Open ChatGPT, Gemini and Google in a private window, logged out, from the country your buyers are in. Ask each one the questions your buyers ask, word for word, in a new chat each time. Save every answer with its sources, then count who gets named and which pages get cited. That is a do-it-yourself AI visibility audit.
The rest of this post covers each step in detail and the common mistakes that make the results useless. It is the same process I use for the free AEO teardown, minus the scoring.
Why check logged out, in a private window?
A logged-in account influences the answer. ChatGPT and Gemini can remember past chats and saved preferences, and Google personalizes results for signed-in users. If you check while logged in as yourself, you see an answer about your company that was partly shaped by you.
A private window with no account is the closest proxy for what a stranger sees. Your buyer may be logged in to their own account, but you can't reproduce their history, so the logged-out answer is the fair baseline.
Why use a VPN?
The same question can get a different answer in a different country. AI tools search the web before they answer, and the results they search are local. A private window doesn't change that: it hides your account and history, but the tools still see where you are connecting from. If your buyers are in the US and you are not, connect to a US server before you start. If your buyers are in Germany, use a German one.
Write down the location you used. A check is only comparable with the next one if it was run from the same place.
Which questions should you ask?
Ask two kinds of question, because they test different things:
| Kind | Example | What it tests |
|---|---|---|
| Comparison, naming you | "Acme vs Globex" | Whether AI describes you correctly and fairly when the buyer already knows your name |
| Buyer question, not naming you | "Best expense tool for a 50-person startup" | Whether AI brings you up at all when the buyer is still building a shortlist |
The second kind is easy to skip, and it matters most. A buyer who asks "best tool for X" and never sees your name won't search for you later.
Write the questions the way a buyer would type them. Your own website is a good guide: who your customers are, what your case studies say, which competitor you lose to. Six questions is enough for a first check: three comparisons with your main competitor and three buyer questions.
How do you ask each question?
- Start a new chat for every question. An earlier answer in the same chat colors the next one.
- Type the question exactly the same way in every tool. Copy and paste it so nothing changes.
- Wait for the full answer, including the sources at the end.
- Save the whole answer as text before you move on.
Do one tool at a time: all six questions in ChatGPT, then all six in Gemini, then all six in Google. For Google, the answer is the AI Overview at the top of the results, or AI Mode if there is no Overview. If Google shows no AI answer, write that down, and note the top three regular results.
Perplexity asks you to sign up after a couple of logged-out questions. Don't sign in to get around it. Note that it stopped, and leave it out of that check.
What should you write down?
For every answer, record the same things, so one month can be compared with the next:
| Field | Why |
|---|---|
| Date, tool, logged out, location | To know later what you were looking at |
| The full answer, word for word | So any quote you use later is exact |
| Every source it cites | The sources are what you can change |
| Whether it names you, and your competitor | The basic count |
| Who the answer favors | Being fifth in a list of eight is different from being recommended |
| What it gets wrong about you | Old pricing, missing features and wrong categories show up often |
The sources need extra care. Logged out, the copy button in ChatGPT and Gemini drops the source links, and Google codes them. Click each source pill and copy the address it opens. It is slow, but the sources tell you which pages AI trusts, and that is where the fixes are.
What does a result look like?
Selah, a pricing tool for creators that I built, is my own test case. On October 24, 2026, I asked ChatGPT, Gemini, Perplexity and Google 20 questions a creator asks before pricing a brand deal. That run went through an API rather than by hand, with the same rules: logged out, US location, the same wording every time.
- When the question named Selah, it was named in 11 of 12 answers.
- When the question didn't name it, Selah was named in 4 of 67 answers. On 14 of the 17 buyer questions, no tool named Selah at all.
- One page did most of the work. For "HashtagPayMe alternatives", three of the four tools cited Selah's own comparison page on selah.fyi.
The gap between 11 of 12 and 4 of 67 is what the check is for. AI knew Selah once the question said its name. It rarely brought Selah up when the buyer didn't know it yet.
How I do it
For teardowns I do this by hand. I turn on a VPN set to the US, or wherever the company's buyers are. I open an incognito window, logged out of everything, and ask the first question. Then I read the answer and go through the source pills one by one, so I know exactly which pages the answer came from.
The sources are usually the most useful part. Most of the time the answer leans on a handful of pages, and many of them are comparison or "best of" pages written by someone other than the company.
How often should you check?
Monthly is enough for most companies. AI answers change as the models update and as new pages get published, so one check is a snapshot. Keep the same questions, the same tools and the same location each time, so a change in the answer means something changed outside.
What a check by hand can't tell you
- How many buyers ask the question. No tool can see what people type into ChatGPT. Google search volume for the same question is the closest guide.
- Whether the answer is stable. The same question can get a different answer an hour later. One check per question is a sample, so look for patterns across questions and tools rather than reading too much into one answer.
- Why the answer is what it is. The sources are the best clue, and the method page explains how I score them.
Want it done for you? Request a free AEO teardown. I run ten questions against one competitor and send you every answer and source, with three fixes. For why this matters for B2B software, see how big a deal AEO is.