How to Check What ChatGPT Says About Your Business

You open ChatGPT, type your business name, and read the answer. It sounds right, or it sounds wrong, or it mentions a competitor instead. Either way, you close the tab believing you now know what ChatGPT says about your business.

You might not. The exact same question, typed by two different people five minutes apart, can come back with two different answers, and the gap often has nothing to do with your website. It has to do with whether ChatGPT actually looked anything up, what it already remembers about the person asking, and where that person happens to be. None of that is about you. All of it can make your check look better or worse than reality.

This isn't a guide to fixing why ChatGPT doesn't know you. That's a longer, separate project covered in our deep diagnostic. This is the narrower thing that has to come first: how to run the check itself so the answer you get is actually about your business, not about your ChatGPT account.

Two different things can happen when you ask

When you ask ChatGPT a question about a business, one of two things happens underneath, and they produce very different kinds of answers.

It can answer from what it already learned during training. That's a fixed snapshot of the web up to a cutoff date, baked into the model itself. No lookup happens. If your business is newer than that snapshot, or anything about it has changed since, this answer is reciting something stale, however confident it sounds.

Or it can search the web right now and build the answer from what it finds. OpenAI's own help documentation puts it plainly: "ChatGPT may search the web automatically when your question would benefit from current information." That word "may" is doing real work. Whether a given question triggers a live search is the model's own judgment call, not something you control by asking nicely.

You can usually tell which one happened after the fact. OpenAI's own documentation says a web-search response "may include citations" and that you can select a Sources control, when available, to view the pages it actually used. If your answer has no Sources button and no inline citation markers, ChatGPT didn't look at anything. It told you what it already thought, not what's currently on your pages.

That distinction matters because a correct-sounding answer built from memory and a correct-sounding answer built from a live search are not the same kind of evidence. One tells you what the model learned a while ago. The other tells you what it can find today. If you're trying to find out whether your current website is reaching ChatGPT at all, only the second one is actually testing that.

Force the search, don't hope for it

Because the automatic trigger is a judgment call, the more reliable move is to not leave it to chance. OpenAI documents a manual path: open a conversation, select "View all tools," select "Search," then ask your question. You can also type / in the message box and choose Search directly.

This matters most for exactly the kind of question you're likely to ask. "What is [business name]?" doesn't obviously "benefit from current information" the way "what's the weather tomorrow" does, so there's a real chance it gets answered from memory even when a search would have told you something more current. If you want to know what ChatGPT finds on your site right now, ask for the search explicitly instead of assuming your question earned one.

Log out, or use a chat with no memory

Here's the part almost nobody checks, and it's specific to ChatGPT in a way that doesn't apply to asking a search engine something.

If you're signed in and Memory is turned on, ChatGPT can draw on saved details and past chats to shape its answer, and that includes the search itself. OpenAI's documentation is direct about this: "If memory is enabled, ChatGPT may use relevant saved memories when rewriting a search query." Its own example: someone who has mentioned they're vegan and live in San Francisco asks about nearby restaurants, and ChatGPT quietly turns that into a search for "good vegan restaurants San Francisco," a more specific query than the one that was actually typed.

Now apply that to your own situation. If you've ever used your ChatGPT account to research your own industry, mention your company, or talk about your location, that context doesn't stay in one conversation. It can follow you into the next one and shape both what gets searched for and how the answer gets worded. The answer you get isn't necessarily what a stranger with no history would see. It can be the model being quietly helpful to you specifically, which is a different thing than being visible to the public.

The fix is built into ChatGPT already. Start a temporary chat and, before you send anything, choose the unpersonalized option rather than the personalized one. OpenAI confirms the mechanism directly: "Before starting a temporary chat, you can choose whether it uses existing memories, custom instructions, and plugins. Select Unpersonalized if you do not want to use them," and "Temporary chats do not create or update memories, even with personalization on." Logging out entirely works too, and removes any question about which account's history might be leaking in.

Your account quietly tells it where you are, too

The same logic applies to location, and it's easy to miss because you never typed a city.

ChatGPT estimates your general location from your IP address by default, and can use a more precise device location if you've shared it, to make local answers more relevant. OpenAI's own example is a restaurant query: ask "what are some good restaurants near me" from the San Francisco area, and ChatGPT may rewrite that into "top restaurants San Francisco" before it searches, without you ever naming the city yourself.

For a local business, that's exactly backwards from what you want to measure. Say you run a plumbing company in Leeds. If you check from a browser that's also in Leeds, a favorable answer might just mean ChatGPT correctly matched your location to the question's location, which tells you nothing about what a customer in Manchester would see when they ask the same thing from there. If you want a result that doesn't depend on where you happen to be sitting, put the location directly into the question itself, "best plumber in Leeds," not "best plumber near me," and the local guess your account would otherwise supply is no longer doing the work for you.

