Choose prompts for your business

Turn customer questions into prompts

Begin with a question people ask before they contact or buy from you. Remove personal details and retain the decision behind their wording.

A repair company might hear about service areas, availability and what a visit covers. Use what your own team hears.

Look at written sources too. Enquiry emails, call notes, reviews and the questions on your own help pages all record how customers put things.

Strip out names and contact details before any question goes near an assistant or a shared sheet.

Keep the wording steady

Make a short list, then give it a version and a date. Don't change the wording between reviews without saying so.

When you add a prompt, start a new version and keep the old records. That way every comparison uses the same set.

Store the set somewhere shared, with the version and the date in the file name. A spreadsheet is enough.

When you compare versions, list the prompts that were added, removed or reworded. That list is the first thing a reviewer needs.

Say what you don't know

A useful prompt isn't necessarily a popular one. Don't call your list demand data unless you have a source and a method.

Note the language, the country and the intended customer beside each prompt, and say why you included it.

If a prompt came from one memorable customer, say so. One vivid enquiry is a reason to test, not evidence of a pattern.

Note too whether a prompt reflects a customer you want more of, or one you already serve. Both are fair reasons, but they are different reasons.

Build prompts from customer decisions

Name the requirement the customer needs to resolve. Keep the offer and context specific enough that you can judge whether the answer helps.

A question about availability has a different purpose from one about suitability. Label each purpose so your list covers more than one decision.

Ask whoever handles enquiries which questions cause confusion, and keep the original meaning when you tidy the wording.

Ask more than one person. Staff who take bookings hear different questions from those who handle complaints or aftercare.

Keep the customer's own phrasing where it is clear, even if it's informal. Tidying too much can change what the question asks.

Write prompts the way customers ask

So write some prompts as full questions, with the customer's circumstances included. A keyword phrase such as kitchen fitter Leeds tests something narrower.

A fuller version asks for a kitchen fitter in Leeds who can start within a month and works to a fixed quote. It carries the details a customer is weighing.

Keep each extra detail real. Add a condition only if your customers actually raise it, such as budget, timing, parking or step-free access.

Include a few short versions as well, as separate prompts. Some customers do type briefly, and the pair shows whether the wording changes the answer.

Keep your own marketing words out of the prompt. A prompt built on your tagline tests your copy rather than the customer's question.

Prompts that name you check how you're described. Prompts that don't show what an assistant says about the wider category.

Note why each prompt is in the set. Drop any that ask about services you don't offer, however tempting they sound.

Avoid leading wording, such as asking for the best or top-rated firm. Those words push the answer towards a ranking and away from the customer's actual need.

Read each prompt aloud. If it sounds like something nobody would say, rewrite it before it enters the set.

• Google guidance on AI features

Expect one prompt to become several searches

Google says AI Overviews may use a technique it calls query fan-out. That means running several related searches across subtopics and data sources to build one response.

So a single prompt can reach several topics at once. For example, a question about a boiler repair could involve prices, call-out areas and reviews.

You don't need a separate prompt for every subtopic. List the subtopics each prompt implies instead, and check that your pages answer each one clearly.

If the assistant shows the searches it ran, save them with the answer. They show which subtopics it looked into for that prompt.

Google also says AI Overviews may use different models and techniques, so their responses and links vary. Treat each as its own surface in your records.

• Google guidance on AI features

Include the prompts that lead to a choice

Include questions that explain a subject as well as questions that compare options. A buying question should describe something your business genuinely supplies.

Write these with care. Use competitors your customers actually mention, and keep the wording neutral so the prompt doesn't lean either way.

Don't build a claim about either business into the prompt. Ask the question plainly and let the answer say what it says.

Reputation prompts are another group. Questions about how a business handles complaints or delays show what an answer says about how you work.

Don't forget existing customers. Questions about aftercare, warranties or changing a booking test whether answers describe your service accurately after the sale.

Label each of these groups. Comparison, reputation and aftercare prompts answer different questions from discovery prompts, so report them separately.

Keep prompts that name you apart from those that don't. They test different things, so mixing them blurs both.

A problem prompt starts from the customer's situation, such as a leaking boiler. An options prompt asks what kinds of help exist. A business prompt names you directly.

Take a dental practice in York. Its set might hold a problem prompt about a chipped tooth and an options prompt about whitening. A business prompt could ask whether the practice takes new patients.

Add one comparison prompt naming a practice nearby, and one aftercare prompt about what to do after an extraction. That gives five prompts, each with a reason you can state.

Version the set instead of editing it quietly

Keep a dated version whenever prompts are added, removed or reworded. Compare periods using the same version wherever possible.

When you add a service, start a new version and keep the old list so earlier results still make sense.

For a before and after comparison, use the prompts that appear in both periods. Report new ones separately. Otherwise a changed list can look like changed visibility.

Record who made each change and why. A short note saves a later reviewer from guessing whether a change was deliberate.

Plan places and languages on purpose

Location can change an answer. Google says it uses your location for more helpful results, and OpenAI's web search documentation accepts an approximate user location.

If you serve several towns, each prompt can be multiplied by each town. Choose the towns where your work really comes from, and note why each one is in.

Write the place into the prompt rather than relying on where you happen to be. That keeps the test the same wherever the review is run.

Keep the country and language in their own columns, even when the place is in the wording. Filtering on those columns makes later comparisons simpler.

If you serve customers in another language, write their prompts in that language from the start. A translation of an English prompt may not match how they ask.

• Google Search Help, Understand and manage your location when you search
• OpenAI API documentation, Web search

Choose priorities without guessing at volume

Use relevance to the business and evidence of an unresolved customer decision. Leave demand unknown when no suitable measurement supports it.

A prompt you wrote is a candidate to test. It isn't a count of what people ask assistants.

If you buy licensed demand data, record the provider, the period and what it covers. Without that, write unavailable.

Give each prompt a simple priority, such as high or normal, with the reason beside it. Look at the priorities again whenever you start a new version.

Use search data as a hint, with its method

Search data can suggest wording, but it measures Google searches, not questions put to assistants. Record which tool you used and what it covers.

In Search Console, the Performance report can group data by the queries people typed. Those are searches where your site showed in Google results, in your customers' own words.

Google Trends shows relative interest, not counts. Its help page says the data is an anonymised, normalised sample, and that it is not a scientific poll.

The same page says Trends leaves out internal searches made by Google products, including AI Overviews. So the searches those features run for a reply aren't counted there.

Your own site search, if you have one, is another source of wording. It records what visitors looked for after they arrived.

Use these sources to choose wording and spot seasonal interest. Note the tool, the period and the place beside any prompt they inspired.

• Search Console Help, Performance report for search results
• Google Trends Help, FAQ about Google Trends data

Retire prompts that no longer earn their place

Remove questions that describe an offer you no longer provide or repeat an existing decision. Preserve their earlier records so past reports remain understandable.

Retire a prompt by leaving it out of the next version, not by deleting it. Its old answers still belong to the period they came from.

Look as well for prompts whose attempts keep failing. A prompt that rarely returns a usable answer gives you little to compare.

Keep the retired prompts in a separate list with the date and the reason. If a service comes back, you can restore the same wording.

Sources and method

A Carqo review method. Provider statements link to the relevant documentation.

• Google guidance on AI features Link checked 4 November 2026
• Google Search Help, Understand and manage your location when you search Link checked 29 October 2026
• OpenAI API documentation, Web search Link checked 4 November 2026
• Search Console Help, Performance report for search results Link checked 29 October 2026
• Google Trends Help, FAQ about Google Trends data Link checked 29 October 2026

Suggest a correction