Track what a buyer asks before they know you.
The tracked set decides what the number means. A set written to flatter you will flatter you, and a set that drifts quietly ends the series it was meant to protect.
Track buyer questions that do not contain your name, spread across four intents: shortlist, comparison, problem and local. Start with five, written the way a buyer types them, and freeze the wording. Add one question that does name you, to check what assistants say about you, but keep it out of the share of voice.
The intents, and how much each counts
Our question library sorts every question into one of five intents. Shortlist questions, such as "What are the best project management tools right now?", and local questions carry full commercial weight, 1.00, because that is the buying moment. Comparison questions carry 0.80, because the buyer is already evaluating a narrower field. Problem questions, such as "What should I ask before buying project management tools?", carry 0.60, because early questions do not always name vendors.
The fifth is self-description, "What is X and what do they do?". It carries 0.00, because it measures whether the answer about you is accurate, not who wins. These weights are a judgment, not a measurement. Another team could choose differently, which is why the weights are stored with each scan and shown in the report.
A starter set for a category such as project management tools might read: "What are the best project management tools right now?", "Which project management tools do small teams usually pick?", "How do the leading project management tools compare?", "What should I ask before buying project management tools?" and "Which project management tools are used most in Colombia?". Two shortlist, one comparison, one problem, one local.
Why your own name stays out of the share
A question containing your name measures recall, and an assistant recalls any company that exists. Buyers who already know your name are not the ones you are losing. Generated questions never name the company, apart from the single self-description question, and they are written in the language of the buyer, not the company.
The language is the buyer's too. A Spanish-speaking buyer gets a Spanish question, because engines draw on different source ecosystems depending on the language. A local question needs a country to read as a question at all, so a set built for a global market skips those instead of inserting the word "global", which nobody types.
Frozen wording is the whole series
Our questions come from templates, not from a model, so the same category always yields the same set and two companies in one category get identical text. New questions are appended and never inserted, because the order is the identity of a set. Reordering would hand an existing brand a different set under the same version and break comparability.
The same rule governs a spreadsheet. Reword one question between months and you have started a new series without noticing. Change wording only by starting a new sheet.
A bad question looks fine and still gets billed
In our own intake, a category typed as the single word "Delivery" was slotted into "Who are the top Delivery for a startup?" and "What are the best Delivery right now?". Both were sent and billed. ChatGPT read the first as a search for a head of delivery and named large consulting firms. It read the second as food delivery in Camden, New Jersey. Nobody has ever typed either.
The check that catches this is structural and needs no model: a bare singular noun in a slot that needs a plural phrase leaves a visible seam. Generated questions are tested for it before use. A set you write by hand needs the same test. Read each question aloud and ask whether a buyer would type it.
Two other shapes fail the same test. A question in the company's first person, such as "our platform", is the company's own copy wearing a question mark. A question that opens with an imperative, such as "Find" or "Buy", is an instruction, not something a buyer asks. Both are rejected for the same reason: no buyer typed them.
A question that names nobody is not a loss
Sometimes an assistant answers and names no brand at all. We count that cell as uninformative and report it separately, not as a zero for you. It means the question did not discriminate, not that you lost it. Scoring it as zero punishes a brand for a badly written question.
A well-built set therefore reports how many of its cells were uninformative, so a reader can tell a set that discriminates from one that returns the same generic list every time. The prompts are visible and editable in the product, so you can read exactly what was asked before you trust a figure built on it.
How many to track
Five is enough to start, and it is what the free scan runs: five questions, each asked ten times, which is 50 calls, in under ten minutes. More questions widen coverage. More samples narrow the band around each one. They are separate dials. A slot is one question in one market, so fifteen questions in one country and five questions across three countries cost the same.
The short answers
What prompts should I track for AI visibility?
Buyer questions without your name, across shortlist, comparison, problem and local intents. Add one self-description question to check accuracy, and keep it out of the share of voice.
Should I track my brand name as a prompt?
Once, as a self-description question, to see whether the answer about you is accurate. Not as visibility, because an assistant recalls any company that exists and the result flatters you.
How many questions do I need?
Five to start, which is what the free scan runs. Add more for coverage and repeat each one for a tighter band. Three repeated readings is the minimum before any change is reportable.
Why do some intents count for less?
Problem questions carry 0.60 because early-stage questions do not always name vendors, against 1.00 for shortlist and local questions. The weights are a stated judgment, shown in every report.
Can I reword a question that is not working?
Only by starting a new series. Rewording in place ends the old comparison without anyone noticing, so the next reading cannot be compared with the last.
When would you tell me not to track at all?
When your category is so small that assistants return the same two names for every phrasing and one is yours. Run the free scan, confirm it, and spend the money elsewhere.
Find out in under ten minutes
One free check, no account and no card. You get the answers the assistants gave, not a score.
Every figure quoted here comes from a real question put to a real assistant, saved with the date and the assistant that answered it. This piece was written to answer the search "what prompts should i track for ai visibility".