How to Choose the Right Prompts to Track Your Brand in AI Search

Every AI visibility report begins with a list of prompts. That list decides everything that comes after it. If the prompts are wrong, your dashboard will be full of clean charts describing a market nobody is actually shopping in, and you will make decisions based on it for months before anyone notices.

Most teams build their first prompt list in about fifteen minutes. They take their old keyword list, add a question mark to each entry, and call it done. The result looks fine and tells them almost nothing. "Best CRM software" is a keyword. It is not how a revenue operations manager at a 60 person company asks ChatGPT for help on a Tuesday afternoon.

This guide covers how to build a prompt set that reflects real buying behaviour, how many prompts you actually need, and which ones to throw out.

Your prompt list is a measurement instrument, not a wish list

Think of your prompt set the way a pollster thinks about a survey sample. The goal is not to include every possible question. The goal is to pick a group of questions that represents the whole population of things buyers ask, so that when your visibility score moves you can trust the movement.

That reframing kills a few bad habits immediately. You stop adding prompts because you want to rank for them. You stop stuffing the list with your product name. You start asking whether the list, taken as a whole, gives you an honest picture.

A good prompt set has three properties. It covers the full buying journey rather than one stage. It uses the language buyers use rather than the language your marketing team uses. And it stays stable enough over time that you can compare this month to last month without wondering whether the change came from the market or from your own edits.

Start with the buyer, not the keyword tool

Keyword tools measure what people type into a search box. They were built for a ten blue links world, and they cap out at the length and phrasing people use when they know a search engine is listening. Prompts are longer, messier and more conversational. People give context. They mention their budget. They explain their situation before asking anything.

Compare these two:

Keyword: project management software small business

Prompt: We're a 12 person design studio switching off spreadsheets. We need something for client projects with time tracking built in, under $15 per user. What should we look at?

The second one produces a completely different answer, names different companies, and rewards completely different content. If your tracked prompts all look like the first one, you are measuring a version of your market that barely exists anymore.

Start instead with a simple question. What is the person trying to accomplish at the moment they open a chat window? Write that out as a sentence. Then turn it into a question. That is your prompt.

The four layers every prompt set needs

A balanced set covers four layers. Skipping any one of them gives you a distorted picture.

Problem layer. The buyer knows something is wrong but has not decided what category of tool or service fixes it. "Our checkout keeps dropping people on mobile, what usually causes that?" You will rarely be the direct answer here, but this is where a model decides which category of solution to recommend, and being present shapes everything downstream.

Solution layer. The buyer knows the category and wants options. "What are the best headless commerce platforms for mid market retailers?" This is the layer most teams track exclusively. It matters, but on its own it is thin.

Comparison layer. The buyer has a shortlist. "Shopify Plus vs BigCommerce for a brand doing $20M a year" or "alternatives to Klaviyo that work with Magento." Comparison prompts are the highest intent prompts in AI search and they are frequently ignored because they mention competitors and feel uncomfortable to track.

Brand layer. The buyer already knows you and wants a verdict. "Is [your brand] any good?" or "[your brand] pricing" or "why do people leave [your brand]?" These prompts tell you what the models say about you when nobody is guarding the narrative. They often surface old Reddit threads, outdated pricing and reviews from three years ago.

A reasonable split for a first list is roughly 20 percent problem, 35 percent solution, 30 percent comparison and 15 percent brand. Adjust once you see where your gaps are.

Where real prompts come from

The best prompts are already written down somewhere in your company. You just have to go get them.

Sales call recordings. Search your call transcripts for question marks. The questions prospects ask on discovery calls are almost word for word the questions they type into an AI assistant before the call.

Support tickets and live chat logs. These give you the post purchase and objection prompts you would never think of, like "does it integrate with NetSuite" or "can I export my data if I cancel."

Search Console queries. Filter for queries of six words or more. Long tail queries are the closest thing traditional search gives you to natural language, and they convert well into prompts.

Reddit, Quora and industry forums. Look at how people phrase requests for recommendations in your category. Copy the phrasing, including the awkward parts.

Your own sales team. Ask three reps to write down the ten questions they hear most. You will get thirty, with maybe eighteen unique ones, and they will be better than anything a tool generates.

Competitor review sites. The comparison pages on G2 and Capterra tell you which head to head matchups people actually consider.

The models themselves. Ask an assistant what a buyer in your category typically wants to know before purchasing. Treat the output as a starting draft rather than a finished list, because models tend to produce tidy generic questions.

