How to choose prompts to track in AI search

Why does prompt choice matter for AI visibility?

The prompts you track are a sample of your market. When you measure your visibility in ChatGPT, Gemini, Microsoft Copilot or Google AI Mode, you cannot see every question that people ask about your category. You pick a set of questions, run them on each engine and read which brands the answers name. Every figure you get afterwards, from mention rate to share of voice, describes that sample and nothing else.

A good sample leads to good decisions. A bad sample leads to confident conclusions that are simply wrong. If your list only contains questions that name your brand, you will look visible everywhere. If it only contains broad questions that nobody asks before buying, you will learn nothing about the moments that bring you customers.

The stakes are real, because many of these questions ask for advice. In OpenAI's study of 1.5 million ChatGPT conversations, published in September 2025, about half of messages (49%) were "Asking": people looking for information or advice rather than getting a task done. When the question is "which tool should I choose?", the answer is a shortlist of brands. Each prompt you track stands for one of those shortlists.

What are the 4 types of prompts to track?

Most buyer questions fall into four types. A useful prompt list covers all four, because each one shows a different stage of the buying journey.

Discovery prompts are the heart of a prompt list. They ask the engine to recommend, so they show whether your brand makes the shortlist when nobody has named it.

Comparison prompts name brands, often your competitors. They show how engines describe you next to the alternatives, and which pages they cite to justify the choice.

Problem prompts come earlier in the journey. The person does not look for a product yet, but the answer often suggests one. Being named here means engines connect your brand with the problem you solve.

Purchase and local prompts come last. They matter most for businesses with a physical location, a price list or a service area, because the answer can send the buyer straight to a provider.

How do you build your prompt list, step by step?

Building a prompt list takes an afternoon if you follow a simple method. The goal is a list that sounds like your customers, not like your marketing team.

1. Collect real customer questions. Read your support tickets, the notes from sales calls and the questions people ask in demos. Look at Reddit threads and forums in your category, the questions on review sites, and the "People also ask" boxes in Google for your main keywords. Write down the questions word for word.
2. Rewrite each question the way people talk to an AI. A Google query is a few keywords: "crm small agency". A prompt to an AI assistant is a full sentence with context: "Which CRM should a 10-person marketing agency use to manage client projects?" Add the details a real buyer would give, such as team size, budget, industry or location.
3. Cover every offer. If you sell several products or services, write prompts for each of them. A law firm that handles both divorces and property disputes needs prompts for both, or half of its business stays invisible in the measure.
4. Cover the four types of prompts. Check that each offer has at least one discovery, one comparison, one problem and, if relevant, one purchase or local prompt.
5. Add the questions where your competitors are named. Ask a few engines your discovery questions and note the brands that come back. Then add comparison prompts with those names, such as "[Competitor] vs [Competitor] for a small team" or "Cheaper alternatives to [Competitor]". A tracking tool saves you this step: Hawkscan lists every brand the answers recommend, including competitors you had not thought of.
6. Group the prompts by intent or by offer. Groups let you see where you are strong and where you are missing, instead of one average that hides everything.

Behind each prompt, engines that search the web often run several hidden searches, a process called query fan-out. Ahrefs found in March 2026 that only 37.9% of the URLs cited in Google's AI Overviews also ranked in the top 10 for the query itself. Whether a page gets cited depends first on how it ranks, including on those sub-queries. Precise, realistic prompts trigger the same sub-queries your buyers trigger, which is why the wording of a prompt should stay close to theirs.

Should you track branded or unbranded prompts?

Track mostly unbranded prompts, and keep branded prompts in a separate group. A branded prompt names your brand or one of your products, such as "Is [Your Brand] worth it for a small team?". An unbranded prompt only describes a need, such as "best project management tool for a small team".

The difference matters because a branded prompt almost always produces an answer about you. Putting your brand name in the prompt guarantees a mention, so it inflates your mention rate and tells you nothing about whether engines recommend you on their own. Google uses the same split in Search Console: a branded query includes "your brand name", its variations or misspellings, while non-branded queries "show how new users find your content".

Branded prompts still have a job. They show what engines say about you: your prices, your features, your reviews and your tone. Use them as a reputation check, not as a visibility score. The branded vs unbranded prompts entry explains how to read each kind.

What are the most common mistakes when choosing prompts?

Most weak prompt lists make the same few mistakes. Each one distorts the measure in a predictable way.

