AI assistants already describe your company to buyers every day. Here is a practical way to see what they say, in about half an hour, using nothing but a few well-chosen prompts.
Daniel Arons · Jun 2026 · 6 min read
Your buyers are asking ChatGPT about your category right now. Some of those answers name your company, some get your facts wrong, and some recommend a competitor instead. Most teams have never actually looked.
This is a do-it-yourself tutorial. You can run it in about 30 minutes, with no special tools, and walk away with a concrete list of what ChatGPT gets right, what it gets wrong, and where you are simply missing. For a structured, repeatable version, you can later move to a proper AI visibility audit. For now, start here.
Set up a clean read first
Before you type anything, get the conditions right. The biggest mistake is auditing in your own logged-in account, where months of personalization and memory have taught the model who you are. That gives you a flattering, distorted picture no real buyer would ever see. Turn off memory and personalization, or open a fresh temporary or incognito chat. The goal is to look like a stranger who has never heard of you, because that is exactly who you are trying to understand.
Keep one more thing in mind. ChatGPT answers differently when it is browsing the live web than when it answers from training data alone. A browsing answer can pull in your latest pages and recent news; a training-only answer reflects what the model already absorbed, which may be months old. Note which mode you are in for each prompt, because they tell you two different stories.
Test as a buyer, not as yourself
Here is the instinct to resist: typing your own company name and reading the summary. That tells you what the model says when it already knows who you mean. It does not tell you whether you show up when a buyer who has never heard of you goes looking for a solution. Real buyers rarely start with a brand. They start with a problem or a job to be done, so your prompts should sound like theirs, phrased in the words your customers actually use, not your internal product language.
“Searching your own name tells you what AI says about you. Searching like a buyer tells you whether AI says anything at all.”
Three kinds of prompt to run
Use three families of prompt, each answering a different question. Category prompts test whether you get named when nobody asked for you: 'what are the best tools for
A worked walkthrough: three prompts, side by side
It helps to see the three prompt types run in sequence. Picture a payroll tool called Northwind that serves construction firms. In a temporary chat with memory off, you run the prompts in the order a buyer would hit them: broad first, specific later.
The category prompt
You start with 'what is the best payroll software for a construction company with field crews'. The answer lists four or five names, a sentence each. The first thing you look for is binary: are you in the list at all? If Northwind is absent, that is the session's most important finding, and you note the phrasing that left you out. If Northwind is present, read where it sits: the first name with a confident reason, or the last name dropped in as an 'also consider' line? Position and the sentence attached to you matter as much as inclusion.
Then change one word and run it again: 'payroll for contractors with union crews', or 'job costing payroll for a small construction business'. Watch whether the cast of names changes. A different shortlist on a near-identical question means there are buyer wordings where you appear and others where you vanish. Both belong in your notes.
The direct knowledge prompt
Next you ask 'what do you know about Northwind' and 'tell me about Northwind payroll', grading accuracy line by line. Suppose the answer calls Northwind 'a general HR and benefits platform'. If you are actually a payroll specialist for trades, that is a category error, exactly the kind of mistake that pushes you off the construction shortlist above. Suppose it lists a feature you retired, or hedges with 'I do not have detailed information about this company'. Each is a separate finding: a wrong fact, a stale fact, or a thin-coverage signal. Read for what is missing too: if the answer never mentions the one thing you are best known for, that absence is data.
The comparison prompt
Run the same prompts across multiple models
ChatGPT is one voice, not the whole conversation. Your buyers also ask Claude, Google Gemini and AI Overviews, and Perplexity. Each was trained on different data and reasons differently, so the same prompt can produce a very different cast of brands. Run your short list of prompts across each assistant you can access. You are not looking for one verdict, but for the pattern across all of them, because that pattern is closer to what your market actually experiences. Understanding how AI assistants form and share these answers is the heart of Contextual AI Presence Mapping©.
Log everything so you can compare
Memory will not save you here, so write things down. Open a simple sheet with a few columns: the prompt, the model, whether browsing was on, whether you were named, which competitors were named, and any wrong facts. Paste the relevant part of each answer next to it. Run each important prompt more than once, in separate fresh sessions, because answers vary and a single run can mislead you. Appearing in two of three runs is a meaningfully different reality than three of three or zero of three.
