How to do AI prompt tracking: size, split and run a prompt set in 6 steps
How to do AI prompt tracking: size the set with budget math, pick cadence by volatility, read rates over many runs, and split a frozen core from exploratory prompts.
Most prompt sets are built backwards. A team brainstorms 200 questions, loads them into a tracker, and finds out at the end of the month that the plan covered two weeks of runs, or that every prompt got so few answers the trend line is mostly noise. The prompts were fine. The arithmetic came last.
This guide puts it first. You'll size the set against a budget, split it into a part you never touch and a part you experiment with, choose cadence per prompt, and read the results as rates rather than single answers. Where a step maps to a feature, we name the Promptwatch feature, because its published prompt and response limits make the math checkable.
Step 1: write the budget equation before you write prompts
Every tracker spends the same unit, one prompt asked on one engine one time. Call it a prompt-run. The monthly total is:
prompts × engines × markets × runs per month = prompt-runs
Plug in your plan before you plug in your ideas. Promptwatch's brand plans come out like this at daily cadence:
| Plan | Prompts | Responses a month | Fits at daily cadence |
|---|---|---|---|
| Essential, $95 | 50 | 6,000 | 50 prompts on 4 engines |
| Professional, $245 | 150 | 18,000 | 150 prompts on 4 engines, or 100 on 6 |
| Business, $579 | 350 | 42,000 | 350 prompts on 4 engines |
Agency plans (Kick-off, Growth, Scale) have unlimited prompts, so the response pool is the only constraint: 10,000, 25,000 or 65,000 a month. Kick-off's 10,000 covers about 83 prompts on four engines daily.
Markets multiply hardest. In Promptwatch a monitor's language and location are fixed once it's created, so a second country means a second monitor with its own runs. If you add personas, budget for them as extra answers too, and confirm the real cost in the Usage Overview after the first week. The order of trimming when the equation doesn't fit: engines your buyers don't use first, then duplicate prompts, then markets that don't drive revenue.
Step 2: split the set into a frozen core and an exploratory edge
A prompt set has two jobs that pull against each other. Trend lines need the same strings, run the same way, for months. Learning needs new prompts all the time. Put both jobs in one list and every edit breaks a chart.
So split it. The core is the set you report on: buyer questions across your main topics, worded the way customers word them, and left alone. Give it most of the budget, say two thirds, so each prompt gets enough runs to be stable. The exploratory edge gets the rest. That's where candidates go to prove themselves for a few weeks. Each month, promote the ones that turned out to matter and drop the ones that didn't. Don't reword a core prompt. Retire it and add the new wording as a new prompt, so the old trend stays readable.
Promptwatch has the plumbing for this. Topics, with their own colors and a Manage tab, group prompts for reporting, and tags plus bulk editing make "core" and "explore" a filter rather than a spreadsheet. Prompt types (Organic, Brand Specific, Competitor Comparison) and intent (Branded, Informational, Navigational, Commercial, Transactional) let you check the core isn't all branded questions. Copy Prompts moves a promoted prompt into the core monitor, and can copy to a monitor in another country if the language matches. New candidates come from Prompt Explorer (suggestions from Search Console, ranking keywords, People Also Ask, manual search or any page on your site) and Prompt Volumes, which adds monthly volume, difficulty and citation rates across AI search, Google and Bing. In Prompt Explorer the language picker sets what the prompts are written in and the country sets the market, which matters if you sell in more than one.
Step 3: choose cadence by volatility, not by importance
The instinct is to run your most important prompts most often. Importance isn't the right test, though. Volatility is. A prompt whose answers barely change between runs gains almost nothing from extra runs. A prompt whose brand list reshuffles every time needs more samples before any single week means anything.
To find out which is which, run the core daily for two weeks and look at each prompt per engine. Count how often your brand appears and how often the set of named brands changes from one run to the next. Stable prompts stay daily. Volatile ones are candidates for a faster cadence on a small, separate monitor, or for a longer reading window if the budget won't stretch.
