Choose the prompts AI Engine Visibility runs

A keyword on its own is not how anyone talks to an AI assistant. Nobody types "emergency plumber Leeds" into an assistant; they ask "who can fix a burst pipe in Leeds tonight?", or they simply say "find me a plumber in Leeds who can come out tonight". So AI Engine Visibility checks each keyword you track as a real prompt, and the prompts you choose are the single biggest lever on how useful your results are. This guide shows you how to set them.

Where you shape your prompts

Open Tracking > AI Engine Visibility, open a location, and choose the Coverage tab on the right of the tab bar (a location that has never been checked opens there). The Keywords and prompts panel lists each tracked keyword with a count of its prompts; click a keyword to see the prompts that will be put to the engines. Prompts are shared across engines, so a prompt you write once is asked of ChatGPT and Gemini alike.

You can also work need by need. Choose Edit coverage on the Coverage tab, or open Research > Prompt Research, and click a need on the research map: its panel lists every prompt asked about it, to reword, remove or add to. Both pages show the same prompts, so a change on one shows on the other.

Every keyword starts with a couple of prompts drafted for you, so you get useful results without doing anything. From there you can accept them, sharpen them, or add your own.

Some keywords get no prompts, and that is deliberate. A keyword marked Not suited is one no prompt could turn into a request that names a local business, so it is not counted in your coverage and nothing is charged for it. A keyword marked Shares prompts with another keyword asks for the same thing, so it is measured by that keyword's prompts instead of paying to ask twice. Use the Needs prompts filter to find keywords that are still short.

A prompt can be phrased either way, and both are worth tracking. Some customers ask a question ("who's the best plumber in Leeds?") and others make a direct request ("find me a plumber in Leeds who can come out today"). The agent drafts a mix, because ChatGPT does not always respond to the two the same way.

Shape each prompt

Each prompt has three small actions next to it:

• Edit (the pencil) rewrites it in your own words. Anything you change is marked "yours"; drafts not yet reworded are marked "ready-made".
• Suggest a different prompt swaps in a fresh phrasing if a draft does not sound like your customers.
• Remove takes it out of the run. A keyword always keeps at least one prompt, so you can never empty it by accident.

If a prompt has already run, editing it or swapping it saves the new wording as a new prompt and retires the old one with its history, so your trend keeps comparing like with like. Removing a prompt that has run retires it the same way. Retired prompts sit under retired prompts on the keyword, and Restore brings one back with its history.

To add another angle to a keyword, use Write your own to type a prompt by hand, or Suggest another to have one drafted for you that avoids repeating what is already there.

You do not have to draft them yourself

When you open the list, any keyword with no prompts is filled, and the agent then rewords those drafts for your business in the background. You will see "Sharpening the wording for your business" while it runs. Keywords you have already shaped by hand are left alone, and so are keywords already being measured, because the trend depends on the wording staying put. If a keyword still shows ready-made prompts afterwards, an Improve wording link under them asks the agent to try again.

So the honest instruction is: open the list, let it settle, then read what it came up with and change the ones that do not sound like your customers.

How many you can add

How many prompts each keyword gets depends on the coverage depth set for the location under Edit coverage: 2 on Ready to buy, 3 on Comparing and 4 on Exploring, plus one for each situation you test. Your plan caps a keyword's everyday prompts at 10 on Starter, 20 on Pro, 30 on Agency and 50 on Enterprise (situation prompts do not count towards this). Each keyword row shows how many prompts it has, and when a keyword reaches the cap the add buttons are replaced by a note saying so.

Your plan also sets how many keywords at each location are checked: 10 on Starter, 50 on Pro, 200 on Agency and 750 on Enterprise unless your contract says otherwise. Past that number, the most recently added keywords are not checked.

That ceiling is deliberately generous, because the real limit is your allowance, not the cap. Every prompt put to one engine once costs one AI visibility check, so running the same prompt on both ChatGPT and Gemini costs two. A scheduled run costs the same. If Brand evaluation is on, its prompts about you by name run once a month and are shown separately. The run bar shows the total before anything runs. So the goal is not to fill the cap; it is to pick the handful of prompts that genuinely tell you something. See How credits and quotas work.

Tips for prompts that earn their check

• Write the way a customer would say it out loud. Full, natural sentences beat keyword fragments. "Can you recommend a good electrician near Didsbury?" is how people really ask.
• Name the service and the place. One clear intent and one location per prompt gives ChatGPT something specific to work with and keeps the result easy to read.
• Cover the different ways people ask. Best, cheapest, near me, open now, same-day, "who do you recommend", and the blunt "find me one". Being named however the prompt is phrased is what the Named figure in the keyword table on the Summary tab shows.
• Start from your money terms. Anchor prompts to the keywords that actually win you work, not vanity phrases you would like to rank for.
• Spend your extra prompts where they matter. Put two or three angles on your most important services, and leave the long tail on a single prompt each.
• Edit, do not settle. If a drafted prompt is close but not quite you, edit it to your real wording rather than tracking something a customer would never say.

Ask as the customer who is actually asking

A prompt is what someone asks. A persona is the constraint they are carrying while they ask it: on a tight budget, needs an evening appointment, does not drive. The same prompt with a constraint attached often returns a different set of businesses, and that gap is usually more useful than the prompt's own score.

In the app these are called situations, and they are set on the location's coverage page: on the Coverage tab choose Edit coverage, then press Add a situation in the plan at the top of the page. See Ask as the customers you actually get.

Then run and read

Once your prompts look right, press Run check now. Every keyword starts ticked, so untick any you want to leave out, and the button shows the checks the run will use. Regular checks run on the schedule set up with the location's coverage, which you can change under Workspace > Schedules. Your results come back as a visibility rate, a share of voice against competitors, and the full response saved word for word. For what those measures mean, see Track AI Engine Visibility, and keep your underlying terms tidy with Build your keyword library. To see which questions your own data says people actually ask, and which of them you are not yet covering, use Research your prompts.