Asking ChatGPT "do you know my company?" is not brand monitoring. That prompt names you. The model will often echo the name back. You learn nothing about what buyers see.
ChatGPT brand monitoring is a repeatable check: same prompts, same engines, same match rules, week after week. You want to know if ChatGPT names you on unbranded buyer questions. You also want to know when it names a competitor instead.
I have run this by hand. Spreadsheets lie fast. Answers drift. Names collide. One lucky chat is not a measurement.
ChatGPT is a next-token machine. If your prompt contains the brand, the model keeps that token in working memory. It is more likely to emit it again.
Discovered Labs audited five B2B prompt sets at onboarding. Every set had anchoring inflation. In the worst case, the reported mention rate was 10.8%. On unanchored prompts (no brand name in the question), the rate was 1.9%. Most of the "visibility" was the prompt asking about the brand.
So split your set:
Prompt type
Example
What it measures
Unbranded
"Best AI visibility tracker for a 5-person SEO team"
Natural recall. This is the number that matters.
Category
"Tools that track ChatGPT citations"
Whether you belong in the category at all.
Comparison
"Profound vs Peec AI vs Otterly"
Whether you make the shortlist when peers are named.
Branded (control)
"What is AI SEO Tracker?"
Accuracy and hallucination, not visibility.
Report unbranded mention rate first. Keep branded prompts as a fact-check. Never blend them into one percentage.
That is the first rule of LLM SEO: the model only knows what the open web repeats. A prompt that already contains your name is not evidence that the web repeats it.
The prompt set to actually run
You do not need 2,000 prompts on day one. You need a set that looks like buyer language, not keyword fragments.
Start with 20 to 50 prompts. Pull them from:
Sales calls and support tickets
Google Search Console queries that sound like full questions
Reddit and review-site wording in your category
"Best X for Y" and "X vs Y" pages that already rank
Write the long version. People type full thoughts into ChatGPT. "Best CRM" is a Google fragment. "Best CRM for a 5-person agency that already uses Slack" is a ChatGPT prompt.
Group them:
Awareness: "What tools track whether ChatGPT mentions a brand?"
Consideration: "Best AI visibility tools for a small SEO team"
Comparison: "AI SEO Tracker vs Profound"
Job-to-be-done: "How do I see if ChatGPT cites my site?"
Keep a written list. Freeze it for at least four weeks. If you swap prompts every scan, you cannot tell whether the brand moved or the question changed.
Add aliases. ChatGPT may say "AISEOTracker", "AI SEO Tracker", or a truncated name. Put every legal variant in the match list. Miss those and you undercount.
Skip leading questions:
Bad: "Is AI SEO Tracker one of the best AI SEO tools?"
Good: "What are the best AI SEO tools for tracking ChatGPT mentions?"
The first one is a test of politeness. The second one is a test of the market.
For the wider discipline, see answer engine optimization. Monitoring is how you know if AEO is working. It is not AEO by itself.
A cadence that survives ChatGPT noise
Same prompt, same model, different day: different answer. That is not a bug. Production endpoints sample with temperature above zero. Batching on GPUs adds more variance.
Discovered Labs treats this as two kinds of noise:
Model noise: the same prompt returns a different write-up on the next run
Sample noise: your 30 prompts are a thin slice of what buyers actually ask
So do not treat one chat as a rank. Treat a week of runs as a sample.
A practical cadence for ChatGPT:
Cadence
When to use it
What you get
One-off audit
First look, or after a site rewrite
A baseline. Not a trend.
Weekly
Default for most teams
Enough snapshots to spot a real shift
Daily
After a launch, a listicle hit, or a model update
Tighter window. More cost. More noise per day.
Weekly is the floor if you want a sustained-shift test. One green week is a coin flip. Three weeks in a row above your baseline is a signal.
Run the same prompts in a logged-out or API context when you can. Personalization and chat history skew the answer toward what you already talked about. Buyers do not have your chat history.
Track ChatGPT on its own line. Do not average it with Perplexity and Gemini. Peec AI's docs split brand visibility from source visibility for a reason. ChatGPT often names a brand with no link. Perplexity often cites a URL. Mixing those into one "AI score" hides the engine that is actually failing.
If you also want Google's layer, pair this with AI Overview tracking. ChatGPT and AI Overviews do not agree often enough to share a number.
How to score mentions without fooling yourself
A mention is not a citation. A citation is not a recommendation.
Score each answer with four fields:
Mentioned? Exact brand or alias in the answer text. Yes or no.
Cited? A URL on your domain, or a named source card that points to you.
Position? First named, second, later, or buried in a list of eight.
Sentiment / accuracy? Positive, neutral, negative, or factually wrong.
Then compute:
Mention rate = answers that name you / answers in the set.
That is the visibility number. Peec calls this Visibility. Same idea.
Citation rate = answers that link or attribute your URL / answers in the set.
You can be mentioned with no URL. You can be cited as a source with no brand name. Track both. HubSpot's AEO write-up draws the same line: mentions build entity association, citations can send a click.
