Ask your sales team which prompts put your brand in front of buyers last month. You will get a shrug, or a guess. Buyers type thousands of category questions into ChatGPT, Perplexity and Gemini, and the answers name a handful of vendors. Whether you are one of them is not visible in any analytics tool you own today.
That blind spot has a cost. A buyer who asks "what are the best tools for X" and gets three competitor names never visits your site, never shows up in your funnel, and never appears in a lost-deal report. Fixing it starts with one question. Which prompts mention my brand, and which ones skip it? Teams that track LLM visibility at the prompt level can answer that in a week.
Why the prompt is the unit that matters
Keywords told you what people searched. Prompts tell you what people want to decide. "CRM" is a keyword. "Which CRM works best for a 20-person agency that sells retainers" is a prompt, and the model answers it with a shortlist.
Your brand can dominate one prompt and vanish from the next one, even when the two look almost the same. That is why a single "are we in ChatGPT" spot check tells you nothing. You need a map.
The four prompt types to track
Most brand-relevant prompts fall into four groups. Cover all four or your picture will skew.
Category prompts
Broad shortlist questions such as "best project management software for agencies". These drive discovery. Appear here and you enter the consideration set.
Comparison prompts
"X vs Y" and "alternatives to X". Buyers use these late in the process. Missing here costs you deals that were nearly closed.
Problem prompts
"How do I stop losing track of client approvals". No product name, just pain. Models answer with a method and then a tool. Being the tool named after the method is a strong position.
Brand prompts
"Is [your brand] any good" and "[your brand] pricing". Here the model describes you directly, so accuracy and sentiment matter more than presence.
How to find which prompts mention your brand
You can do this by hand to learn the shape of it, then automate it once you see the value.
1. Build the prompt set
Take your top 30 commercial keywords and turn each into the question a buyer would ask. Add comparison and problem versions. Pull real phrasing from sales calls, support tickets and community threads. Fifty to one hundred prompts is enough to start.
2. Run it across engines
Run every prompt on ChatGPT, Perplexity, Gemini and Claude. Answers differ by engine, sometimes sharply. A brand that wins on Perplexity can be invisible on Gemini.
3. Log three things per prompt
Were you mentioned. Were you cited as a source. Who else was named. That last column is the one most teams skip, and it is the most useful.
4. Repeat the runs
One run proves little because answers shift between runs. Run each prompt several times and record how often you appear. A mention rate of 40 percent beats a screenshot of one lucky answer.
Reading the results
Sort the sheet three ways and you will see where the work is.
Prompts where you always appear. Protect these.
Prompts where you sometimes appear. Cheapest wins.
Prompts where competitors appear and you never do. Your gap list.
For every gap prompt, check which sources the engine cites. Those pages are your target list. Get listed there, earn a mention there, or publish a clearer answer than the one the model is pulling from.
Why spreadsheets stop working
A 100-prompt set across four engines, run five times each, is 2,000 answers a cycle. Multiply by weekly re-runs and the sheet collapses. This is the point where a dedicated LLM brand visibility tracker earns its place, because it runs the set on a schedule, flags changes and shows share of voice against named competitors. If you want prompt-level data without the manual work, AI visibility tracking is what we built for it.
Three mistakes that distort the data
Testing only branded prompts makes you look great and tells you nothing about discovery. Testing from a logged-in account with chat history skews answers toward what the model already knows about you. And treating one answer as truth ignores that models vary run to run.
What to do this week
Write 25 prompts your buyers would really ask. Run each on two engines. Mark every answer where you are named and every answer where a competitor is. By Friday you will have your first gap list.
Want a new prompt-level breakdown like this every week, with real data from ChatGPT, Perplexity and Gemini? Subscribe to the newsletter below and the next one lands in your inbox.