Someone asked whether ChatGPT recommends your product, and you didn't have an answer. That's a normal place to start. ChatGPT handles roughly 2.5 billion queries a day, and each answer leans on only a handful of sources, often two to seven. So the real question sitting under all of this is fair and a little uncomfortable: is ChatGPT citing my brand, or is it citing someone else?
This guide takes you through ChatGPT citation tracking from your first manual test to a weekly habit you can actually keep. You'll finish with a number instead of a hunch, and you'll know what that number means.
Nine steps. Most of them take minutes. Let's go.
Step 1: Separate a mention from a citation before you measure anything
These two words get used interchangeably, and that one habit ruins more AI visibility reports than anything else.
A mention is ChatGPT naming your brand in the answer text. No link. Something like "tools such as Acme and Brand X are popular here."
A citation is ChatGPT linking to a URL on your domain, either as an inline footnote or in the Sources panel at the bottom of the answer.
They come apart in both directions. You can be mentioned without being cited, named in the prose while the links all point elsewhere. You can also be cited without being mentioned, your URL sitting in the Sources panel while your brand name never appears in the answer at all.
There's a third thing worth naming: a recommendation. That's ChatGPT actively suggesting you in a best-of list. It's the outcome most teams actually want, and it's downstream of the other two.
You'll know this step is done when: you can write down, in one sentence each, what counts as a mention for your brand and what counts as a citation. Include product names and trademarks, not just the company name.
Where people go wrong: blending the two into a single "visibility score." Mention Rate and Citation Rate move independently, and a tool or spreadsheet that averages them will hide the thing you needed to see.
Step 2: Put ChatGPT in Search mode before you judge anything
This is the step that decides whether the rest of your work means anything, so take an extra minute here.
ChatGPT answers in two very different ways.
In training-data mode, the default, it answers from what the model already knows. It can name brands. It can even produce URLs. Those URLs are reconstructed from memory, which means they're often stale, wrong, or invented outright. There's no Sources button, because nothing was retrieved.
In Search mode, ChatGPT queries the live web while it answers. You get inline footnote-style links tied to specific claims, plus a Sources button that opens a panel listing every URL it pulled. ChatGPT Search arrived in late 2024 and has expanded steadily since, and it works on the free tier in supported regions.
Only Search mode produces citations you can audit. Everything else is the model talking from memory.
One genuinely useful thing about Search mode: it retrieves from the open web rather than from Google's index. A page that ranks nowhere on Google can still get cited if it answers the prompt cleanly and is easy to find. If you've been assuming your AI visibility is capped by your rankings, it isn't.
You'll know this step is done when: you can see either footnote numbers in the answer or a Sources button underneath it. If neither appears, you're in training-data mode.
Where people go wrong: mixing modes across a test. Half your prompts in Search mode and half in default mode gives you a trend line built on two different behaviors. Pick one mode, write it down, and stick to it.
Step 3: Set up a clean test session
ChatGPT knows you. That's lovely for daily work and terrible for measurement, and every attempt to monitor brand in ChatGPT answers depends on getting this right.
Memory, custom instructions, your prior conversations, and in some setups your location all shape what you see. If you've spent weeks asking ChatGPT about your own category, your account has quietly learned which brands you care about.
Here's the fix, and it takes about thirty seconds.
- Open a private or incognito window.
- Make sure you're logged out.
- Clear site data and cookies for the domain.
- Start a fresh thread with no prior messages.
- On mobile, use a private browser tab rather than the app.
If your buyers cluster in one country, add one more control: use a VPN pinned to that region and keep it pinned every time. Regional answers differ, and you want them to differ consistently.
You'll know this step is done when: you're looking at a blank chat, logged out, with no custom instructions loaded and no history in the sidebar.
Where people go wrong: running the test in the browser they use every day. It's the single most common reason two people on the same team get different answers to the same prompt and end up arguing about which one is real.
Step 4: Build a prompt set of 10 to 12 buyer questions
You're not testing ChatGPT. You're testing the questions your buyers actually type.
Write 10 to 12 prompts in plain language, spread across the moments where a buyer makes a decision:
- Discovery: "What are the best [category] tools?"
