Most advice on choosing prompts to track ends at the list: mix discovery, comparison, and decision questions, keep branded prompts apart, and start with a few dozen. The list then gets treated as if every prompt counts the same. They don't.
In a Zumi prompt run on Indian stock trading, 702 prompts were built from Google keywords carrying 1,004,700 searches a month. The 205 prompts built from the smallest keywords, 29% of the list, carried 1.28% of those searches. The top 50 prompts carried 67%.
A mention rate averaged across that list gives a prompt built from a keyword searched a few dozen times a month the same weight as one built from a keyword searched tens of thousands of times. This guide covers how to choose prompts so the list shows where the demand sits, and how to report results so that weight is visible.
Key takeaways
- Every tracked prompt should record the Google keyword it came from and that keyword's monthly searches, so the list carries its own demand record.
- Prompt count and demand are different things: in the 702-prompt run, 29% of prompts carried 1.28% of the searches and the top 50 carried 67%.
- Each rate should be reported twice, once with every prompt counted equally and once weighted by searches, because a brand mentioned often in low-volume questions and rarely in high-volume ones looks healthy on the first view and weak on the second.
- Low-volume prompts stay in the list. In the same run, whole topics such as bonds and debt existed only below 100 searches a month, and a volume floor would have removed them.
- Prompts outside what the brand sells are removed before anything is counted, and categories the run never measured are reported as unknown, not as zero.
Why should each prompt carry its search demand?
AI engines don't publish how often a question is asked. Google does, through Keyword Planner, which reports average monthly searches for a keyword. A prompt built from a keyword can borrow that number as its weight.
So each row in the prompt list holds two things beside the prompt itself: the keyword it came from and that keyword's monthly searches. The intent label, the branded flag, and a stable prompt ID still apply. The baseline guide covers those fields, and they stay as they are.
The searches figure is a proxy. It measures how often people type the underlying need into Google, not how often anyone asks ChatGPT or Gemini the exact prompt. That makes it a sound basis for weighting one prompt against another and the wrong basis for any claim about market share.
How does a keyword become a prompt?
A keyword is written for a search box, without articles or question framing. A prompt is what a person types or says to an AI engine. Copying the keyword into the prompt column produces a search term, not a prompt.
Each prompt moves at least one step away from its keyword, through a question word, a comparison, natural word order, a buyer situation, or an added constraint. From the stock-trading run, where a demat account is the account Indian investors hold shares in:
| Keyword | Monthly searches | Prompt tracked |
|---|---|---|
| demat account | 60,500 | which demat account should a first-time investor in india open |
| demat system | 60,500 | what is the best demat system in india for someone just starting out |
Both prompts keep the keyword's searches as their weight. Several prompts can share one keyword, which is how a run covers the different ways buyers ask about the same need.
Which prompts should be left out?
A prompt counts only if the brand could reasonably appear in the answer. The test is the brand's own website: its category navigation shows what it sells and who it sells to. Three kinds of prompt fail that test:
- Categories the brand does not sell. A skincare brand with no fragrance line gains nothing from fragrance prompts.
- Audiences the brand does not serve. A brand with no men's range is not missing from men's grooming answers. It was never in that market.
- Filters the brand cannot meet. A prompt asking for the best Korean skincare brand excludes every brand that isn't Korean, whatever its quality.
These prompts come out before any rate is calculated, and the searches they carried are stated in the method note. A brand left in those prompts would show as absent from answers it was never eligible for, and that absence would drag every rate down for no reason.
The reverse also holds. A category the run never measured is reported as unknown. Reporting it as zero tells the reader the brand was checked and missing, when it was never checked.
Should low-volume prompts be cut?
No. Cutting prompts below a volume floor looks like tidying, and it removes whole parts of the market.
In the stock-trading run, three topics existed only below 100 searches a month: sub-brokers and franchises, bonds and debt, and ETFs and index funds. A floor at 100 would have deleted them without a trace.
These low-volume prompts stay in, with their true weight. That weight is small, which is what the weighted view is for.
Low-volume prompts answer a narrower question than high-volume ones. Keyword Planner rounds small volumes heavily: all 205 prompts under 100 searches a month in the run sat on one of six values, from 10 to 90 searches a month. A keyword at 50 and one at 70 may differ by rounding alone, so low-volume prompts support present or absent findings and are never ranked against each other.
How should the results be reported?
Every rate is reported twice: once per prompt, where each prompt counts equally, and once weighted by the searches behind each prompt. Where the two disagree, the disagreement is the finding.
A worked example uses invented numbers.
A list has 100 prompts built from keywords carrying 10,000 searches a month. The 20 high-volume prompts carry 400 searches each, and the 80 low-volume prompts carry 25 each. A brand is mentioned in answers to 2 high-volume prompts and 48 low-volume ones.
| View | How it is calculated | Result |
|---|---|---|
| Per prompt | 50 prompts mentioned, out of 100 | Half |
| Weighted by searches | (2 × 400) + (48 × 25) searches mentioned, out of 10,000 | One fifth |
The per-prompt view says the brand appears in half the category. The weighted view says it appears where a fifth of the searching happens. Both are true, and only the pair shows that the brand is strong in specialist questions and weak in the high-demand ones.
The method note also states how demand spreads across the list, in figures. In the stock-trading run it reads:
| Monthly searches per keyword | Prompts | Share of prompts | Share of searches |
|---|---|---|---|
| None measured | 45 | 6.4% | 0% |
| 1 to 99 | 205 | 29.2% | 1.28% |
| 100 to 999 | 293 | 41.7% | 9.7% |
| 1,000 or more | 159 | 22.6% | 89% |
A reader who sees "702 prompts" without that table will read every total as 702 equal parts of the market.
How many prompts does a list need?
Published advice ranges from about 10 prompts to 1,000, and no single number fits every category. The demand record answers the question more usefully than a count does: a list is large enough when its prompts cover the categories the brand sells and the method note can say what share of the measured searches they carry.
Prompt count also needs care in the report itself. Rates are written in answers, and the method note discloses the prompts and repeat runs behind them, because ten runs of one prompt are not ten independent observations. The same rule applies to citation figures.
When is this not the answer?
Demand weighting needs Google demand to weight by. A new category, or a narrow B2B market where buyers rarely search Google for the problem, can return no measurable searches for most keywords. Weighting then adds nothing, and the list is better built from sales calls, support questions, and the brand's own site search, with the per-prompt view reported alone and that limit stated.
Keyword Planner also reports by country and language. A brand selling in several markets needs a separate demand record for each one, and a weighted figure that mixes markets says which ones went into it.
Zumi, an AI Search Intelligence Platform, tracks a brand's prompt list across the AI engines in its plan, and plans differ by the number of prompts and engines tracked. How it works covers the workflow. Book a demo to see the prompt list a brand's own category would need.
Sources
- Zumi prompt run, Indian stock trading category, August 10, 2026: 702 prompts built from Google Ads Keyword Planner keywords for India, 1,004,700 average monthly searches in total. A category run with no client data. Unpublished; figures recorded in Zumi's claims register.
- Google Ads Help, "About Keyword Planner forecasts and historical metrics".