How to choose the right prompts to track in AI search

So getting the right prompts is critical to ensuring your tracking tells you anything useful.

This guide shows you how to find those prompts, sort them and know when to add more.

Prompt tracking means running a fixed set of questions through AI assistants on a schedule, then recording what comes back. Does your brand get named? Which competitors appear? Which websites get cited as sources?

It's the AI version of rank tracking. But a prompt isn't a keyword.

One thing to note, no one has access to a database of exact prompts that users type into the AI models like ChatGPT. Unless the big providers like Anthropic, OpenAI and Gemini release that data, no one has access to that data.

Keep in mind: prompt tracking is one part of AI visibility tracking. The prompts are the input. Mentions, citations and share of voice are what you measure.

Once your prompts are running, the same answers let you see which competitors AI recommends.

Most teams spend weeks choosing a tool. Then they spend an afternoon writing the prompts.

That's backwards.

Three mistakes we see coming up again and again:

• Too many branded prompts: What is [your brand]? will name you almost every time. It makes your visibility look great and tells you nothing about how buyers find you.
• Keywords pasted in as prompts: crm software uk isn't how anyone talks to ChatGPT. You end up measuring a question nobody asked.
• One check, one conclusion: you ask ChatGPT once, see your brand, and tick the box. In our data, the same prompt named at least 64% more brands across a week of runs than in any single answer. One answer does not give you any meaningful data to act on.

Your tool measures. Your prompt list decides what's worth measuring.

Your keyword list should be just one data point used to craft your prompts but it should not be the only source.

You already have the data you need in your business it is just scattered in different departments. You just need to collect them.

Here's where to look:

• Sales calls: your sales calls and data is a goldmine. The questions prospects ask on a first call are often the same ones they ask ChatGPT the night before.
• Support tickets: these show what customers didn't understand before they bought. That's a prompt someone else is asking right now.
• Search Console: filter for queries of seven words or more. These are people already typing full questions into Google.
• Reddit and forums: search your category and copy how people describe the problem, word for word.
• AI follow-up questions: ask ChatGPT or Perplexity your main question and note the follow-ups it suggests. AI also splits one question into several searches behind the scenes. We explain how in query fan-out explained.
• People Also Ask (bonus one): the people also ask questions in Google also provide valuable insights which you can use.

Once you have your prompt research, which should be a list of potential prompts, it's time to organise them. First by topic: these should be the areas that are important to your business. They will be different for every business, and only you can decide what they are. However, you must get them right, because topical coverage across your core areas is crucial. For a general project management tool, that might be task management, resource planning and integrations.

Then by stage. Buyers ask different questions as they get closer to buying. You want to ensure you get good coverage across your customer buyer funnel.

Most teams only track the comparison stage. Best X for Y prompts feel like the money prompts.

However, discovery prompts are where AI decides which brands get into the conversation at all. Miss those and you're not considered when the comparison happens.

Aim for every topic to have prompts at all stages of your own funnel.

Track the whole journey, not just the shortlist.

A branded prompt names you. An unbranded prompt doesn't.

• Branded: Is [your brand] good for agencies?
• Unbranded: What's the best project management tool for agencies?

Branded prompts matter. They show how AI describes you, and whether it gets your pricing or features wrong.

However, they don't show whether you get found. Your buyers start with the unbranded question. If AI doesn't name you there, they never reach the branded one.

Here's how to split it:

• At least 70% unbranded: these show whether you're in the conversation at all.
• Up to 30% branded: these show what AI says about you once you're there.
• Report them separately: branded prompts nearly always name you, so mixing them in makes your visibility look better than it is. AIS Insights will tag these for you.

Unbranded prompts tell you if you're found. Branded prompts tell you what's said once you are.

People don't talk to ChatGPT in keywords. They explain their situation.

So add the context a real buyer would give: their role, team size, budget, location or the problem they're stuck on.

Keep in mind: once a prompt is live, don't keep rewording it. Change the wording and you will pollute your data. You can't tell if your visibility moved or your question did.

If a prompt needs a big change, retire it and add a new one.

We recommend keeping a core set of prompts that don't change, plus a few spare slots for adding or testing new ones.

Write it once, the way a buyer would say it. Then leave it alone.

This is one of the most common questions people ask when they start tracking prompts, and it's a fair one.

It depends on a few things. If you're just testing a tool, start with 3 to 5 topics and 30 to 50 prompts. We'd always say have at least 10 prompts per topic.

If you're a retailer, you'll need a lot more. You could need thousands of prompts to cover all your departments.

The goal is to think about the topics that matter most to your business and make sure you have coverage across all of them.

