Prompt research for AI search, how to choose the prompts to track

What is prompt research for AI search?

Prompt research is choosing the questions you will track in AI engines. Each question is a prompt, and the full list is your prompt set.

It is the AI search version of keyword research, with one big difference. AI engines do not publish what people ask them. There is no search volume for a prompt. So the job is to get as close as you can to real buyer questions, using the sources you do have.

Get this right and tracking tells you something useful. Get it wrong and you can look highly visible for questions nobody asks, while missing the ones that bring in sales.

How are prompts different from keywords?

Prompts are longer and more specific. People type short phrases into Google. They ask AI engines full questions, often with context.

Your keyword list is still a good starting point. Take each important keyword and ask how a buyer would put it as a question to a friend. That is usually close to how they ask an AI engine.

Where do the best prompts come from?

From the people who buy from you. Here are the sources we find most useful, in rough order of value.

Sales enquiries and call notes

This is the richest source and the most overlooked. Every enquiry form, email and call note holds questions in the buyer's own words. They come right before a purchase, so they are the questions that matter most.

Ask your sales or reception team for the ten questions they hear most often. Read through a month of enquiry emails. Strip out personal details and keep the questions.

People Also Ask

The People Also Ask box in Google shows questions related to a search. Search your main services and note every question. Click a few to open more. These show the follow-up questions people have around a topic.

Google autocomplete

Type the start of a question, such as "how much does a" or "best dentist in", and note the suggestions. Google says autocomplete is based on real searches, showing common and trending ones. It also reflects location, so suggestions differ between towns.

Reddit threads

Reddit shows the questions people ask each other, including worries they would not put to a salesperson. Search your topic with the word reddit, or read the subreddits that come up in AI answers. The guide to Reddit and AI search covers finding the right communities.

Search Console

Google Search Console's Performance report shows the queries that bring people to your site. Filter for longer queries and ones that start with how, what, which, best or can. These are questions people already ask about your topic.

How do you turn raw questions into good prompts?

Rewrite each one as a buyer would ask it, and keep it neutral. A few rules help.

1. Keep the buyer's words. If customers say "worktops", do not change it to "work surface solutions".
2. Do not lead the engine. "Why is Example Windows the best choice" is a leading prompt. "Who are the best kitchen fitters in Leeds" is fair.
3. One question per prompt. Split questions that ask two things, unless buyers really ask them together.
4. Add useful context. "For a three bedroom semi" or "for a small agency" makes a prompt more realistic.
5. Remove duplicates. Several phrasings of the same question add cost without adding much.

How should you group your prompts?

Tag each prompt by service, buying stage and location. That lets you see where you are strong and where you are missing.

Buying stages are simple to set up.

• Discovery. "Best", "top" and "recommend" questions, where buyers look for a shortlist.
• Comparison. "A versus B", "is it worth it" and "pros and cons" questions.
• Detail. Price, timing, finance, guarantees and process questions.
• Brand. Questions that name you, used to check how you are described.

Most prompts should sit in the first three groups. A good rule of thumb is that no more than one in ten prompts names your brand.

How do local modifiers work in prompts?

Put the place in the prompt. "Best dentist near me" gives unreliable results in tracking, because a tracking tool has no real "me" to locate. "Best dentist in Harrogate" gives a clear, repeatable question.

For a business in several towns, write a core set of prompts and repeat it for each one. Then you can compare towns side by side. You will often find you are well known in one place and missing in the next. The local SEO and AI search guide covers what to do about that.

Use the place names buyers use. People might say "West Yorkshire" for some searches and a specific town for others. Include both where it makes sense.

How many prompts do you need?

Most businesses do well with 30 to 100 prompts to start. That is enough to cover the main services and stages without becoming a list nobody reads.

A rough guide by business type follows.

• A single-location service business, 30 to 50 prompts.
• A business with several services or a few locations, 50 to 100 prompts.
• A multi-location brand, a core set of 30 to 50 repeated per market.
• A B2B software company, 50 to 150 prompts across use cases, industries and competitors.

More prompts cost more to run and take longer to review. It is better to track 60 well chosen prompts each week than 600 once a quarter. Repeat runs matter, since answers vary between runs.

What mistakes make a prompt set less useful?

Most problems come from writing prompts for the result you hope to see. A few are very common.

• Too many branded prompts. They make visibility look high, because engines almost always mention a brand you name.
• Leading questions. Prompts that assume the answer tell you nothing about how buyers find you.
• Only easy wins. Tracking only the niche questions where you already lead hides the big ones where you do not.
• No competitors in mind. If you never check who gets named instead of you, you miss the most useful part of the data.
• Constant changes. Rewriting the set every month means you can never compare one month with the next.

A quick test helps. Show the list to someone in sales and ask whether these are the questions buyers really ask. If they start adding questions, your list is not finished.

How do the results feed back into content?

The prompts where competitors are named and you are not are your content gaps. Look at which pages the engines cite for those prompts. Then ask whether you have a page that answers the same question as well.

Sometimes the fix is a better page. Sometimes it is a listing on a site the engines trust. Either way, your prompt set is doing its job when it points to specific work. The guide on how to measure AI visibility covers how to read the results, and an AI visibility audit is a good way to test a new prompt set by hand first.

How often should you review the prompt set?

Every quarter is a good rhythm. Keep a stable core that never changes, so your trend lines mean something. Then add and retire prompts around it.

Add prompts when you launch a service, enter a new area or hear a new question often. Retire prompts for services you have dropped. If a group of prompts shows you are always named first, it may be time to add harder questions in that area.

How Axiom GEO helps

In AI visibility tracking you keep a prompt set per brand and run it on a schedule across the engines you choose. Prompt suggestions propose new prompts to fill gaps in your coverage. On the Premium plan and above, QA Finder pulls People Also Ask questions, Google autocomplete and Reddit questions, and extracts questions from pasted enquiries without storing the text. Local AI visibility runs the same prompts per country or region and splits share of voice by market.