- Head to head: "Compare [your brand] and [competitor] for [use case]."
- Alternatives and switching: "I'm using [competitor] but want to switch. What are the closest alternatives?"
- Local intent, if your category has it: "What are the best [category] providers in [city]?"
- Purchase intent: "I have budget approved and need a [category] solution this week for a mid-sized team. Recommend one and name a runner-up."
Mix the shapes too. Open-ended, comparative, and transactional prompts pull different answers and different sources.
Then do the thing almost nobody does: give each prompt a stable ID and a version tag. P-007, v1.0, dated. Save it somewhere you won't lose it.
Pro tip: if you can't think of 12, look at your sales call notes and your support inbox. The questions people ask a human before buying are the questions they now ask ChatGPT first.
You'll know this step is done when: every prompt is written out, numbered, version-tagged, and frozen. Frozen is the important word.
Where people go wrong: rewording prompts between rounds. Swapping "best" for "top," or "tools" for "software," genuinely changes the answer. Every reword resets your trend line to zero. If a prompt truly needs to change, retag it as a new version and start its history fresh. A repeatable prompt set is the whole foundation here.
Step 5: Run each prompt and check ChatGPT mentions and citations
Now the actual work. One prompt per fresh thread, Search mode on, and three things captured for each.
- Mention, yes or no. Does the answer name your brand in any form, including product names and trademarks?
- Citation, yes or no. Does a clickable link to your domain appear inline or in the Sources panel?
- Position. Where did you land in the list? Being first of seven is a different business outcome than being seventh of seven.
Alongside those three, save a screenshot of the full response with the Sources panel open, the date and time, and the model plus mode you used. Put the date in the filename. Future you will need it.
That's the manual method, and it works. It's also where honesty helps: twelve prompts, run properly, verified, and logged takes most people an hour. Do it weekly across a few competitors and you've built yourself a part-time job.
This is the point where a tracker earns its keep. DeepSmith's AI Visibility module runs your prompt set on a schedule and records mention and citation results for every run, so the capture step happens whether or not anyone opens the app. Discover Prompts will even generate a starter set from your product, persona, and buyer-stage context if staring at a blank page is what's stopping you. The Pro plan is $99 a month and tracks ChatGPT with 50 prompts, which is roughly four times the working minimum from Step 4.
You'll know this step is done when: every prompt has a complete row. Mention, citation, position, screenshot, date, model, mode. Blanks become noise the moment you try to compare two weeks.
Where people go wrong: logging only what's in the prose and never opening the Sources panel. Citations hide there constantly. You can be a source for an answer that never says your name.
Step 6: Verify every cited URL before you count it
A URL that looks like yours isn't the same as a URL that is yours.
Open every single one. Check three things: does it load, is it genuinely your page, and is it the right page for what the prompt asked?
Watch for these:
- URLs that 404 or bounce through a redirect.
- URLs that point to a competitor or a third-party review you assumed was yours.
- URLs where ChatGPT invented a plausible path, something like yourdomain.com/best-pricing-2026, that has never existed.
Hallucinated URLs are convincing. Right domain, sensible slug, correct-looking structure. They pass a glance every time, which is exactly why glancing isn't enough.
While you're in there, read what ChatGPT says about you. If the description of your product is out of date or plain wrong, that's a separate problem worth logging, and correcting outdated brand information is a different job from earning a citation.
You'll know this step is done when: every cited URL has been clicked and marked as verified or flagged.
Where people go wrong: counting the URL string as a win. An unverified citation isn't a citation. It's a screenshot.
Step 7: Read the results without overreacting
You've got data, and you finally have a defensible answer to "is ChatGPT citing my brand" for this week. Take a breath before you push it further, because this is where good measurement goes sideways.
Six rules that will keep you honest:
One run is a snapshot, not a trend. Today's answer tells you where you stand today. It says nothing about direction.
Keep Mention Rate and Citation Rate apart. Rising mentions with flat citations means ChatGPT knows who you are but doesn't consider your pages worth linking. That's a content problem, not an awareness problem, and averaging the two would have hidden it.
Position is not binary. "We showed up" and "we showed up first" are different results. Track the number.
