What prompt types your library needs to cover
Four prompt types cover the questions a B2B buyer actually puts to an assistant. A set missing one of them is blind to that question, whatever its size.
| Type |
What the buyer asks |
What it tells you |
Example |
| Informational |
How the problem works |
Whether AI treats you as a source on the problem itself |
“how does AI search visibility tracking work” |
| Comparative |
Who solves it best |
Whether you make the shortlist |
“best AI visibility agency for B2B SaaS” |
| Brand evaluation |
Whether you specifically are any good |
Whether the answer describes you accurately |
“is Rampiq worth it” |
| Transactional |
What it costs and how to choose |
Whether you appear at the decision point |
“how much does a GEO audit cost” |
SE Ranking’s guide splits the same ground into five categories and Aleyda Solis names ten. Both agree on the part that matters. Each group answers a different visibility question, so a missing group is a missing answer rather than a smaller number.
Map each type onto a buying stage as you go. The awareness, consideration and decision split stops a set drifting to the bottom of the funnel. Sets drift there by default, because decision-stage prompts feel closest to revenue.
For candidates, start with the questions your sales and support teams already answer every week. Then rephrase your non-brand Search Console queries as questions a person would type into a chat window.
How many prompts to track, and why more isn’t automatically better
Beyond that starting range, the set grows with the number of products and markets you sell into.
SE Ranking’s Yevheniia Khromova recommends starting with 20 to 40 prompts spread across buying stages, with brand-evaluation prompts held in their own group (April 2026). Aleyda Solis puts the opening range higher, at 30 to 50 commercially relevant prompts. For SaaS and services businesses she scales that to 100 to 250 (updated June 2026).
Her summary of the tradeoff is blunt: “A smaller, well-structured prompt library is better than a large, random one.”
The real ceiling is your run budget. Solis’s measurement rule is that a single answer is one sample of a distribution, not a fixed ranking. Each core prompt therefore needs three to five runs before its result carries weight.
Every prompt you add multiplies against that number. At 30 prompts and five runs you’re reading 150 answers a cycle. At 120 prompts you’re reading 600. Or you quietly drop to one run each and start reporting noise as a trend.
Signal quality degrades before the count does. That’s why “we track 200 prompts” is a weaker claim than it sounds. Comparing what different AI visibility tracking tools charge per run makes the budget explicit.
Add prompts when the business adds something, not when the dashboard looks thin:
- A second product line with a different buying committee
- A second market or language
- A competitor set that changes by segment
- A pricing or packaging change that creates new objections
Branded vs. unbranded: why they can’t share one bucket
A branded prompt contains your company name: “Is Rampiq worth it?” or “Rampiq vs another agency”. An unbranded prompt, also called a category prompt, describes the problem without naming anyone: “best AI visibility agency for B2B SaaS”.
Branded prompts are near-certain hits. As SE Ranking’s guide puts it, “when your brand name is in the prompt, AI visibility is nearly guaranteed, which inflates your numbers if these get mixed in with category-level tracking.”
The arithmetic bites hardest on a small set. Take seven tracked prompts where two are branded. Both come back positive almost every run.
Two hits out of seven is 29%. A dashboard averaging all seven together reports that as visibility, while the brand appears in none of the five category prompts.
That number is real. It describes your company name rather than your market position.
Unbranded prompts are the only place a competitive gap is visible at all.
Worked example: auditing a real prompt set
Rampiq tracks 7 prompts. Two are branded. All seven sit in a single category: recommendation and comparison queries about the best AI visibility agency. There’s no funnel split, so every prompt asks a decision-stage question of the same shape.
Against the four types, the set looks like this:
| Prompt type |
Rampiq’s set today |
A rebalanced starter set |
| Informational |
0 |
8 to 12 |
| Comparative |
5 |
10 to 15 |
| Brand evaluation (tracked separately) |
2 |
4 to 6 |
| Transactional |
0 |
3 to 5 |
| Total |
7 |
25 to 38 |
Two of the four types have nothing in them. The branded share is what produces the inflated average above.
A set like this can tell you whether you’re on best-agency lists. It cannot tell you whether an in-house marketing lead, researching how AI citation tracking works, has ever met your name. No prompt asks.
Four questions to run against your own set:
- What share of your prompts carry your company name? Above roughly a fifth, and the headline number is mostly self-reference.
- How many of the four types have zero prompts in them?
- Does every prompt ask a decision-stage question?
- Which buying problem could this set never surface, because nothing in it asks?
If question four has an answer you don’t like, the prompt set is the cheapest thing to fix. A set that’s been wrong for a year usually points at deeper gaps too. That’s what a full AI search visibility audit is for.
We’re rebuilding ours against that table. A tracked set drifts toward whatever the team checked first, and for an agency that’s whether it shows up on best-agency lists.
Keeping the library current as buying questions shift
Refresh cadence is how often you revisit a tracked prompt set as buyer questions and your positioning change. Tie it to prompt group rather than to one calendar rule.
Solis sets cadence by group. Priority commercial prompts get a monthly check. Competitor comparison prompts run monthly or quarterly, product and pricing prompts after any material change, persona prompts quarterly. A full library review lands every six to twelve months.
SE Ranking’s guide sits in the same range for the core set, at every 30 to 60 days. It adds a floor worth respecting: track for at least 30 days before drawing any conclusion.
Keep a stable core so the numbers stay comparable. Swap the whole set each quarter and you get a fresh baseline each quarter, which is the same as having no trend.
Which triggers should move a prompt in or out?
- A competitor starts appearing in answers where it previously didn’t
- You ship a feature or change pricing, creating an objection nobody was raising before
- A fan-out query stops recurring across runs
- Sales starts fielding a question no prompt in the set asks
That last trigger catches emerging concerns, because the platforms won’t catch it for you. Tracking tools observe which sub-queries an assistant expands your prompt into. They log which of those recur across runs, and which sources get cited for them.
A rising concern shows up as a fan-out query that keeps returning with a citation set you’re absent from. Nothing fans out from a question nobody asked. A set that only asks about best agencies will never expand into contract length or data handling.
Ongoing AI visibility optimization without a representative set is optimization against a sample you chose by accident.
A spreadsheet carries this until two product lines and three markets need their own sets, and nobody owns the refresh. That’s where in-house AI search training earns its place. Ours runs in under four hours of live time, and ends with a custom AI visibility audit, a monitoring dashboard and a prompt library your team owns outright.