Write prompts the way people actually type them

Real prompts are imperfect. They contain context, constraints and sometimes typos. Your tracked prompts should look the same, because phrasing changes the answer.

Include the details that shape a recommendation. Company size, industry, geography, budget, existing tech stack, timeline. "Best ERP" and "best ERP for a European manufacturer with 200 employees already using Dynamics" return different lists of vendors, and only one of those lists matters to you.

Use natural verbs. People write "help me choose," "what should I use," "which one is worth it," "is it worth switching." They do not write "top 10 best solutions 2026."

Vary the format. Some prompts should be direct questions. Some should be statements followed by a request. Some should include a constraint the model has to work around. This variety matters because assistants respond differently to different structures, and a set built entirely from one template will overstate how consistent your visibility is.

Add the modifiers that actually move answers

Once you have a core list, create variants by layering modifiers. This is where prompt tracking gets genuinely useful, because modifiers reveal where you are strong and where you disappear.

Useful modifier types include location ("in Germany," "for the UK market"), company stage ("for a startup," "for enterprise"), role ("as a CTO," "for a marketing team without developers"), price sensitivity ("on a tight budget," "free options"), and technical constraint ("that works with Shopify," "with an open API").

You do not need every combination. Pick the three or four modifiers that genuinely change who wins your deals, and apply them to your ten most important prompts. That gives you thirty to forty high value tracked prompts without an unmanageable list.

How many prompts do you need

Fewer than most people think, and more than most people start with.

For a single product in a single market, 40 to 60 prompts gives you a stable read. For a company with three product lines across two regions, expect 150 to 250. Past roughly 300, you are usually adding noise rather than signal, and the reporting becomes so heavy that nobody looks at it.

The number matters less than the stability. Changing a third of your prompts every month destroys your ability to see trends. Lock a core set of prompts that you will not touch for at least six months. Keep a smaller experimental set alongside it that you can rotate freely.

Also run each prompt more than once. Answers vary between runs even with identical wording, so a single check on a single day is a snapshot with a wide error bar. Multiple runs across several days give you something closer to a real rate.

Prompts worth leaving out

Prompts nobody asks. If a phrase only exists because a keyword tool assigned it 90 monthly searches, skip it.

Prompts where you can never win. If a question is about a use case your product genuinely does not serve, tracking it just adds a permanent zero to your average and hides real movement elsewhere.

Prompts that are really navigational. "[Your brand] login" tells you nothing about visibility.

Overly broad prompts. "What is marketing automation" will be answered by encyclopedic sources for the foreseeable future. Definitions are worth owning as content, but they make poor tracking prompts.

Prompts written to flatter you. If a prompt includes so many qualifiers that you are the only possible answer, you have written an ad, not a test.

Group prompts into clusters so the data means something

A flat list of 200 prompts produces one number. Clustered prompts produce insight.

Group by buying stage, by product line, by persona, by region, and by competitor. Then look at visibility per cluster. The interesting finding is almost never "our overall score is 34 percent." It is "we appear in 61 percent of comparison prompts but 9 percent of problem prompts," which tells you exactly what content to build next.

Clustering also makes reporting survivable. Your CMO does not want 200 rows. They want six clusters and a trend line.

Refresh the list on a schedule, not on impulse

Review quarterly. At each review, ask three questions. Have any new competitors appeared in answers that you should now track head to head? Has our product changed enough that new use cases deserve prompts? Are any prompts consistently returning irrelevant answers, which usually means the phrasing is off?

Retire prompts by moving them to an archive rather than deleting them, so you keep the historical comparison if you ever want it back.

A 60 minute workflow to build your first list

Set a timer and work through this in order.

First ten minutes: write down the five problems your product solves, in the customer's words, not yours.

Next fifteen: pull twenty real questions from sales calls, support tickets and long tail Search Console queries.

Next ten: list every competitor a buyer might realistically shortlist against you, then write one comparison prompt for each.

Next ten: write five brand prompts, including at least two uncomfortable ones about pricing, complaints or reasons to leave.

Next ten: pick three modifiers that matter in your market and apply them to your ten strongest prompts.

Final five: sort everything into clusters and mark which prompts belong to your locked core set.

You will finish with something in the range of 50 to 70 prompts that reflect how people actually buy. That is a real measurement instrument, and it will be more useful on day one than a list of 500 keywords with question marks glued on.