How to choose AI tracking prompts that reflect buying decisions

Choose AI tracking prompts by starting from the decisions your buyers make, not from keywords. For each decision, write the full question a buyer would ask, tag it with an intent, a buying stage, an audience and a location, and keep it only if it passes a short set of inclusion rules. The result is an editorial prompt set: a list you chose because it represents your market. It is not a measurement of how often anyone asks those questions, and the difference matters for how you read the results.

This post gives you a starter worksheet, example prompts, and the rules we use to decide what stays on the list.

In short

• A tracked prompt is a full question your buyers might ask an AI engine, not a keyword fragment.
• Classify every prompt by intent, buying stage, theme and audience, and decide its locations, before you decide whether to track it.
• Your prompt set is an editorial choice. Search demand data can help you rank it, but it does not tell you how often AI assistants are asked a question.
• Start small and cover the decision stages evenly. Add a prompt only when it tests something the current set does not.
• Each location, audience and language you add multiplies the number of runs and the prompt slots used, so plan the grid before you fill it.

What is a tracking prompt, and how is it different from a keyword?

A tracking prompt is a complete question that a monitoring tool runs against AI engines on a schedule, so you can see whether your brand is named or cited in the answer. A keyword is a short phrase you might target in a search engine, with no question, no location and no schedule attached.

The distinction changes how you write them. "document review software" is a keyword. "What is the best document review software for a small compliance team?" is a prompt: it has a buyer, a situation and a decision inside it. In DiscoveredBy, a prompt is stored as its own row and is what actually runs against the engines you track, while a keyword is only a text label you can attach to a prompt to file it (see Prompts and Key terms). The glossary entry for prompt has the short definition.

Which buyer questions belong in a starter set?

Include the questions a buyer asks at each point on the way to a decision, from "what is this category" to "is this vendor safe to pick". A starter set that only covers "best X for Y" questions will tell you about one moment in the buying process and nothing about the others.

Work through the buyer's path in this order:

1. Problem and category questions. What is the problem, and what kind of product solves it?
2. Shortlist questions. Which products should I look at, and what are the alternatives?
3. Comparison questions. How do these two or three options differ on the things I care about?
4. Evaluation questions. What does it cost, how hard is it to implement, what does it integrate with, what are its limits?
5. Decision and objection questions. Is it right for a team like mine, and what would make me hesitate?

Enlarge image

The product uses a closed vocabulary for the same idea. Each prompt can carry an intent (informational, navigational, commercial, comparison or troubleshooting), a buyer stage (awareness, consideration or decision), a theme (competitor alternative, pricing, use case, implementation, troubleshooting, regional opportunity, comparison or other), and branding (whether the prompt names a tracked own or competitor brand, product or alias). Those lists come from the Prompts docs. Using the same labels in your worksheet means its values map directly onto the fields in the tool later.

How do you write the worksheet?

Write one row per buyer question, and fill every column before you decide whether the row goes on the tracked list. The columns force the decisions that are easy to skip: who is asking, from where, and why this question is worth a slot.

Copy this into a spreadsheet:

Buyer question (full sentence) | Intent | Stage | Theme | Branded? | Audience | Locations | Source | Keep? | Reason

What each column asks:

Two columns deserve extra care.

Audience. General is the plain, unbranded audience every prompt starts with. A persona (a named buyer profile with a short description) adds an audience block to the message sent to the engine, asking it to answer for that person. Use one only when the question genuinely differs by buyer, for example a compliance lead versus a procurement lead. Read the Personas page before relying on them: persona variants are not sent to Google AI Overviews, Google AI Mode, ChatGPT (app) or Gemini (app), which receive only the prompt text. The General variant of the same prompt still runs there.

Locations. A prompt tracks at least one country, and can also track cities. The location is information passed to the engine, and each engine decides how much weight to give it; a city target does not reproduce what a customer standing in that city would see (see Engines and measurement). City variants are not sent to ChatGPT (app), so a city-level question will not be measured there. Add a location because your buyers really differ by market, not because the option exists.

What are the inclusion rules?

A prompt earns a place on the tracked list if it passes every rule below. The rules exist to keep the set small, distinct and tied to a decision.

• It is a full question a buyer would plausibly ask. If you cannot say who asks it and when, cut it.
• It maps to a stage and a theme. A prompt you cannot classify is usually too vague to interpret.
• It tests something no other prompt in the set tests. Near-duplicates measure the same thing twice and use up slots. The Suggestions screen flags a card as "similar to a tracked prompt" when tracking it too would likely duplicate a measurement rather than add a new one.
• A good answer would change something you do. If you would not update a page, brief a writer or investigate a source based on the result, it is curiosity, not monitoring.
• It does not depend on a fact you cannot state. If you do not know what the correct answer is, you cannot judge the response.
• Its locations, audiences and languages are justified. Every combination is its own run and uses a prompt slot.
• It is neutral enough to be a fair test. A question that names your product and asks for its praise tells you what the engine does with your name, not whether you are chosen. Keep some unbranded prompts so you can see whether you appear when nobody asks for you.

Keep some prompts branded on purpose. Branded prompts (for example "Is Quillstone good for legal teams?") tell you how an engine describes you when asked directly. Unbranded prompts tell you whether you are in the conversation at all. You need both, and you should not average them into one number.

Is an editorial prompt set the same as measured demand?

