How to track brand mentions in ChatGPT

Most people "track" ChatGPT by opening it, typing "what are the best project management tools," reading the answer, and either celebrating or panicking. That test is broken in at least four ways, and understanding why is the whole job. Once you understand why the manual check fails, the systematic method more or less designs itself.

Why asking ChatGPT once proves nothing

A single ChatGPT answer is a sample, not a fact. Language models generate text by sampling from a probability distribution over next tokens. Ask "best CRM for small teams" five times in five clean sessions and you can get five different lists, with different brands, in different orders. Nothing changed about your brand between 10:03am and 10:04am. The dice just landed differently.

That's the core problem. Four things compound it:

Personalization. Your ChatGPT account knows you. If you've spent three months asking it about your own product, your competitors, and your category, that history colors what it surfaces. You are the single most biased tester of your own brand's visibility.

Memory. ChatGPT's persistent memory stores facts about you across sessions, and custom instructions inject text into every prompt. Both quietly change the input. You think you're testing "best CRM for small teams." You're actually testing "best CRM for small teams, asked by someone who works at a CRM company and has mentioned Acme fourteen times."

Retrieval variance. ChatGPT decides per query whether to run a live web search. Sometimes it answers from parametric memory (what's baked into the weights), sometimes it browses and cites sources. Those two paths produce very different brand lists. You don't control which one fires, and you often can't tell from the output which one did.

No history. Even if you got a clean answer, you have one number with nothing to compare it to. Visibility is a trend. "We appear in 40% of category queries" means nothing until you know last month was 25%.

Put those together and the manual check gives you a random draw from a distribution you can't see, run through a personalized session, with no baseline. People make budget decisions on this.

What "a brand mention" actually means in ChatGPT

A mention is your brand name appearing in the model's generated answer to a prompt you did not seed with your name. That last clause is the whole thing.

There are three distinct events people mash together, and they're worth separating because they have different value and different fixes:

Track the second one. Report on the second and third. Ignore the first, and treat any tool or screenshot that leans on prompted recall as noise.

Position matters too. Being named first in a list of ten is not the same as being named tenth, and buyers behave accordingly. Any tracking worth doing records where in the answer you appeared, not just whether.

The manual method, done as well as it can be done

If you're going to do this by hand, do it properly. Here's the least-broken version:

1. Open a logged-out or temporary chat. In ChatGPT, use a Temporary Chat so memory is off and nothing is written back. Better still, use a fresh browser profile with no OpenAI login. Personalization is the biggest single contaminant.
2. Turn off custom instructions. If you're logged in, they still fire. Check them.
3. Write your prompts as a buyer, not as a marketer. "Best CRM for small teams" is a real query. "Is Acme the best CRM?" is not.
4. Run each prompt at least three times. Record how many of those runs mention you. That fraction is your mention rate for that prompt.
5. Record the full response text, not just yes/no. You want the competitor names and their order.
6. Do it on a schedule. Same prompts, same day of week. A snapshot with no cadence is a screenshot, not a measurement.

Do that across 20-25 prompts and you have a defensible baseline. You also have a spreadsheet you will abandon by week three, because 25 prompts × 3 runs × 4 engines is 300 manual queries per cycle. This is exactly the point where people either give up on measurement or automate it.

The systematic method

Systematic tracking means the same prompt set, on the same schedule, in a clean context, with every raw response stored and diffed. Four components, and none of them are optional:

A fixed prompt set

Your prompts are the measurement instrument. Change them and you break the trendline, the same way swapping thermometers mid-experiment ruins the temperature log. Lock a set of 20-30 buyer questions, version them, and only add new prompts as additions, never as swaps. If you don't have a good set yet, start here before you track anything, because a bad prompt set produces beautifully consistent, completely irrelevant data.

A clean context

No login, no memory, no custom instructions, no prior turns in the conversation. Every run starts from zero. This is the thing that's genuinely hard to do by hand and trivial to do with an API call, because the API has no memory by default.

One caveat worth knowing: the API and the ChatGPT consumer app are not the same system. The app has a system prompt, a browsing tool, and product-level scaffolding that the raw API doesn't. If you measure via the API only, you're measuring the model, not the product buyers actually use. Good tracking hits the surface real users hit.

Repeated sampling

Run each prompt multiple times per cycle and report a rate. If "best CRM for small teams" mentions you in 2 of 5 runs, your mention rate is 40%. Next month it's 3 of 5, or 60%. That's a real, movable number. A single binary yes/no per prompt is a coin flip you'll misinterpret as a trend.

Stored responses

Store the full text. Six months later you'll want to know not just that you dropped from 60% to 20% on a prompt, but which competitor took the slot and what the model said about them. You can't reconstruct that from a boolean. The raw text is where the diagnosis lives.

Manual vs systematic, side by side

What to do with the data once you have it

Three numbers carry almost all the signal. Mention rate is the share of your tracked prompts where you appear at all. Average position is where you land when you do appear. Share of Voice is your mentions as a fraction of all brand mentions across the set, which is the only one of the three that tells you whether you're winning or just present.

Then look at the prompts where you're absent. That list is your content roadmap. If you never show up for "alternatives to [Competitor]" but you own "best tool for [use case]," you know precisely what to go build. We see this in scan data constantly: brands are strong on their own category terms and invisible on comparison and alternatives queries, which is exactly where the buying decision happens.

Worth being clear about why any of this matters commercially. 51% of B2B software buyers now start their research in an AI chatbot, up from 29%, and 69% picked a different vendor than they'd originally planned based on AI guidance, per a G2 survey of 1,076 B2B decision-makers. Being absent from the answer isn't a missed impression. It's a missed shortlist.

Should you track ChatGPT only?

No. ChatGPT is the biggest surface, with 900M weekly active users as of February 2026, but it isn't the only one, and the engines disagree with each other more than you'd expect. A brand strong in ChatGPT can be invisible in Perplexity, which leans heavily on live retrieval and citations, or absent from Google AI Overviews, which now appear on roughly 48% of tracked queries and on 82% of B2B technology queries.

That disagreement is useful. If you're in ChatGPT but not Perplexity, your problem is probably crawlable, citable web content. If you're in Perplexity but not ChatGPT, your problem is entity presence in the training-and-memory layer: mentions on sites the model absorbed, not just pages it can fetch today. Tracking one engine hides that diagnosis entirely.

This is where a tool earns its keep. Spottlo runs your prompt set through ChatGPT, Perplexity, Gemini and Google AI Overviews on a weekly cycle, in clean sessions, and stores every response, so you get the mention rate, the rank and the competitor set per engine rather than a screenshot. All four engines are on every plan, which is worth checking against the rest of the category — most tools gate engines behind higher tiers, and a single-engine view is the exact blind spot described above.

What to do next

1. Write down 20-25 real buyer questions in your category. Pull them from sales calls and support tickets, not from a keyword tool.
2. Run five of them three times each in a Temporary Chat, logged out, and record the mention rate. That's your unfiltered baseline, and it usually stings.
3. Note every competitor that appeared. That list is your real competitive set in AI, which is often not the one on your battlecards.
4. Pick the three prompts where you're most absent and treat them as a content brief.
5. Get a free scan at spottlo.com to see the same prompts run across four engines with no signup, and compare it to what your logged-in ChatGPT told you. The gap is the point.