How to write prompts worth tracking

Start with the words your buyers use

It's tempting to paste in your keyword list, but keywords are written for search boxes. People tend to give an assistant more to work with, including who they are, what they need and what they've already ruled out.

Compare these two:

• Keyword: standing desk small space
• Prompt: What's the best standing desk for a small flat if I work from home full time?

The second one sounds like a person, and it's also more useful to you, because it tells you which situation you're winning or losing in. To find the real wording, read your sales call notes, support emails, reviews and the forums your customers use. Write down the phrases they repeat, including the ones you wouldn't use yourself.

Cover the whole decision

Buyers ask different things at different stages, and a brand can do well at one stage and be absent from the next. Aim for a mix:

• Learning: "What is a standing desk converter?" or "How do I choose a standing desk?"
• Comparing: "Standing desk or converter for a small flat?" or "[Your brand] vs [a rival]"
• Buying: "Best standing desk under my budget for a small flat" or "Where can I buy a standing desk near Leeds?"

If you could only keep one group, keep the comparing and buying prompts, as these are closest to a sale. The learning prompts still earn their place, though, because they show who assistants treat as the authority on your subject.

Keep most prompts free of your brand name

A prompt that mentions you will usually get a reply about you, so it says little about whether you'd have been considered in the first place. Prompts without your name are the real test, and they often start low. In Pratyush Kumar's study of more than 100,000 prompt responses, niche and small brands appeared in just 11% of relevant replies on their first runs, against 73% for global household names.

Branded prompts still have a job to do. Keep a small group of them to check what assistants say about you: what you sell, who you serve and whether any of it is out of date. Read them across several runs, because the same study found tone flips about 6.7 times more often than whether a brand is mentioned at all.

Be specific, and don't collect rewordings

Prompts with a concrete need, like "what size do I need" or "which is better for a small flat", tell you more than broad ones like "ideas for a home office". They're also closer to the moment someone decides. If you sell in particular places, add the location to some prompts and compare the results by place.

You don't need a dozen rewordings of the same prompt. Assistants already split one prompt into several related sub-queries when they search, so close variations tend to add cost rather than insight. Spend your prompts on different situations instead: another type of customer, another use, another objection.

Expect some prompts to return no citations

Not every prompt sends an assistant to the web. As of February 2026, ChatGPT activated web search on 34.5% of queries, down from 46% in late 2024. On the rest it replies from what the model already knows.

That's worth noting prompt by prompt, because it changes what you can do about the result. Where the assistant cites sources, the pages it reads can make a difference. Where it doesn't, editing your own page won't help much, and how widely your brand is known matters more. Prompts with no citations are still worth keeping if buyers use them. Just don't expect a quick page fix to move them.

Hold the set steady and run it repeatedly

Two habits make the numbers trustworthy. The first is repetition. AirOps found that only 20% of brands remain present across five consecutive runs of the same prompt, so run each prompt at least five times and report a rate rather than a single result.

The second is patience. Keep the same set for about a quarter, because you can't compare this month with last month if the prompts have changed. Then review it. Drop prompts nobody would really use and add any new wording you've picked up from customers. If you must add something urgent in between, track it separately so that your trend stays clean.

A quick test for any prompt

Before a prompt goes in your set, it should pass all five of these:

1. Would a real buyer type this, in roughly these words?
2. Does it match a stage of the decision you care about?
3. Would a good or bad result change what you do next?
4. Is it a different situation from the prompts you already track?
5. Could you sell to the person asking?

If you're new to this, read what AI visibility tracking is first. Once your prompts are running, how to earn more citations from AI assistants covers what to do with the results. In Suggestible, the plan you choose sets how many prompts you can track, so a tight, well-chosen set goes further than a long list.