No. An editorial prompt set is a list you chose. Measured demand is a count of something people actually did. The two are easy to blur, so state which one you are relying on.

Your worksheet is editorial: it reflects your judgement about your buyers. Nothing in DiscoveredBy measures how often a question is put to ChatGPT, Gemini or any other assistant, and you should not present a prompt list as a picture of what the market asks AI.

What the product can show is Google demand, when Search Console is connected. That figure is the total impressions Google recorded over the last 28 days for the search queries a prompt covers. It counts what people typed into Google, not what anyone asked an AI assistant. It is useful for ranking your own prompts against each other, since a question people already search for on Google is usually worth tracking before one nobody searches for. Three things to keep in mind, all from the Prompts docs:

• A query counts toward a prompt only when every meaningful word of the query appears in the prompt, and the query has at least two meaningful words.
• The 1 to 5 score is relative to your own project, so a 5 on one project is not comparable with a 5 on another. With fewer than three matching prompts there is no score, only the impressions.
• A dash means no query matched. It does not mean no demand; it may mean your site does not rank for that question yet.

The Suggestions queue is a third input, and a different kind again. Each card was written by a research run that read your site and built a profile of the business, and it arrives classified with a theme, buyer stage and intent, an opportunity score and a rationale. Treat suggestions as candidates for your worksheet, not as a replacement for it. Your best Search Console queries are given to the research as extra context, but a Google query never becomes a suggestion on its own.

A worked example

Quillstone sells document-review software to mid-sized legal and compliance teams. Its competitors in the example are Brieflane and Clausewise. The team drafts ten prompts and runs them through the inclusion rules. Here is a cut of the worksheet.

Enlarge image

The keepers are eight prompts. Now the slot arithmetic, using the rule that each country, audience and language combination is its own target and uses one prompt slot:

• Prompts 1, 3, 4, 5 and 6: one country each, General, As written, so 1 slot each, 5 slots.
• Prompt 2: two countries (US, GB), so 2 slots.
• Prompt 7: one country, one persona, so 1 slot. Its General variant is not being tracked here, since General was deselected in favour of the persona.
• Prompt 8: one country (GB) plus one city (London), so 2 slots.

That is 5 + 2 + 1 + 2 = 10 slots for eight prompts.

Enlarge image

The team then checks Google demand, having connected Search Console. Prompt 5 shows 1,240 impressions, prompt 2 shows 380, prompt 6 shows 95, and prompt 8 shows a dash. Only three prompts matched, so the column has just enough spread to rank them, and prompt 5 comes out on top. The dash on prompt 8 is left alone: the team keeps it, because the UK regulated-firm buyer is a market decision, not a search-volume decision. The point of the column was to order the list, not to veto it.

One more note from the example. Prompt 7 is tracked as a persona only. Persona variants are left out on Google AI Overviews, Google AI Mode, ChatGPT (app) and Gemini (app), so this prompt has nothing to run on those four. If the team needs that buyer's question tested there, they also need to keep General selected.

Common mistakes and what this cannot tell you

• Treating the prompt list as market demand. It is your editorial sample. Say so in any report that uses it.
• Reading Google demand as AI demand. It counts Google impressions only.
• Adding many near-identical prompts. Rewording a question usually adds little, and each copy uses slots. If you want to audit an existing list rather than build one, see Prompt coverage, which recommends prompts you could pause while what they see regularly is still covered.
• Only tracking branded prompts. You learn how you are described but not whether you are found.
• Multiplying locations and audiences by default. Nine targets from one prompt is easy to reach (3 countries times 3 audiences), and each one is a slot.
• Changing the set constantly. Adding or removing prompts changes what a result is measured over. Note the date of any edit so you can explain a shift later.
• Expecting a result. A well-chosen set tells you where you appear and where you do not. It does not make an engine name you.

Frequently asked questions

How many prompts should I start with?

Start with a set you can review by hand, then grow it. There is no universal right number. Your plan sets how many prompt slots you have (see Plans and limits), and remember that each location, audience and language combination uses one slot, so a small number of prompts can fill a plan quickly.

Should I use my brand name in my prompts?

Use both kinds, and report them separately. Unbranded prompts show whether you are part of the answer when nobody asks for you. Branded prompts show how an engine describes you when asked directly.

Can I just import keywords from Search Console?

Not as they stand. A query is a fragment; a prompt is a full question. Search Console queries feed the research that proposes suggestions and can rank your prompts by Google demand, but you still write or approve the question.

When is a city worth tracking?

When your buyers in that city really get a different answer to the question you care about, such as local providers or regional regulation. A city target is a location passed to the engine, not a customer standing there, and city variants are not sent to ChatGPT (app).

How often do prompts run?

Once a day, a scheduled job runs each active prompt target on every engine available on your plan that the target runs on. Accepting a prompt from Suggestions also queues a fast first run outside the daily cycle. See Your first scan for how the first run behaves.

What if the prompt I care about has no Google demand?

Keep it if it represents a real buyer decision. A dash in the Google demand column means no matching query was found, not that nobody cares.

Next step

Put your worksheet into DiscoveredBy, then review the Suggestions queue against it and track only the cards that pass your inclusion rules. Once your baseline is running, your first 30 days of AI search optimization is a plan for what to do with it, and what each number tells you explains how to read the results. For the wider features, see AI visibility.

Sign in or create an account to build your prompt set.