How to Check If ChatGPT Recommends Your Business

Most businesses have never checked this. They assume that because they rank well, AI assistants must be mentioning them. As covered in our post on why ranking first no longer guarantees an AI citation, that assumption is often wrong.

The good news is you do not need software to find out. The method below is the same one we run at the start of client engagements, stripped down to what you can do manually. It takes about two hours for a first pass.

First, the mistake almost everyone makes

Do not open ChatGPT and ask it “do you recommend my business?” or “what do you think of [your company]?”

Three things go wrong when you do.

• You have primed it. By naming your company in the prompt, you have guaranteed the answer is about you. That tells you nothing about whether you would surface unprompted.
• Models tend toward agreeableness. Ask a language model to evaluate something you clearly have a stake in and you will usually get a diplomatic answer. It is not a reliable critic of you.
• Your account is personalised. If you are logged in, ChatGPT may be drawing on memory of previous conversations, including ones where you discussed your own business. Your results will not match a stranger’s.

The question you actually need answered is different: when someone who has never heard of you describes their problem, does your name come up?

Step 1: Build the prompt set

This is the step that determines whether the whole exercise is useful. Test how buyers ask, not how you would phrase a keyword. Real prompts are conversational, often eight words or longer, and frequently describe a situation rather than name a service.

Cover all five categories below. Twenty to thirty prompts total is enough for a first pass.

Step 2: Clean the session

Personalisation will corrupt your results if you skip this.

• Use a temporary or incognito session. ChatGPT offers a temporary chat mode that does not use or update memory. Failing that, log out entirely, or use a private browser window.
• Check your location is right. Local answers depend on where the engine thinks you are. If you are testing “SEO company in Denver” from an office in Denver, that is correct. If you are testing a market you do not sit in, results will differ from a real buyer’s.
• Do not chain prompts. Start a fresh chat for each prompt. Asking five questions in one thread means answers two through five are influenced by answer one.

Step 3: Run each prompt at least three times

This is not optional, and it is the step most people get wrong.

Generative answers are probabilistic. The same prompt can return different companies on different runs. Ahrefs published research titled “AI Overviews Change Every 2 Days (But Never Change Their Mind)”, finding that the specific content and citations shift frequently even while the overall stance stays stable.

A single run tells you almost nothing. Three runs per prompt per engine gives you a usable signal. If you appear in one of three, that is a real and different result from appearing in three of three.

Which engines to test

Test at least ChatGPT, Google AI Overviews and Gemini. Add Perplexity and Copilot if you have time. The engines disagree with each other far more than people expect, so testing only one gives a misleading picture. Ahrefs found only 6.82% of ChatGPT results overlap with Google’s top 10, and that Google’s own AI Overviews and AI Mode shared only 13.7% URL overlap in December 2025.

If you have to pick one, pick ChatGPT. SE Ranking’s 2026 analysis found it accounts for roughly 74.78% of all AI referral traffic, well ahead of Gemini at 11.56% and Perplexity at 7.23%.

Step 4: Record four things per run

Use a simple spreadsheet. Four columns are enough.

Step 5: Read the result honestly

Your share of voice number on its own is not the interesting part. What matters is the pattern, because different patterns point to different underlying problems.

Five things that will invalidate your results

1. Testing while logged in. Memory and personalisation will show you a version of reality only you see.
2. Running each prompt once. Answers vary between runs. One data point is noise, not signal.
3. Using keywords instead of questions. Nobody types “SEO company Denver” into ChatGPT. They type a sentence.
4. Only testing prompts you expect to win. The uncomfortable prompts are the informative ones.
5. Testing once and never again. This is a trend metric. A single snapshot cannot tell you whether you are improving.

What to do with the results

Once you have your baseline, the priority order is usually the same.

• Fix factual errors first. If a model is stating something wrong about your business, that is both the fastest fix and the highest risk item. Correct your details everywhere they appear publicly.
• Look at the sources column. Whichever pages keep appearing in the citations are the pages that decide your category. Getting listed on them is usually the single highest-leverage action available.
• Check your coverage against the prompts you lost. Absent from problem prompts usually means you have service pages but no content answering the questions that precede a purchase.
• Re-test monthly. Same prompts, same method, same day of the month. The trend is worth more than the starting number.

Frequently asked questions