• Writing prompts like Google keywords. "Best CRM" is a keyword, not a question someone types into ChatGPT. Short prompts produce generic answers that favor the biggest brands of the category, whatever your real position with your buyers.
• Tracking many prompts that say the same thing. Ten versions of "best CRM for small business" weigh ten times more in your average than one question about pricing. SparkToro found that across 994 answers to very different wordings of one need, the same brands showed up 55% to 77% of the time. Engines capture the intent behind the words, so near-duplicates add cost without adding information.
• Forgetting purchase intent. A list full of problem and discovery questions shows awareness, not sales. Add the questions people ask right before they pay: prices, providers, locations, "is it worth it".
• Putting your brand name in the prompt. As explained above, a branded prompt always names you and inflates your visibility.
• Never updating the list. Your offers, your competitors and your customers' words change. Review the list every few months, and add a prompt each time sales hear a new question.

How many prompts should you track?

There is no agreed number yet. SparkToro, which studied the consistency of AI recommendations in January 2026, lists "how many are needed" among the open research questions. What the research does show is that a handful of prompts is not enough, because a single answer moves too much from one run to the next.

As a practical starting point, think in prompts per offer rather than in a total:

Then repeat each prompt, because the answers change every time. In SparkToro's study, two answers listing the same brands in the same order came up less than once in 1,000 runs, and SparkToro estimates that you need to ask a question "usually at least 60-100X" to know an AI's set of recommendations. This AI answer variability is why prompt tracking runs the same list on a schedule, every day or every week, and why your mention rate across many runs means more than any single answer. Done by hand, this quickly gets out of reach: 30 prompts asked 60 times on four engines make 7,200 answers to read. An AI visibility tool such as Hawkscan runs the list for you on a schedule and counts the mentions in every answer.

What should your prompt checklist include?

Before you start tracking, check your list against these points. A list that passes all of them gives you a sample you can trust.

• Every prompt is a full question with context, as a buyer would write it.
• The prompts come from real customer questions, not from your own keyword list.
• Each offer or service has its own prompts.
• Each offer is covered by discovery, comparison, problem and, if relevant, purchase or local prompts.
• Unbranded prompts make up most of the list.
• Branded prompts sit in a separate group.
• Comparison prompts name the competitors that engines actually mention.
• No two prompts ask the same thing in different words.
• Prompts are written in the language of each market you sell in.
• Each prompt will run many times, on every engine your buyers use.

What does a good prompt list look like?

Here is a template for a fictional company, a scheduling software for dental clinics in the United States. It has one main offer and a secondary one, online payments. The list below is a starting point of 12 prompts that you can adapt to your own business.

Notice how each prompt reads like a question a clinic manager would really ask, with the details that make an answer useful. Run on four engines every day, these 12 prompts produce 48 answers a day and more than 1,400 a month, enough to read a trend by group.

Common questions

What is a prompt in AI visibility tracking?

A prompt in AI visibility tracking is a question that a brand runs on AI assistants on a schedule to see whether the answers name it. It is usually a buyer question, written the way a potential customer would type it into ChatGPT, Gemini, Microsoft Copilot or Google AI Mode, with enough context to get a useful answer.

How is choosing prompts different from choosing SEO keywords?

Choosing prompts differs from choosing SEO keywords because people talk to AI assistants in full sentences, with their own context. SparkToro asked volunteers to write a prompt for the same need and found that, across 142 prompts, the overall semantic similarity was just 0.081. You cannot track every wording, so you track intents, while keyword research looks for exact phrases with known search volume.

Can you find out which prompts people actually ask?

You cannot see the exact prompts people ask, because AI assistants do not publish their prompt volumes. You can get close by reading your customers' own words in sales calls, support tickets, forums and reviews. The questions people type into Google, such as those in "People also ask", are another good signal of what buyers want to know.

Should prompts include a location or a language?

Prompts should include a location when location changes the answer, as it does for local services, and they should be written in the language of each market. OpenAI explains that ChatGPT search "may use an approximate location based on your IP address", and Chen et al. found that AI search services differ in their "cross-language stability". Track each market separately.

Hawkscan shows the full answers from ChatGPT, Microsoft Copilot, Gemini and Google AI Mode for every prompt you track, word for word, so you can check that your prompts bring back the kind of answers your buyers see. When you create a project, it reads your website and suggests prompts for your offers, and you choose which ones to run. Enter your domain below to get your first answers and see where your brand already shows up.