Prompt tracking: why your first prompt list is wrong

Prompt tracking is only as good as the list behind it. The first list of prompts you write will be wrong, and it will be wrong in a predictable direction. You will write the questions you wish people asked — the ones where your product is obviously the answer, phrased the way your marketing site phrases things.

Real questions are messier, and they rarely contain your category's proper name. Getting them right matters more than anything else you set up, because every visibility figure is calculated from them.

Start from the problem, not the product

Someone who ends up buying a coffee subscription did not begin by asking a model to compare coffee subscriptions. They began somewhere further back: the office coffee is undrinkable, or they have run out again on a Tuesday, or a cafetière turned up as a birthday present and now they need something to put in it.

Write those down first. The purchase-intent questions matter, but they are the last few steps of a longer walk, and they are also the ones every competitor has already thought of.

Write prompts the way people ask them

Models are asked things in the second person, with contractions, half-formed and often slightly rude. "What's actually worth paying for" is a real prompt. "Which subscription coffee service offers the best value proposition" is a sentence from a slide deck.

If you would feel self-conscious saying a prompt out loud, it is probably not a prompt. A keyword is written for a search box, a prompt for something that answers back, and that difference is most of what separates AEO from SEO.

Track the unflattering questions too

The prompts you least want to track are the ones worth tracking most. Anything shaped like "is X worth it", "X alternatives", or "why is X so expensive" is being asked about you already, and the answer is being assembled whether or not you are watching.

Finding out what a model says when someone asks whether you are any good is uncomfortable exactly once. After that it is just information.

How many prompts to track

A list of four hundred prompts is a list nobody looks at. The point of tracking is that you notice when something moves, and you cannot notice movement across four hundred rows.

Start with something you could read over a coffee — a couple of dozen, spread across the awkward early questions, the comparison questions and the unflattering ones. Add prompts when you have a reason to, not to fill a quota.

The count matters less than the spread. Twenty prompts that ask the same thing in slightly different words will tell you one thing twenty times. Twenty that come at the problem from different directions will tell you twenty things, and when one of them moves you will know which part of the journey it belongs to.

What a good list looks like

Most useful lists cover three kinds of question.

Awareness questions start from the problem: the undrinkable office coffee, the empty cupboard on a Tuesday. Nobody asking them has a brand in mind yet, so they show whether you are in the running before anyone has drawn up a shortlist.

Comparison questions are where the shortlist gets drawn up. "Is a coffee subscription better than buying beans at the supermarket", "best coffee subscription for someone who only drinks decaf". The brands named here are being weighed against each other, and the order they come in matters.

Purchase questions come last: "which coffee subscriptions let you pause deliveries", "cheapest way to get fresh beans delivered". They are the obvious ones and the most contested, so they belong on the list, but they should not be most of it.

Then there is branded versus unbranded. An unbranded prompt never mentions you, which makes it the honest test: either the model brings you up or it doesn't. A branded prompt, such as "is Acme worth it", names you by definition, so it can't tell you whether you are visible. What it can tell you is what gets said about you once someone already knows your name. Keep the two apart. Referenced leaves branded prompts out of the visibility score and reads them for sentiment instead, because counting them would pad the number with questions you were always going to appear in.

How often to re-run prompts, and when to change the list

Answers move without anyone touching them. Ask the same question twice, as anyone who has audited a single answer by hand will have seen, and the brands named can change, so a prompt checked once tells you very little. Tracking means asking the same question over and over, so that a real change stands out against the usual wobble. Referenced re-runs every active prompt about once a day.

The list itself should change much less often. Every prompt you swap out ends the line it was drawing, and a list that gets rewritten every month never builds up enough history to show a trend. Revisit it every quarter or so, or when something real changes: a new product, a new market, a competitor you had not been watching.

When you do revisit it, look first at the prompts whose answers never name you or anyone you track. Sometimes that is a gap worth closing. Sometimes it means the question belongs to a different category from yours. If you are not sure which, pause the prompt rather than deleting it, so you can bring it back later.