Buyer prompts: how to choose the questions worth tracking

A buyer prompt is the question somebody asks an assistant while they are deciding who to buy from. Choosing the right twenty is the hardest part of AI visibility work and the part nobody explains, because a prompt set chosen badly produces a number that moves for no reason and points at no action. A good prompt earns its place on four properties: real people ask it, the answer names companies, it is close to a purchase, and it is stable enough to compare month to month. This page gives you where to find candidate prompts, how to score them, how to phrase them, and what to leave out.

Where real buyer prompts come from

Invented prompts produce invented findings. Five sources give you the real wording, and the first two are the best.

• Your sales inbox and call notes. The questions people actually ask before they commit, in their own words, including the awkward ones about price and contract length. Read the last thirty enquiries and write down every question verbatim.
• The chat log on your website. People type to a chat box the way they type to an assistant, which makes this the closest match you own.
• Search Console queries. Filter to queries of five words or more. Long queries are phrased as questions and they transfer to assistants almost unchanged.
• People Also Ask boxes. Search your category and expand the box twice. You get the sub questions that Google already knows get asked.
• The assistant's own follow ups. Ask a broad category question and note the clarifying questions the assistant asks you. Those are the dimensions buyers vary on.

Aim for forty candidates. You are going to keep about twenty.

The five families, and why you want one of each

Family	Shape	What it tells you
Shortlist	Who are the best providers of X for a business like mine	Whether you are in the competitive set at all
Constrained category	X for a company of 30 people that needs Y	Whether your specific claim is public and readable
Comparison	A compared with B for a named job	Whether anybody except your competitor has written the comparison
Price and terms	How much does X cost, what is in a typical contract	Whether your commercial facts exist in crawlable text
Verification	Is this kind of supplier trustworthy, what should I check first	Whether your credentials and accreditations are stated anywhere but your homepage

A set made only of shortlist prompts reads like a league table and tells you nothing about why. A set with one of each family tells you which part of the work is failing.

Score each candidate out of twelve

Score every candidate prompt on four properties, nought to three each. Keep the top twenty.

• Asked (0 to 3). Did you find this wording in your inbox, your chat log or Search Console? Three if you have it verbatim from a real person, one if you inferred it, nought if you invented it.
• Names companies (0 to 3). Run the prompt once. Three if the answer names several businesses, one if it names a category of business, nought if it returns general advice. A prompt that never names anybody cannot measure whether you are named.
• Close to money (0 to 3). Three if somebody asking this is choosing a supplier this month. Nought if they are a student doing research. The distance from the question to an invoice is the whole point.
• Stable (0 to 3). Run it twice. Three if both runs name a broadly similar set, one if they disagree completely, nought if one run returns no names at all. An unstable prompt makes a chart that jumps without meaning.

That one pass through forty candidates takes about an hour and it is the single highest value hour in the whole programme. It also gives you, for free, the first reading of your baseline.

How to phrase them

• Never name your own business. This is the rule that everything else depends on. Ask about your own company and the model will talk about your company, which tells you nothing about whether it would have raised you unprompted. The full reasoning is in why you must never name your own brand in a tracking question.
• Write a sentence, not a keyword. "best crm" is a search query. "which crm should a 12 person recruitment agency use if we live in Gmail" is a buyer prompt, and it is the second one that triggers a search and produces names.
• Put exactly one constraint in. Team size, or budget, or a compliance requirement, or a location. Two constraints at once narrows the answer to a set so small that it stops being representative.
• Name the place when place matters. An assistant with no location signal answers generally, so a local business must put the town in the prompt the way a customer would.
• Keep the wording frozen. Rewriting a prompt resets its history. Park improvements in a candidate list and swap them in at the quarterly review.

What to leave out

• Branded prompts. Your own name, and also competitor names on their own. These measure recall of a name you already supplied.
• Head terms. Two word category phrases are often answered from model training with no sources attached, so there is nothing to act on when you lose them.
• Questions with one correct answer. Factual questions about standards or definitions do not name suppliers, so they cannot measure supplier visibility.
• Prompts nobody asks. If it scored nought on "asked", it is your marketing language rather than your buyer's.
• Anything you would not want screenshotted. Prompts that fish for criticism of a competitor produce answers you cannot put in a report.

How many, and how often to revisit

Fifteen to twenty prompts for a single business, run three times each across two or three assistants, every month. That is 90 to 180 readings a month, which is enough for a rate with a range around it. Multi location or multi product businesses need a set per market or per line rather than a longer single set.

Review the set quarterly and change at most three prompts at a time, noting the change date in the report so a jump can be read against it. AI Knows Us builds and runs the set across five surfaces and then writes the pages for the questions you are losing, and its free scan needs no account.

This page is prompt selection. For a ready made map of 100 named prompts with scoring rules, see the buyer question topic map. For the definition in plain words, see what a buyer question is.

Asked and answered: buyer prompts

How is a buyer prompt different from a keyword?

A keyword is the shortest phrase that gets a results page. A buyer prompt is a full sentence describing a situation, because that is how people talk to assistants. The same intent produces a two word query in Google and a twenty word question in ChatGPT.

Should I track prompts my competitors win?

Those are the most valuable ones in the set. A prompt where a competitor is named and you are not has a clear next action: write the page that answers it, and get your facts onto the sources that answer cited.

How many runs per prompt?

Three is the working minimum, because two runs of the same prompt can disagree completely. More runs narrow the range, and the range is what tells you whether a change is real.

Can I use keyword volume to pick prompts?

Volume is useful for sizing the Google side of a topic and it does not exist for assistant prompts, since no platform publishes a prompt count. Use volume as one input to the "asked" score and let your own inbox outrank it.

What if my category is too small to have prompts?

Small categories are the easiest ones to win, because few companies have written the page that answers the question. Take the five questions your last five customers asked and start there. A niche prompt with four askers a month can still be worth a contract.

What to do first

• Read the last thirty enquiries and write every question down in the sender's own words.
• Pull Search Console queries of five words or more and add them to the candidate list.
• Get to forty candidates, then score each one out of twelve using the four properties above.
• Keep the top twenty, with at least one from each of the five families, and freeze the wording.
• Run all twenty three times across two assistants this week, and save every cited source as your off site target list.