Somewhere today, a potential customer asked ChatGPT a question you would love to answer. The model named a few brands. Maybe yours, probably not. This guide covers how to find out, and how to set up tracking that tells you whether it is getting better.
In short: define the prompts your customers actually use, run them against the models on a schedule, score the answers with a consistent formula, and review the trend weekly. You can do a rough version by hand in an afternoon. The six steps below cover both the manual route and the automated one.
A note on our own numbers: for a single tracked brand, Trackbase logged 4,393 brand mentions and 12,017 cited URLs across 4,004 domains in the past 30 days. That volume is why manual checking only takes you so far. We will be honest about where a tool becomes necessary.
Step 1: Decide which AI platforms matter for you
ChatGPT gets the headlines, but your buyers spread across more models than you think. Start from your audience:
ChatGPT: the default for consumers and most professionals. Always track it.
Google AI Overview and AI Mode: appear on top of normal Google searches, so they touch everyone who still googles. Track both.
Gemini: Android users and Google Workspace companies.
Perplexity: researchers and tech-savvy buyers. Cites its sources, which makes it useful for diagnosis.
Copilot: embedded in Windows and Microsoft 365, common in corporate environments.
Claude, Grok, DeepSeek, Llama, Mistral: smaller individually, but together a meaningful share, and their answers reveal what the underlying training data says about you.
Manual tracking forces you to pick two or three. Automated tracking makes the choice irrelevant: Trackbase queries all 11 daily for the same effort.
Step 2: Build your prompt library
The single biggest mistake in AI visibility tracking is only testing your own name. "What is [brand]" tells you almost nothing, because people who already know your name were never the problem. Build 20 to 30 prompts in four groups:
Brand prompts (5 to 7). "What is [brand]", "Is [brand] reliable", "[brand] reviews". These catch misinformation and sentiment.
Category prompts (8 to 10). "Best [category] for [audience]", "Top [category] tools in [country]". These are the money questions where recommendation lists get formed.
Competitor prompts (5 to 7). "[Competitor] alternatives", "[Competitor] vs [brand]". Buyers ask these late in their decision.
Problem prompts (5 to 8). The questions people ask before they know the category exists: "How do I [problem your product solves]".
Write prompts the way customers talk, not the way your marketing does. Nobody asks ChatGPT for "an innovative omnichannel solution".
Step 3: Run the prompts on a schedule
For a manual baseline: open a private browser window, run each prompt once per model, and paste the answers into a spreadsheet with the date. Private mode matters, because your chat history personalizes answers and flatters your brand.
Two things will frustrate you immediately. Answers vary between identical runs, so a single pass per prompt is statistically thin. And by week three, running 25 prompts across 3 models by hand is a half-day job that quietly stops happening.
This is the point of automation. A tracking platform runs every prompt against every model daily, stores the full responses, and keeps the history. Whichever tool you use, the schedule matters more than the tooling: monthly snapshots hide the trend you are looking for.
Step 4: Score the answers consistently
Reading responses is not measuring. Use a fixed scoring model so week 12 is comparable with week 1. The core metric:
AI Visibility Score = (prompts where your brand appears ÷ total prompts tracked) × 100
If you track 25 prompts and appear in 4 answers, your visibility is 16 percent. Alongside it, record:
Position: when the model lists brands, where do you sit? First and seventh are different worlds.
Sentiment: is the mention positive, neutral or critical? A weighted version works well in practice: +3 for a positive accurate mention, +1 neutral, -1 for a negative or inaccurate one.
Share of Voice: your mentions divided by all brand mentions in your category, times 100.
Citations: does the model link your site, or describe you from third-party sources only?
Trackbase calculates all five automatically per model per day. Doing it by hand works for the baseline, and that first spreadsheet is genuinely worth the effort: it turns "I think we are invisible" into "we appear on 2 of 25 prompts, both times behind three competitors".
Step 5: Diagnose why you score what you score
The score is the thermometer. The diagnosis lives in the sources. For every prompt where you are absent, look at what the model cited instead. Patterns emerge fast:
The same three comparison articles keep appearing, and you are in none of them. That is an outreach list, not a mystery.
Models quote review platforms, and your last G2 or Trustpilot review is from 2024.
A competitor blog answers the exact question in the prompt, with a clear structure and a recent date. You have a content gap with a specification attached.
The model confuses you with a similarly named company. That is an entity problem: tighten your schema markup, your About page and your directory listings.
Perplexity and Google AI Overview show sources openly, which makes them the best diagnostic models. The Trackbase source view aggregates this across all models and ranks domains by citation frequency, which turns an afternoon of clicking into a sorted list.
Step 6: Build the weekly routine
Tracking only pays off as a loop. A routine that works and takes under an hour:
Monday: check the dashboard, or your spreadsheet, for movement. New mentions, lost positions, sentiment shifts.
Note anything that moved more than a few points and find the cause in the responses. A drop usually traces to a source change.
Pick one action from the diagnosis: one article to write, one listicle to pitch, one review push, one schema fix.
Monthly: review the prompt set itself. Add prompts for new products, retire ones that no longer match how customers search.
Expect the first visible movement after six to eight weeks, and plan on three to six months for structural change. Models refresh at different speeds, and the third-party content that feeds them takes time to spread.
FAQ
How do I see what ChatGPT says about my company right now?
Ask it directly, in a private window: your brand name, your category question, and "[brand] reviews". That gives you a snapshot in five minutes. The snapshot is the easy part; the trend is the useful part.
Why does ChatGPT give a different answer every time?
Language models are probabilistic and sample their output. The list of brands for the same question can differ per run. That is exactly why single checks mislead and daily averaged tracking is the standard.
Can I track ChatGPT visibility for free?
Yes, manually, with the spreadsheet method in steps 2 to 4. Free checker tools give you a quicker start. The costs come later, in the hours the routine takes every week.
Does my chat history influence what ChatGPT says about my brand?
Yes. If you use ChatGPT while logged in, your history shapes answers, usually in your favor. Always test in a private window or through the API.
Which metric matters most?
If you pick one: presence rate on your category prompts, meaning the money questions. You can have a perfect brand page and still lose every "best tools for X" answer, and those answers are where buying decisions form.
How is this different from Google rank tracking?
A ranking is a position in a list of links. An AI answer is a synthesized recommendation with three to five winners and no page two. Being absent is invisible in your analytics, which is why you have to measure at the model, not on your site.
How many prompts should I track?
Twenty to thirty covers most SMB cases well. Below fifteen, single answers swing your averages too hard. Hundreds only make sense for enterprises with many product lines.
When should I switch from a spreadsheet to a tool?
When the routine breaks, which in our experience happens around week three. If the choice is between imperfect manual tracking that actually happens and a tool, keep the spreadsheet. If it stopped happening, that is the signal. Trackbase starts at €79 per month with a 7-day trial, and the setup takes about ten minutes.
