AI Prompts
The panel of questions RankX AI asks the assistants. Writing good ones, the three statuses, the per-plan cap, and why the prompt list is your credit bill.
A prompt is a question RankX AI asks the AI assistants on your behalf, on a cadence you set for that prompt, so it can record whether your brand gets named in the answer. The AI Prompts screen is where that panel is managed, and it is the single most consequential screen in the product: your prompt list decides both what you measure and what you spend.
Every prompt has its own schedule and its own assistants
Nothing runs every night across every assistant unless you set it to. Each prompt carries two settings of its own, and both start out following the Website:
| Setting | What you can choose | Default |
|---|---|---|
| Schedule | Daily, every three days, weekly, fortnightly or monthly | Use the Website default |
| Assistants | Any of the assistants the Website has switched on | Use the Website's list |
So one prompt can run on ChatGPT every night while another runs on Perplexity once a week, and you pay for each accordingly. The next scheduled check is shown beside every prompt.
A prompt can only ever narrow the Website's list, never widen it. If Claude is switched off for the whole Website, a prompt that targets Claude does not check it, and RankX AI says so rather than silently returning four assistants where you asked for five.
Choosing slower than your plan's floor is always allowed. A floor is a lower bound on the interval, which is an upper bound on how often you spend. See check cadence for the two floors and which wins.
Your prompt list is your credit bill
Every active prompt is checked on the assistants it targets, on every run it is due for, and each of those pairs is charged. Twenty-five prompts on all six assistants is 150 charged checks per run, not twenty-five, and 125 on five. A prompt targeting two of them is 2 checks per run rather than six.
So the arithmetic that matters is:
prompts x assistants x runs per month
That is most of the recurring spend on almost every account. Pruning the panel is the single most effective thing you can do about your balance, and it costs you nothing but the prompts you were not going to act on anyway.
The three statuses
| Status | What it means |
|---|---|
| Active | Runs on its own schedule, and is billable when checked |
| Inactive | Paused. Does not run, costs nothing, keeps its history |
| Suggested | A proposal RankX AI generated that you have not accepted |
Pausing is the reversible way to stop a prompt costing credits. It keeps every measurement already taken, and resuming picks up where it left off. Prompts are paused rather than deleted, by design, so a prompt's history is never lost.
Buyer stage
Every prompt carries a stage, which is what the buyer is doing when they would ask it:
| Stage | The buyer is | Example shape |
|---|---|---|
| Researching | Learning about the problem | "Why do my pages not appear in AI answers?" |
| Comparing | Weighing up options | "Best AI visibility tracking tools" |
| Ready to buy | Choosing who to go with | "Who should I use for AI search tracking in Manchester?" |
A panel that is all one stage tells you about one moment in a purchase. The useful shape is a spread, weighted towards the stage where your brand should be winning.
Writing a prompt that measures something
RankX AI generates a starting set during onboarding from your brand profile, topic clusters and competitor set. The generated ones are a good baseline; the ones you write yourself are usually better, because you know how your buyers talk.
Ask what a buyer would ask. "Best CRM for small law firms" is a prompt. "CRM software solutions" is a keyword, and a keyword asked of an assistant produces a vague answer that mentions nobody in particular.
Do not put your brand name in it. A prompt that names you is a question about you, and an assistant will name you in the answer. That measures nothing. The whole point is whether you come up unprompted.
Be specific enough to have an answer. Geography, segment, use case, price band, integration. A prompt with no constraints returns the four best-known names in the category every time, on every assistant, and never moves.
One question per prompt. A compound question gets a compound answer and the verdict becomes ambiguous.
Write the ones you would want to win, not the ones you already win. A panel tuned to make the numbers look good is a panel that tells you nothing.
The per-plan cap
Plans cap AI visibility prompts per Website, and the cap is enforced by the database rather than by the interface, so a bulk add that exceeds it is refused rather than silently trimmed. The per-plan figures are in plans and limits.
AI Shopping prompts are a separate allowance with its own cap, and the two do not draw on each other. See AI Shopping.
Reading a prompt's performance
Each prompt shows its mention rate per assistant, over the window you choose.
A blank rate means no check in the window produced an analysed verdict, not zero. Reporting it as 0% tells you that you are invisible when the truth is that nothing was measured. See null is not zero.
A rate of zero on a prompt you should win is the most useful row on the screen. Open it in the AI Chat Feed and read what the assistants said instead: the brands they named and the sources they leaned on are the whole brief for what to do next.
Do not read day-to-day movement on a single prompt. It is noise. A prompt is a sample of one question, and assistants answer the same question differently on consecutive runs.
Running a check manually
You can run one prompt, or all of them, at any time, as often as you like. A manual run ignores the schedule entirely: a human asking for a check has said what they want. It still honours the prompt's assistant list, so it costs exactly what a scheduled run of that prompt costs, and it confirms the price before it starts.
It is worth doing after a real change to your site or your presence somewhere else, and not worth doing to see whether the number moved overnight.
Managing the panel over time
A prompt panel is not set once. A reasonable rhythm:
- After the first scheduled re-check, prune. Pause anything that is not a question your buyers ask, and anything you would not act on if you lost it.
- Monthly, read the zero-rate prompts and decide which are real gaps and which are prompts nobody asks.
- When you launch something, add prompts for it before you start measuring whether it worked.
- When a cluster stops mattering, pause its prompts rather than deleting the cluster.
Driving this from an assistant
Every part of this screen has an MCP tool:
list_prompts to read the panel, create_prompt to add, set_prompt_status to
pause and resume, get_prompt_visibility to read the rates, and
run_prompt_check to run one now.
"Show me my tracked prompts with their mention rates, recommend five to pause, and tell me what that saves per run. Do not change anything yet."
Where to go next
- The AI Chat Feed, to see what the answers actually said.
- Topic clusters, which is what a prompt attaches to.
- Credits and metering, for the spend model in full.
Last updated
AI Visibility
How RankX AI measures whether the six tracked AI assistants name your brand, and what each screen in the AI Visibility section is for.
AI Chat Feed
The raw answers the assistants gave to your tracked prompts. The most useful screen in RankX AI, and the first one to read when a number surprises you.