Read the sources, not just the prose. Which domains does ChatGPT lean on in your category? Review sites, forums, encyclopedias, vendor blogs? That mix tells you where the answer is really being formed.
A clean zero is a finding. If you don't appear anywhere across twelve prompts, that's not a failed test. That's the clearest possible baseline, and honestly the easiest one to improve from.
Never compare across rewritten prompts. Same wording or no comparison. This is Step 4 coming back to collect.
Where people go wrong: taking one bad answer to the boardroom. One prompt is an anecdote. Patterns across a prompt set, over several runs, are evidence.
Step 8: Track ChatGPT citations on a fixed cadence
Everything so far was an audit. To monitor brand in ChatGPT answers over time, you need three things locked down and one cadence you don't break.
Lock the inputs. Same prompt set, same version, same model, same mode, same region. Write the configuration down in the same place you keep the prompts.
Lock the cadence. Weekly suits most teams. Biweekly is fine for smaller prompt sets. Daily adds noise without adding signal, and monthly means you find out about a drop four weeks late. Run on the same day and roughly the same time each week.
Then track five metrics, and only five to start:
- Mention Rate: the share of tracked prompts where your brand is named.
- Citation Rate: the share of tracked prompts where at least one source URL points to your domain.
- Share of Voice: your slice of total mentions or citations across the set, measured against a defined competitor list.
- Visibility Trend: the period-over-period change in any of the above.
- Pages Cited: which specific URLs of yours get cited, and how often.
That fifth one quietly matters most. Once you know which pages ChatGPT trusts, you know which format, depth, and structure it rewards on your site specifically. That's a content brief you didn't have to guess at.
Two more worth adding when you have room: the source mix in your category, and the sentiment of the language around your brand when it appears.
Doing all of this in a spreadsheet is possible. It's also the reason most teams stop by week three. A system built to track ChatGPT citations on a schedule removes the part that breaks: DeepSmith reports Mention Rate, Citation Rate, Share of Voice, and Visibility Trend with the period-over-period delta already calculated, keeps full answer history per prompt, and shows in the Pages view which of your URLs earn citations and which prompts drove them. No tool controls what ChatGPT cites, including this one. What it removes is the manual capture and the spreadsheet math, which is what actually stops people from measuring at all.
You'll know this step is done when: two consecutive runs exist under identical conditions and you can state the change in one sentence.
Where people go wrong: starting with fifteen metrics. Get Mention Rate and Citation Rate running reliably for a month first. The rest can wait.
Step 9: Turn the numbers into your next move
Measurement that doesn't change what you publish is just an expensive hobby. Here's how to spend what you found.
If your Citation Rate is near zero but mentions exist: ChatGPT knows your brand and isn't linking to your pages. Look at the pages that should be cited and ask whether they answer the prompt directly, near the top, in plain language.
If both are near zero: you have a baseline and a blank page, which is a cleaner starting position than it feels like. Pick the three prompts closest to purchase intent and work only on those.
If competitors dominate your prompts: open their cited pages. Not their homepage, the exact URL ChatGPT chose. The format of that page is your brief.
If one of your pages is doing all the work: study it. Then build two more like it.
Look at the sources ChatGPT keeps pulling from in your category, too. If a review site or a forum shows up in most answers, your presence there is part of the picture, not a side quest.
The gap between knowing and fixing is where most AI visibility programs stall. DeepSmith closes that loop by producing the content against the gaps it finds, using your stored brand context so what comes out sounds like you rather than like a generic AI draft. Tracking and producing sit in the same platform, which means the finding and the fix don't live in two different tools with two different owners.
What to do next
Pick one hour this week. Build twelve prompts, run them logged out in Search mode, verify every URL, and write down two numbers.
That's it. That's your baseline. Next week you run the same twelve and you have a trend, which is more ChatGPT citation tracking than most of your competitors are doing right now.
Six weeks of that and "is ChatGPT citing my brand" stops being a question anyone in your company has to guess at.
If the weekly run is the part you know you won't keep up, hand that piece to a system. You can start a free DeepSmith trial and see real tracking data on your own prompts in seven days, no contract, no cancellation fee. Take the first step manually anyway. You'll read the dashboard better for having done it by hand once.
