The twenty questions: how to build the prompt set you track

Every AI visibility project starts in the same place. Not with content, not with schema, but with a list of questions.

If you cannot say which questions you want to be the answer to, you cannot tell whether anything you do is working. Rankings gave us that for free in search. Nobody hands it to you here. You write the list yourself.

Here is how we build one, and how you can build your own this afternoon.

Keywords are not prompts

A keyword is a fragment. A prompt is a sentence with constraints in it.

"Leggings" is a keyword. What a buyer actually types is closer to "squat proof leggings that don't go see through, under 80 dollars, for hot yoga". That sentence contains four filters: opacity, price, activity and implicitly fit.

This is the whole difference. Search rewarded you for owning the fragment. Answers reward you for satisfying the constraints. A page built around the fragment often never states a single one of them.

So write the sentences, not the fragments. Long, awkward, specific ones, the way people actually type when they are talking to something that understands them.

The five intents

Every category I have worked in sorts into the same five buckets. Use them as a checklist so you do not end up with twenty variations of the same question.

Use case. What the buyer is trying to do. Best air purifier for wildfire smoke. Mattress for couples with different sleep positions. Spices for making biryani at home.

Fit or suitability. Whether it works for their particular situation. Moisturiser for eczema-prone skin without essential oils. Joint supplement for a senior dog with stiff back legs. Cookware for an induction hob.

Comparison. Two named options, or two materials, or two approaches. Latex versus memory foam in a hot climate. Bhringraj oil versus onion oil. Cast iron versus carbon steel for a beginner.

Price. Almost always with a number attached. Good running shoes under 5,000 rupees. Eid gift perfume under AED 300. Best value organic cotton T-shirt.

Trust or authenticity. Where it came from, whether it is real, whether the claim holds. How to tell genuine handblock print. Organic certification that actually means something. Direct from farmer spice brands.

Two or three prompts per intent, per product line, and you have your set.

How many, and why more is worse

Twenty to forty. Not two hundred.

The temptation is to track everything, and it is a trap. A large set takes longer to ask, produces a number that moves so slowly you cannot tell signal from noise, and buries the ten questions that actually drive revenue in a crowd of ones that do not.

Pick the questions where being named would change your month. If a prompt gets asked constantly but the buyers asking it never convert, it belongs in a secondary list, not the main one.

How to ask them properly

This is where most self-run tracking falls apart, and it is worth being pedantic about.

Log out. Your own account has months of history telling the assistant what you like, including your own brand. Answers from a logged-in session are about you, not about your market.

Fix the location. Answers differ by country and often by city. Pick the market you sell to and ask from there, consistently. If you sell to four countries, that is four sets, not one.

Fix the schedule. Same day of the week, same rough time, every week. Answers move around. You are looking for a trend, and a trend needs a stable rhythm underneath it.

Save the whole answer. Not a tick or a cross. The full text, with the citations. Six weeks from now the useful insight will be something you did not think to record, and you cannot go back and get it.

Ask it fresh. Open a new conversation each time. Continuing a thread carries context forward and contaminates the result.

What to record

Five fields per answer. This is the entire scoring sheet.

Was the brand named, yes or no. Where did it appear, first, second, third. Was one of your own pages cited as a source. Were the facts about you correct. And, if you want one more, was the tone positive, neutral or negative.

Add the competitors named in each answer while you are there. Over a few weeks that column becomes the most interesting thing in the sheet, because it tells you who you are actually competing with in this channel, which is often not who you thought.

What does not work

Asking once, in your own account, and drawing a conclusion. I have seen brands commission a strategy off the back of a single Tuesday afternoon of prompting. Answers vary between sessions even with everything else held constant. One observation is an anecdote. Six weeks of the same twenty questions is a measurement.

Also: writing the prompt list from your keyword research export. Your keyword tool knows what people typed into a search box. It does not know what they say when they can speak in full sentences, and the difference between those two things is the whole opportunity.

Where to start

1. Write ten questions from memory, before you look at any data. The ones your customer service inbox answers every week are usually the best.
2. Sort them into the five intents and fill the gaps.
3. Ask all of them once, logged out, and record the answers properly.
4. Put a repeating half hour in your calendar for the same slot next week.

If half an hour a week is half an hour you do not have, our free self-test writes the question set for you and scores the first round.