There is no shortage of prompts in any category. If you sell running shoes, a model can generate ten thousand plausible questions about running shoes before lunch. So “track everything” is not a strategy, it is a way to spend a budget. The useful question is narrower: how many distinct, decision-shaped questions do you need to watch to get a trustworthy, actionable read on your brand?
Across the AI-visibility projects we run, the answer lands in a consistent place. We package it as a simple ladder.
The definitive answer
- 10 prompts - a first look. Enough to sanity-check whether AI assistants know you exist and where you stand on your most obvious commercial questions. Do this before you buy anything.
- 100 prompts - ongoing tracking. The default for most brands monitoring one product line in one market. Big enough to be representative, small enough to read every week.
- 1,000 prompts - a full setup. When the same prompt set also feeds your content roadmap and you are covering multiple products, personas, and locations. In practice most “full” sets settle around 300–600 after pruning.
How many prompts to track, by stage
Recommended working set size (log scale)
Source: Aiso benchmark across client AI-visibility projects, 2025–2026.
If you only remember one number, make it 100. It is the right size for the vast majority of brands tracking a single category, and it is where the cost-to-insight ratio is best.
Why not just track more?
More prompts is not more truth. Three things degrade as your set grows past what you can actually use:
- Overlap. “cheap running shoes for beginners” and “affordable beginner running shoes” are the same question wearing two outfits. Past a point, new prompts are paraphrases, not new signal.
- Dilution. Adding hundreds of top-of-funnel informational prompts that never name a brand drags your headline share-of-answer number toward zero and hides movement on the prompts that convert.
- Cost and cadence. Every prompt is run repeatedly across multiple assistants (see our companion piece on how often to run your prompts). Tracked count multiplies run cost, so a bloated set quietly forces you to run each prompt less often, which makes every number noisier.
What a well-targeted working set looks like
Abstract advice is easy to nod at and hard to copy. So here is the shape of a typical full set, drawn from how we build them across projects. A good foundation for both content and answer tracking usually lands in the few-hundred-prompt range, blended from several real demand sources: questions people actually asked ChatGPT, leads from a brand’s own web-chat assistant, Bing Webmaster query data, and long-tail informational queries from Google Search Console (an increasingly good proxy for what triggers AI answers).
4+
real demand sources blended into one set
~85%
unbranded, where share of answer is contested
~90%
mid & bottom of funnel by design
The shape of the set matters more than its size. We deliberately inverted the natural mix of AI conversations, which skews top-of-funnel and informational, and weighted toward the prompts where a brand can actually win the answer.
How a well-targeted working set breaks down
Typical weighting, biased toward the prompts where a brand can win the answer
Source: Aiso benchmark across client AI-visibility projects, 2025–2026.
By funnel stage
By intent
The logic is simple and worth stealing. A “what should I look for in running shoes” answer rarely names a brand; a “best running shoes for a marathon, Brand A vs Brand B” answer almost always does. Bottom-of-funnel, commercial-intent prompts are both more likely to mention a brand and closer to a purchase, so they earn the bulk of the tracking effort. Keep a layer of informational prompts for coverage, not because you expect a mention there.
The right number falls out of a filter, not a target. Prune every prompt for things you do not sell, and every prompt that never names any brand. Whatever survives - can you win the answer, and do you want the click - is your set.
How to pick your own number
- Start at 10. List your ten most obviously commercial questions (comparisons, “best X for Y”, pricing, location). Run them. This tells you whether you have a problem worth measuring at all.
- Grow to ~100 by adding angles, not synonyms. Each new prompt should cover a new product, persona, location, or competitor - not reword an existing one.
- Only go to the hundreds or 1,000 when the set does double duty. A large set is justified when it is also your content map: every tracked prompt you lose is a content brief waiting to be written.
- Prune ruthlessly. Drop prompts for things you do not sell and prompts that never name any brand. A focused 300 beats a sprawling 1,500.
Get the set right and everything downstream - share of answer, competitive gaps, content priorities - gets easier. Get it wrong and you will either miss the questions that matter or drown them. The companion question, once you have your set, is how often to run it; we cover that in how often should you run your prompts.