False positives to strip:
Your legal name inside a competitor's sentence ("unlike SmallerBrand, HubSpot…") if SmallerBrand is you and the model is not recommending you
A matching substring in another brand ("OpenAI" vs a product called "AI")
The model repeating your name because the user already said it in a follow-up
A hallucinated product that shares three letters with yours
False negatives to catch:
Abbreviations and old product names
"They" after a previous mention in the same answer (count the first mention, not every pronoun)
Misspellings that still point at you
Keep screenshots or stored answer text. ChatGPT will not show you last Tuesday's reply. If you cannot reopen the answer, you cannot audit a dispute.
For the citation layer across engines, see what AI citations are. ChatGPT is sloppier about showing sources than Perplexity. Do not grade it on Perplexity's rules.
Manual monitoring vs a real monitor
You can do the first audit by hand.
Open ChatGPT. Paste prompt 1. Record the four fields. Repeat. Do Perplexity next if you have time. Budget two to three hours for 20 prompts on one engine.
That is useful once. It is a bad weekly job.
Problems show up fast:
You change wording without noticing
You forget an alias
You score "kind of mentioned" as a yes on a tired Friday
You cannot show a client the raw answer from three weeks ago
You never run Gemini, Copilot, or AI Overviews, so ChatGPT becomes the whole story
AI SEO Tracker is the monitor I want for this job. You add the prompts. You name the brand and the competitors. It runs ChatGPT, Perplexity, Gemini, Copilot, and AI Overviews. You get mention rate, citation URLs, and an engine split instead of one blended score.
Start with the $49 audit if you have never measured this. Fifty prompts, five engines, a report you keep. If ChatGPT never names you, you do not need a daily dashboard yet. You need pages, listicles, and consistent brand copy.
If you already know the gap, weekly tracking starts at $79/mo. MCP is there if you want the same data inside an agent. Page Inspector sits on the higher plan when you need to see how a specific URL reads to a model.
That is different from asking ChatGPT about yourself in a new chat every Monday.
A 30-minute setup you can repeat
Do this once. Then let the monitor run.
Write 25 unbranded prompts and 5 branded control prompts. Freeze the list.
List aliases (legal name, product name, common misspelling).
List 3 to 5 competitors that already show up in answers. Be honest. Include the ones you wish were irrelevant.
Run the $49 audit or a manual pass if you insist on doing it once yourself.
Score mention rate on unbranded prompts only. Put ChatGPT on its own row.
Pick three gaps. A prompt where a competitor is named and you are not is worth more than a vanity branded query you already win.
Fix the page or the off-site mention, then wait. Do not rescan 12 hours later and call it a win.
Rescan weekly for four weeks before you change the prompt set.
If mention rate is 0% on unbranded "best X" prompts, stop optimizing titles for a week. Get named on a third-party list. ChatGPT copies consensus. Your homepage talking about yourself is a weak consensus.
That is why listicles and how you rank in ChatGPT are the other half of this work. Monitoring without distribution is a scoreboard with no game.
What to report to a human, not a dashboard
A founder does not need 40 prompt rows. They need four lines:
ChatGPT named us in N of 25 unbranded prompts this week (was N-1 last week)
Cited our URL in N of those
Named Competitor A on the prompts we missed
Top source URLs in those answers (so you know where to earn a mention)
If those four lines are stable for a month, your ChatGPT monitoring is working. If they swing 20 points every week, your prompt set is too small or too branded. Fix the set before you fix the site.
For share of voice math across competitors, use share of voice measurement as the sibling metric. Mention rate is "do we appear." Share of voice is "how much of the named set is us."
ChatGPT-specific traps
Memory and custom GPTs. If you test inside a logged-in account that already knows your brand, you are not testing the public model. Use a clean session.
Search vs training. ChatGPT with browsing on will pull live pages. ChatGPT with browsing off will lean on training data and prior consensus. Your buyers use both. Track the product they actually use (ChatGPT with search is the one that can cite a URL this week).
Shopping carousels. For ecommerce prompts, ChatGPT may show products instead of a prose shortlist. A brand mention in the prose and a product card are different events. Score them separately if shopping is in scope.
Follow-ups. "Tell me more about the second one" is not a new prompt. Do not add follow-ups to the mention-rate denominator unless you designed the study that way.
Copycats. If your category has a similarly named tool, add a negative match. False positives here are worse than misses. You will ship a blog post celebrating a mention that was not you.
When monitoring is enough, and when it is not
Monitoring answers "are we in the answer?" It does not write the page that gets you in.
If ChatGPT never names you:
Align the one-sentence description of the product everywhere (site, LinkedIn, directories)
Earn a spot on a third-party "best of" page that already gets cited
Put a direct answer in the first two sentences of the section that matches the prompt
If ChatGPT names you with wrong facts:
Fix the conflicting sources. Price on the site vs price on G2 is a classic split.
Add a clear facts block (price floor, engines, who it is for)
Recheck branded control prompts until the lie disappears
If ChatGPT names you but never cites you:
You have entity recognition. You do not have a quotable page. Tighten structure. See AI search visibility.
The monitor tells you which of those three problems you have. Guessing from one chat does not.
Run the audit. Freeze the prompts. Ignore branded vanity. Watch ChatGPT as its own engine. That is brand monitoring. Everything else is a demo.
ChatGPT
Perplexity
Claude
Grok
