as written. It’s the single most important sentence in the prompt. | Removing it. You’ll get a tidy competitor list, half of it fictional, and you’ll benchmark against ghosts. |
| The “don’t flatter me” line | Leave it. Assistants default to encouraging, and encouraging is the enemy of an audit. | Removing it. Invisibility gets reported as “some room to grow”. |
Below is an illustrative run for a composite business — a commercial refrigeration repair company in Adelaide. It's a representative shape of result rather than a specific client's data, and it's the pattern we see most often.
| # | Named us? | Named instead | What earned them the mention |
|---|---|---|---|
| 4 | No | Two national chains + “your local emergency refrigeration service” | The chains publish emergency-callout pages that name the symptom and promise a response time. The generic phrase is a hedge — it means the model has no specific local knowledge. |
(a) — no knowledge of this business at all. Across 12 questions the business was named 0 times without browsing, and 2 times with browsing on. The category's local layer is essentially empty: for 7 of 12 questions the model reached for national chains or a generic placeholder, because nothing in Adelaide has published anything connecting a symptom to a name.
Zero mentions sounds like a disaster and is actually the best-case starting point. The competitors named weren't beating them on reputation — they were the only ones who'd written anything down. When a model falls back to "your local emergency refrigeration service", it's telling you nobody owns that answer yet. Nine times out of ten the fix isn't outspending anyone; it's publishing the specific, symptom-shaped pages that nobody in the category has bothered to write.
We've watched all of these happen, usually within the first ten minutes.
The most natural question and the most useless. It's leading, the model is agreeable, and you'll get a confident description of a company that may not resemble yours. Never mention your name until the audit is finished.
Output varies between sessions on identical input. One run is an anecdote. The finding isn't "we were absent" — it's "we were absent in four out of four runs across two models". That's a claim you can spend money on.
Step 4 asks the model to diagnose, and its diagnosis is a hypothesis — a well-informed one, but not a fact about its own weights. Check it against reality. If it says a competitor out-published you, go and look at what they published.
Those five items are a starting hypothesis, not a plan. At least one will usually be wrong for your business — chasing a question you can't win or shouldn't want. Judgement is the part that can't be prompted.
Almost everyone scores zero the first time. It doesn't mean your business is failing — it means nothing on the open internet explains what you do in the words your customers use. That's a content problem, and content problems are among the most fixable ones you have.
The audit is diagnosis, not treatment. It's genuinely satisfying to run and it changes nothing on its own. The value is entirely in what you publish afterwards — and in re-running it next quarter to see whether it worked.
GEO — generative engine optimization — is the work of becoming the business AI assistants name. It's what we do for clients across Nepal, Australia and the USA. If you ran this and didn't like the answer, that's a good reason to talk. If you ran it and liked the answer, run it again next quarter, because your competitors are reading this page too.