Do the cost check first. One prompt on one engine is about 30 runs a month daily, 60 every 12 hours, 120 every 6 hours and about 720 hourly. Ten volatile prompts on four engines moved from daily to every 6 hours go from 1,200 to 4,800 runs a month. In Promptwatch, every 12, 6 or 1 hours is available on Business, Agency Scale and Enterprise. Existing monitors stay daily until you change them, you can't pick a frequency your remaining quota can't cover, and lowering a frequency always works. Our hourly vs daily guide covers when the faster options are worth it.
Step 4: read rates over many runs, not single answers
An AI answer is one draw from a spread of possible answers. Reporting it as "we're third in ChatGPT" treats the draw as the truth. The number to report is a rate: of the responses for this prompt on this engine over four weeks, what share named you, and where?
Sample size decides how far you can trust the rate. With 30 runs, a single answer moves your mention rate by about 3.3 points. With 120 runs, it moves it by under one point. That's why the answers-per-prompt figure from step 1 matters so much, and why a tight set run often beats a sprawling one run rarely. Our post on whether a single run is enough goes further into the sampling side.
Promptwatch's metrics already work as rates. The Visibility Score (0 to 100) counts responses without your brand as zero, so absences pull it down. Share of Voice gives your slice of all brand mentions, one per response per brand. Position is averaged only across responses that mention you, so read it next to the other two and never alone. If you want to compute your own rates, the Responses export lists every brand mention with its position, in the form "Acme (1), Globex (2)", one row per response.
Step 5: check what the engine searched, then cut redundant prompts
Query fan-outs are the sub-queries an engine runs before it answers. Promptwatch shows them per model, the exact searches behind each response. They help in two ways.
First, diagnosis. If a prompt never names you, the fan-outs show what the engine went looking for, which is usually a page type or angle you don't cover. Second, budget. If two prompts fan out into the same sub-queries, they're measuring the same thing. Keep one and give its runs to a prompt that covers new ground.
Fan-outs shift, so re-check them. Promptwatch's data shows ChatGPT's average fanouts per response rising from about 1.08 to about 1.83 on August 8, 2026 (report). After a change like that, a redundancy check you ran in July is out of date.
Step 6: watch the quota, and review the set on a schedule
Quota problems are quiet. You notice them as a gap in a chart three weeks later. Promptwatch sends usage alerts at 85% of a quota and has a per-project Usage Overview, so you can see which monitor is eating the pool. If a plan changes, owners pick which prompts keep running on a downgrade, which is another reason to keep the core clearly tagged: you'll know what to protect.
Then set a rhythm. Monthly, promote or drop exploratory prompts and check the Usage Overview against the equation from step 1. Quarterly, look at the core as a whole. Add the questions sales and support keep hearing, retire prompts that no longer map to anything you sell, and keep a dated log of every change so a step in the trend line can be traced to an edit rather than to the engines.
Common mistakes
Loading every brainstormed prompt on day one is the big one; the budget spreads so thin no prompt gets a stable rate. Close behind: mixing markets in one monitor, rewording core prompts in place, and moving everything to hourly because it sounds more thorough. More runs give you a sharper picture of the answers. They don't make you more visible.
Where to go from here
This post is about the prompt set. For where your brand lands inside the answers, see how to do AI rank tracking. To compare tools on the limits this method depends on, read AI prompt tracker compared, and for a ranking by what each tracked prompt buys, the best AI prompt tracker 2026. The general list of trackers is at what are the best AI rank trackers.
Every step above maps to a feature in Promptwatch: published prompt and response caps for step 1, topics and Copy Prompts for step 2, per-monitor frequency for step 3, rate-based metrics and the Responses export for step 4, fan-outs for step 5, and usage alerts for step 6. Write your equation, pick the plan it fits, and start the 7-day free trial with your core set loaded on day one. A week of daily runs gives you a first read on which prompts are volatile, and the Usage Overview shows whether your equation held.