“One answer is an anecdote. The same prompt run several times, logged side by side, is data you can act on.”
Turn it into a structured gap list you can track over time
Convert your scattered observations into one structured list. Do not keep it as prose; make each finding a row, structured enough that anyone on your team can understand it cold weeks later. Group the rows into three buckets. The first is where you are absent: the prompts where you never appeared, each with the exact phrasing, the assistant that produced it, and which names appeared instead. The second is where facts are wrong: every hallucinated, stale, or mislabeled claim, written as the model stated it next to the correct version. The third is where you are framed weakly: comparisons where the model treated you vaguely, with the specific competitor claim you could not match.
Name the competitors and their likely sources
For the absent and weak-framing rows, add a column for the competitors the model named and a column for where that material probably came from. You will not know the exact source, but you can usually reason about it: a competitor described in crisp, current language likely has clear pages of their own, listings or roundups that name them, review-site presence, or recent coverage. Writing down your best guess turns a complaint into a lead, because the source is what you would study and try to match. Note too whether each finding came from a browsing answer or a training-only one, because a browsing miss points at pages the model can read today while a training-only miss points at older, more widely repeated material, and the two call for different fixes.
Record it so you can track change over time
Stamp every row with the date you ran it and how many times you saw it, such as two of three runs. That turns the list from a snapshot into a baseline. When you make a change and re-run the same prompts weeks later, add a fresh dated row beside the old one and compare: did you move from absent to mentioned, did the wrong fact soften, did the framing get more specific? Without the date and run count you are guessing whether anything moved. With them, you can see it.
This list is the real output of the exercise: the difference between a vague worry that 'AI might be saying something' and a prioritized set of problems you can fix and then verify. To understand the discipline behind closing these gaps, the generative engine optimization approach gives you the wider frame.
Know what this method can and cannot tell you
A manual audit is the right first move, but be honest about its limits. You are sampling a handful of prompts at one moment, by hand, and answers shift between sessions, accounts, and browsing modes. You cannot watch every question your buyers ask, and you cannot easily tell whether a fix changed anything.
That is where the manual version reaches its ceiling and a continuous, mapped approach takes over: many prompts run repeatedly across multiple assistants, tracked over time, tied to the actions that change what AI says. You can start that today by seeing how Aethon maps and improves your AI presence. But even before that, run the 30-minute audit above. The clearest way to find out what ChatGPT says about your company is to go ask it, like a buyer would, and write down every answer.
Frequently asked questions
Why shouldn't I just search my own company name in ChatGPT?
Searching your own name only shows what the model says once it already knows who you mean. It hides the more important question: whether you appear at all when a buyer describes their problem without naming you. Test category and use-case prompts in your buyers' words, then check your name directly as a separate step.
Why do I get different answers each time I ask ChatGPT the same question?
AI assistants generate answers probabilistically, so responses vary between sessions even with identical prompts. Personalization, memory, and whether browsing is enabled add more variation. Run each important prompt several times in fresh sessions and log the results so you can see the pattern rather than reacting to a single run.
Should I turn off memory and personalization before auditing?
Yes. Memory and personalization teach the model who you are, which produces a flattering answer no real buyer would see. Use a temporary or incognito chat, or turn memory off, so you read the model as a stranger would. That stranger's view is the one that matters for new buyers.
Does it matter whether ChatGPT is browsing the web or not?
It matters a lot. A browsing answer can pull in your latest pages and recent news, while a training-only answer reflects older absorbed knowledge that may be months stale. The two modes can disagree, so note which mode produced each answer and treat them as separate signals.
What do I do after I find wrong information about my company?
Add it to a structured gap list grouped into wrong facts, missing category appearances, and weak comparisons, each tied to the prompt that produced it and stamped with the date and run count. Wrong facts usually trace back to thin, stale, or inconsistent material about you, so the fix is clearer, more consistent information in the places models rely on. Re-running the same prompts later then lets you confirm whether what AI says actually changed.
Written by
Daniel Arons
Co-founder & CEO, Aethon AI
Daniel co-founded Aethon AI in November 2025 to close the gap between how marketers measure AI visibility and what AI is actually doing with their brands. Before Aethon, he spent eight years building digital marketing programs in New York across SaaS, financial services, and consumer brands. He holds an MPA from Baruch College and a BA in Public Relations from SUNY Oswego.