Run the check properly, once

Put the three controls together and the process looks like this:

1. Open a temporary chat and choose Unpersonalized, or log out. Either removes your own chat history and saved memories from the answer.
2. Open the Search tool explicitly (View all tools > Search, or / then Search) rather than typing your question into a plain chat and hoping it searches.
3. Name a location in the question itself if your business is local, instead of relying on "near me" and your account's inferred position.
4. Ask your question, then check for a Sources button or inline citation markers before you read anything else into the answer. No sources means no search happened, memory-only, regardless of how specific the answer sounds.
5. Read what came back against what's actually on your site. If it cited a page, open that exact page and check it says what the answer claims.

Here's what the difference looks like in practice. Say two people at that same Leeds plumbing company each ask ChatGPT "who's a good emergency plumber in Leeds." The owner asks it casually, signed into a personal account that's been used for months of general chatting, from a browser physically in Leeds, with the Search tool never explicitly opened. The answer names the business, confidently, with no Sources button anywhere. The owner takes that as proof the business is visible. It isn't proof of anything beyond the fact that the model recognizes the name, possibly from training data months old, possibly helped along by nothing more than the account's own location matching the question. Run the same question through an unpersonalized temporary chat, with Search opened manually, and it may come back with a completely different, unflattering answer, built from a live page that's actually out of date. That second answer is the one worth acting on, because it's the one a stranger would actually get.

What you ask matters too, but less than you'd think for this specific purpose. A direct "What is [business name]?" tests whether ChatGPT can find and recognize you at all. A category question with no brand name, "who's a good [category] in [city]," tests whether it recommends you unprompted, a harder and more useful bar. Other guides on this site tell you not to lead with your brand name when measuring whether strangers get recommended to you, and that's the right advice for that specific goal. Here the goal is different: you're checking what ChatGPT currently says about you by name, so asking by name first is the correct test, not a mistake. For the full multi-question, multi-engine version of this check, see our guide to improving AI visibility, and for what to do once you've found a real gap in how ChatGPT describes you, see entity grounding.

One clean check isn't the end of it

A single answer, run properly, tells you where things stand today. It doesn't tell you whether that's typical. A model update, a retrieval change on OpenAI's side, or something a competitor just published can move the answer without anything on your site changing at all, which is exactly why a visibility number is a rate measured over repeated checks, not a single screenshot. How many checks are enough and how often to repeat them is covered in full in our breakdown of real AI metrics in AI search. The three controls above fix what you're measuring each time you run it. That post covers how often to measure it going forward.

Automating the clean version of this check

Once you know the controlled check is worth repeating, doing it by hand every week stops being realistic. We built this directly into the AI Schema Gen dashboard as AI Visibility tracking, available on Pro plans and above: add a question once, in the exact words a customer would use, and a weekly automatic check runs it and records whether your site was one of the sources the answer cited, where it came in that list of sources, and which other websites were cited alongside it. You can also trigger a check on demand instead of waiting for the weekly run, and turn on an email alert for the moment an AI assistant starts, or stops, citing your site.

Worth being straightforward about where this stands today, in the spirit of the honesty the rest of this site asks of any tool reporting an AI number: the dashboard lists ChatGPT, Claude, Gemini, Perplexity and Google AI Overviews, and ChatGPT is the one returning real, searched answers with sources right now. The others show plainly as "not connected" rather than being quietly counted as "not mentioned," which is a different, more honest thing than a dashboard that blurs the two together. As more engines come online, the same tracking extends to them automatically, with no change needed on your end.

This is a different job from a plugin that writes you a playbook for getting into those answers in the first place. If what you actually need is the step-by-step work of earning a citation rather than just checking for one, that's covered separately in getting cited by ChatGPT.

Common mistakes

Trusting an answer with no Sources button. If nothing in the response shows a citation, you got a recollection, not a lookup. Treat it accordingly, however specific and confident the wording sounds.

Checking only while logged in, with months of your own research sitting in your chat history. The more you've personally used ChatGPT to look into your own industry, the less your own check tells you about what a stranger sees.

Assuming "near me" means the same thing for everyone. It means wherever your account happens to be, not wherever your customer is. Name the city if the city actually matters to the answer.

Running it once and calling it a trend. One answer is a data point. A real signal comes from the same question, asked the same controlled way, more than once over time.

Confusing a clean reading with the deeper diagnostic. This guide is about getting an honest check. If the honest check comes back bad, that's a different and longer project, covered in the deep diagnostic linked above.