How to check if ChatGPT recommends your business

Why one quick question is not a check

Most owners test this by typing their own business name into ChatGPT. That tells you whether ChatGPT can describe you once you have already named yourself. It says almost nothing about whether a stranger looking for your kind of service would hear about you. The useful question is the unbranded one: when someone asks for a recommendation in your category and area, does your name come up?

Answers also move around. OpenAI says ChatGPT search results can be shaped by things like your approximate location, IP address and, where enabled, memory, and that no position in an answer is guaranteed. So a single answer is an anecdote. A small, fixed set of prompts, run the same way and saved, is evidence you can compare over time.[1]

Set up a fair test in five minutes

1. 1Start a separate fresh conversation for each prompt, not five follow-up questions in one chat. Where available, choose Temporary, then Unpersonalized before sending. Temporary Chat alone can still use memories and custom instructions when Personalized is selected. Record the personalization setting you used.
2. 2Select Search from the tools menu and confirm a search actually ran. ChatGPT can also search automatically, so do not infer the mode from the absence of citations alone. Keep the same search setup for every run.
3. 3Write the location and language into the prompt itself (for example "in Leeds, UK" and ask in English) rather than relying on where ChatGPT thinks you are. Note your real browsing location too.
4. 4Do not mention your business name, website or any hint about it. You want the answer a stranger would get.
5. 5Copy the full answer and the listed sources into your record before you close the chat. Screenshots are fine as a backup, but text is easier to search later.

OpenAI's help article describes choosing Search in the tools and checking the citations and Sources panel on an answer; that panel is where you confirm whether your own site, or a page about you, was actually used.[1]

OpenAI currently documents both Personalized and Unpersonalized Temporary Chats. Unpersonalized does not use memory, custom instructions or plugins; it does not eliminate location effects or safety context. Check the available controls rather than assuming that a new or temporary chat is automatically neutral.[3][1]

Five prompt templates that mirror real customers

Pick prompts that match how buyers research, compare and decide. Replace the bracketed parts once, then never change the wording between runs, or your before-and-after numbers stop being comparable.

Example fill-in for illustration only: [service] = emergency boiler repair, [city/country] = Bristol, UK, [constraint] = weekend call-outs. Use the words your customers actually use, not your internal service names.

Mention, recommendation and citation are three different things

• Mention: your business name appears anywhere in the answer, including in a list of many options or a passing reference.
• Recommendation: the answer actively suggests you, for example as a top pick, a good fit for the stated constraint, or one of a short set it tells the reader to contact.
• Citation: a page is linked as a supporting source for the answer. Record links to your own domain separately from links to directories or review sites that mention you. Open the source to check whether it actually supports the claim: a displayed citation can be wrong.

Keep them separate in your notes. You can be cited without being recommended (your page was used for background), recommended without your site being cited (the facts came from a directory), or mentioned only as a contrast. Each points to a different next step.

Record sheet and how to do the arithmetic

Hypothetical arithmetic, not measured data: if you ran all five prompts and your business was recommended in 2 of them, your recommendation rate for that run is 2 of 5 = 40%. Calculate mention rate and citation rate the same way, separately.

Use only successful answers as the denominator. If one prompt errored, refused, or returned no answer, your denominator is 4, not 5, and you should note the failure rather than count it as a miss. Small samples swing a lot, so treat a change from 2 of 5 to 3 of 5 as a hint to keep watching, not a verdict.

Repeat the same prompts and keep the real answers

• Re-run the identical five prompts on a fixed schedule, such as weekly, with the same setup.
• Save the full answer text and sources every time, so you can see what changed rather than relying on memory.
• Click the cited sources. If a competitor keeps appearing, look at which pages are cited for them; that tells you what material ChatGPT is drawing on.
• Log any change you made to your own site or listings with a date, so you can line it up with later runs.

If this is too much to do by hand, the free scan runs 5 ChatGPT answers once per email address through Bright Data against live chatgpt.com. The $19 USD one-off full check asks 25 questions across 3 assistants (75 answers) using their APIs with web search, which may differ from consumer chat apps, and adds audit details and an action plan.