Once you are set up:

• Add a topic only when you'd act on it: if a result wouldn't change a page, a campaign or a sales pitch, don't track it.
• Give new prompts two weeks: in our data, unstable prompts needed 12 to 14 daily runs before the result settled. Don't judge a prompt on its first few days.
• Don't cut prompts that never move: some vendors tell you to delete them. That just dresses up your numbers. First work out why you're not gaining visibility on that prompt. The only time we suggest removing one is when the AI models almost never name or cite any brand for it.

Start with the prompts you'd act on. Add more when you've acted on those.

Ask ChatGPT the same question twice and you won't always get the same answer.

AI assistants write each answer fresh. They search again, pick sources again and decide again which brands to name.

We wanted to know how much that matters, so we looked at our own data. Here's what 170,893 daily runs of 543 prompts showed over 12 weeks:

• About half of prompts are unstable: 53% named the brand in some runs and not in others.
• A mention doesn't carry over: a brand named in one answer was missing from the next run of the same prompt about 1 time in 7.
• Sources churn: two runs of the same prompt in the same week shared fewer than a third of their cited sources. More than half of cited domains showed up in just one run.
• First place isn't safe: the brand gained or lost the top spot in about 1 run in 7.
• One answer shows part of the field: across a week, the same prompt named at least 64% more brands than any single answer did.

It happens on every platform we track:

So what does that mean for your prompts?

It means a single check tells you almost nothing. You ran it once, you saw your brand, but the next run might not show it.

Instead of reading single answers, read trends. Run each prompt daily and look at the mention rate over two weeks or more. In our data, unstable prompts needed 12 to 14 runs before the rate settled. That is why you need a tool like AIS Insights.

One answer is an anecdote. Two weeks of answers is data.

How we measured this: AIS production data, 10 August to 1 November 2026. A prompt counted on a platform for a given week only if it ran at least five times that week. Internal accounts are excluded. Citation figures leave out Copilot, and ChatGPT citations after 8 October, because of known data capture issues.

Track the platforms your buyers use, not every platform that exists.

For most brands, that means starting with a few:

• ChatGPT: the assistant most people mean when they say I asked AI. See how to track your brand in ChatGPT.
• Google AI Overviews: it sits on top of normal Google results, so your buyers see it even when they didn't choose AI.
• Google AI Mode: your buyers already use Google for complex, research-heavy questions. We also think AI Mode will become the default way to search on Google.

Then add the others when you have a reason:

• Gemini, Copilot and Claude: worth tracking if you sell B2B.
• Perplexity: useful for technical and research-led audiences.

Use the same prompts on every platform. That's how you compare like with like.

Same prompts, every platform. Otherwise you're comparing two different questions.

Not directly. AI companies don't publish how often each question gets asked.

Some tools estimate it. Those estimates can help you rank ideas, but treat them as a rough signal, not a fact like Google search volume.

In practice, demand shows up in other places: the questions in your sales calls, the threads on Reddit, the queries in Search Console. If real people keep asking it, it's worth tracking.

Keep a core set of prompts that doesn't change. Review it once a year if you need to, and only change it if your core business has changed.

The trick is to keep most of your list fixed. Trends only mean something if the questions stay the same.

Here's a simple rhythm:

• Every week: read the results. Don't touch the list.
• After a launch or campaign: add prompts for the new product, feature or market.
• Always: keep a core of at least 80% of prompts unchanged, so you can compare this quarter with last.

Change the list slowly. Read the results often.

Let's put it together.

Tallyboard is a made-up project management tool for UK creative agencies. It handles projects, time tracking and client sign-off.

Here's how its team would build a first list:

1. Research: pull questions from 20 recent sales calls, the last quarter of support tickets, Search Console queries of seven words or more, and two agency subreddits.
2. Group: sort them into three topics that match what the product does and what sales talk about.
3. Balance: 10 prompts per topic, with 7 unbranded and 3 branded in each. That's 70% unbranded.
4. Spread: cover all four buying stages in every topic.
5. Platforms: run the same 30 prompts daily on ChatGPT, Google AI Overviews and AI Mode. Because Tallyboard sells B2B, add Gemini, Copilot and Claude too.

The final list:

Notice prompt 30. It's the one most teams skip, and it's the one buyers ask right before they sign. If AI repeats an old complaint or a wrong price, you want to know.

A good list reads like your sales call notes, not your keyword report.

Run your list through this before you start tracking:

• Every prompt came from a real source: a sales call, a ticket, a search query or a forum thread
• Prompts are grouped into the number of topics you'd act on
• Every topic covers discovery, evaluation, comparison and decision
• At least 70% of prompts are unbranded
• Branded prompts are reported separately
• Prompts read like full questions, with real context, not keywords
• At least one prompt asks about your downsides or pricing
• The same prompts run on every platform you track
• Every topic has at least 10 prompts, and you haven't added more than you'll act on
• You'll wait two weeks before judging any new prompt
• Your core prompts are fixed, with a few spare slots for testing new ones