How to check if ChatGPT recommends your business

To check if ChatGPT recommends your business, ask it the questions your buyers ask, leave your name out, use a clean session, repeat each question several times and write down every answer in full. Then count how often you are recommended, merely mentioned or absent. One reply tells you very little, but a pattern across repeats tells you where you stand.

What questions should you ask?

Ask discovery questions, the kind a person types before they know your name. For an agency that might be "what's a good agency for B2B SaaS content?" For a software product it might be "which tool should I use to do X on a small budget?"

Aim for a spread:

• Best and recommend: "what's the best way to..." and "which tool should I use for..."
• Compare and alternatives: "alternatives to [a well-known competitor]".
• Specific needs: a use case, a budget, a size of company, a platform.
• Price: "how much does X cost?"
• How to: "how do I do X?", to see whether you appear when no product was asked for.

Write them the way customers do: first person, problem first, no leading words. How to find the questions your customers ask AI shows where to get them.

Keep your own name out of these. Asking "is [your brand] good?" only checks whether ChatGPT has heard of you, which is a different and less useful question. Branded questions are worth asking separately, to see what it says about you, but they do not measure whether you are discovered.

How do you set up a fair test?

Control what you can.

1. Use a clean session. Sign out, or use a fresh account or private window, and make sure memory and custom instructions are off. Otherwise your own history shapes the answer.
2. Choose one mode and stick to it. If you test with web search on, test with it on every time. Note which you used.
3. Fix the wording. Use the exact same question text each time you re-test.
4. Start a new chat for each run. A follow-up in an existing conversation is influenced by what came before.
5. Record the date and setup. Keep a note of the mode and any settings.

How to test AI answers properly covers these habits in more depth.

How many times should you ask?

Several. ChatGPT generates its answers, so identical questions can produce different brand lists. Ask each question multiple times, in separate chats, and count appearances. Three runs is the least that shows any variation, and more gives a steadier read. How many times should you ask an AI discusses the trade-off.

What do you record?

For each run, save:

• the question, exactly as asked;
• the full text of the answer;
• the names of brands in it, and whether each is recommended, neutrally mentioned or warned against;
• any sources or links shown;
• the date, and whether search was on.

Saving the whole answer is the important part. A tally without the underlying text cannot be checked later, and you will not be able to recount if you change your mind about a category.

How do you read the result?

Sort each appearance into a bin: recommended, neutral mention, negative mention, or absent. Then calculate, per question, how often each applied across your runs. A summary might say you were recommended in two of five runs for one question and absent from all runs for another.

Then look at the sources. If ChatGPT searched, it may cite pages. Note which ones recur and whether you are on them. Those pages show where your recommendation is decided. Mentioned, recommended or cited explains why the bins matter.

What if ChatGPT says something wrong about you?

You may find the answer names you but gets facts wrong: an old product name, the wrong audience, a feature you do not have. Note it exactly, with the question and the date, and look at the sources if any were shown. The error often traces to a page that is out of date or ambiguous.

You cannot edit ChatGPT directly. What you can do is fix the sources: update your own pages so they state the facts plainly, correct the third-party listings you can influence, and make sure descriptions of you agree with one another. Then retest later to see whether answers change.

Be cautious about conclusions from one error. Ask the relevant question several times to see how often the mistake appears. A mistake that shows up in most runs is a pattern worth fixing. One that appears once may not recur. Keep a log of errors with the full answers, so you have a clear list of what to correct and a record of whether it worked.

What do you do about what you find?

If you are absent, check whether the problem is your own pages, which may be unclear or hard to fetch, or missing third-party presence. If the cited roundups never list you, that is a clear target. If you are mentioned but not recommended, look at how your competitors are described and whether your own pages state your strengths clearly.

Do not draw big conclusions from one session. Repeat the whole exercise on a schedule, such as monthly, to see movement. For one assistant you can do this by hand. For four assistants and sixty questions, a systematic process is worth setting up. The Citewake checker offers a quick first look, and the methodology shows how a full benchmark is run.

What is the short version?

Ask unbranded buying questions, in clean sessions, several times each. Save the full answers. Sort every appearance into recommended, neutral, negative or absent, and note the sources. Repeat on a schedule. That is all there is to a fair check, and it is enough to show you where you stand and what to fix first.