- A prompt set built from SEO keyword permutations tests phrasing, not buyer intent. The two are not the same exercise, even though they produce lists that look superficially similar.
- IAB's own published framework formalizes four intent categories for prompt-type coverage: informational, comparison, recommendation, and transactional, the same structure independently converged on across multiple vendor guides for building an AI-visibility prompt set.
- IAB's quality tiers tie directly to this: directional measurement requires at least two of the four intent types with the distribution disclosed; decision-grade requires all four, segmented and disclosed.
- Converging practitioner guidance lands on a 20-50 prompt range as a reasonable working set size, with IAB's own floor treating anything under 50 total queries across a whole program as merely exploratory.
- Filter a draft list against four criteria before locking it: competitive relevance, influenceability, business intent alignment, and scope. A shorter, filtered list beats a longer, unfiltered one.
AI visibility KPIs establishes that citation frequency, sampled across a fixed prompt set, is the foundation every other KPI sits on. It doesn't cover how to actually build that prompt set. This piece does, starting from buyer intent rather than keyword variation, structured around the same four intent categories IAB's own measurement framework formalizes.
Why doesn't a keyword-variation list work here?
Classic SEO keyword research produces lists like "AI visibility tool," "best AI visibility tool," "AI visibility tool pricing," the same underlying query, permuted by modifier. That approach tests phrasing sensitivity, which matters for a different reason (a generative engine's answer can genuinely shift based on how a question is worded), but it doesn't map onto how a real buyer actually interacts with a chat interface. Nobody types "AI visibility tool pricing" into ChatGPT the way they might type it into a Google search box. They ask a fuller question, in their own words, shaped by where they actually are in a decision.
A prompt set that's supposed to predict something real about buyer-facing visibility needs to start from that decision journey, not from a spreadsheet of modifier combinations.
What are the four intent categories to build around?
IAB's Measuring Visibility in the AI Era framework requires prompt-type coverage across four intent categories for its quality tiers: directional measurement needs at least two, with the distribution disclosed; decision-grade needs all four, segmented and disclosed:
- Informational. The buyer knows they have a problem but hasn't settled on a solution category yet: "how do I know if AI models are citing my brand correctly," not "which AI visibility tool should I buy."
- Comparison. The buyer has narrowed to a solution category and is now narrowing a shortlist: "what's the difference between AI visibility tracking tools," or a comparison naming specific categories of solution rather than specific vendors.
- Recommendation / transactional. The buyer is asking which specific option to pick. The prompt most directly analogous to a branded or near-branded search in classic SEO, and the one most vendors' demo screenshots come from.
- Post-purchase / support. A category some practitioner guides add on top of IAB's three. A buyer who's already chosen is now asking how to use, troubleshoot, or get more out of what they picked. Less central to a pre-purchase visibility program, but relevant if retention or expansion citations matter to your category.
Independent practitioner guides, including one published directly by Profound, a vendor in this exact measurement category, converge on the same four-part structure without citing IAB directly, which is worth noting as convergent practice, not proof of anything beyond that: multiple parties building or buying this measurement have independently landed on the same intent taxonomy.
How many prompts should actually be in the set?
There's no single peer-reviewed number here, and any specific figure a vendor quotes should be read as practitioner guidance, not settled research. What's converged on across the vendor guides checked for this piece is a working range of roughly 20 to 50 prompts, distributed across the intent categories above rather than concentrated in one. IAB's own floor (fewer than 50 total queries across an entire measurement program counts as merely exploratory) sets a related but distinct threshold: that's total program volume (prompts × platforms, not just prompt count), and it's the level below which even a well-built prompt set isn't measuring enough to call directional.
The practical takeaway from both: more prompts isn't automatically better. A prompt set padded with marginally-relevant queries dilutes the signal from the ones that actually matter, and running fewer, well-chosen prompts more times each (see how many prompts do you actually need for the runs-per-prompt math) is a better use of a fixed measurement budget than maximizing prompt count.
How do you actually filter a draft list down?
Draft broadly first. Pull real questions from sales call transcripts, support tickets, and the actual language prospects use, not a keyword tool. Then filter every candidate prompt against four criteria before it makes the final set:
| Filter | The question to ask |
|---|---|
| Competitive relevance | Would a real, named competitor plausibly appear in the answer to this prompt? If a query is so narrow or so broad that competitive dynamics don't apply, it's not testing what you need it to test. |
| Influenceability | Is this a query your own content, entity work, or earned-media presence could realistically move, or is the answer dominated by factors entirely outside your control (a platform's own product bias, a structural category fact)? |
| Business intent alignment | Does this map to an actual stage in your real buyer journey, or is it a plausible-sounding question nobody in your category actually asks? |
| Scope | Is the prompt specific enough to run identically, repeatedly, and get a comparable answer each time, or so broad that "did we appear" becomes a subjective judgment call from one run to the next? |
A shorter list that survives all four filters outperforms a longer list that doesn't. The same instinct behind trimming a keyword list for search intent, applied to a fundamentally different kind of query.
What happens after the set is built?
The set itself is only half the methodology. How many prompts do you actually need covers the other half: how many times each prompt needs to run, over what window, before the resulting frequency means anything. A well-built, buyer-language prompt set run once tells you almost nothing, per the same research; the set and the sampling discipline are both required, not substitutes for each other.
Frequently asked questions
What's wrong with building a prompt list from SEO keyword variations?
A keyword variation ("AI visibility tool", "best AI visibility tool", "AI visibility tool pricing") tests phrasing, not intent. A prompt set built to predict anything real needs to start from what an actual buyer types into a chat interface at a specific stage of their decision, in their own words. That's a different exercise than permuting a keyword.
What are the four intent categories a prompt set should cover?
The same four IAB's Measuring Visibility in the AI Era framework uses for prompt-type coverage: informational (the buyer knows they have a problem but hasn't picked a solution category), comparison (narrowing a shortlist), recommendation/transactional (asking which option to actually pick or buy), and a fourth, post-purchase/support-style intent some practitioner guides add on top. IAB's own quality tiers require at least two of these for directional measurement, all four, disclosed and segmented, for decision-grade.
How many prompts should be in the set?
There's no single validated number, but converging practitioner guidance (including vendor guides published by companies building this measurement) lands in the 20-50 range, and IAB's own floor treats fewer than 50 total queries across an entire measurement program as merely exploratory. More prompts isn't automatically better if they're not filtered for relevance; a shorter, well-filtered list outperforms a long, unfocused one.
How do I filter a prompt list down once I've drafted it?
Filter against four criteria before locking the set: competitive relevance (would a real competitor plausibly show up in the answer), influenceability (is this a query your own content and entity work could actually move, or is it dominated by structural factors outside your control), business intent alignment (does it map to a stage in your actual buyer journey), and scope (is it specific enough to be testable and repeatable, or so broad it can't be tracked consistently).
Sources referenced
- IAB: Measuring Visibility in the AI Era (Aug 2026)
- Profound: How to Design Prompts for AI Visibility Tracking in 7 Practical Steps
This piece is the prompt-set construction methodology behind the citation-frequency KPI in AI visibility KPIs. For the sampling discipline to apply once the set is built, see how many prompts do you actually need; for the literal build once both are decided, see how to build your own AI visibility tracking without a vendor tool. For a smaller, productized version of this same filtering discipline, see what is an AI Shortlist Scan?
