Do-it-yourself game plans · No. 06 / 22
Stage 2: What prompts should you track for AI search visibility?
Last updated
Start with buying questions from customers and sales conversations, then fill documented gaps with proposed questions. Every prompt should have an origin and an intended decision use.
Use these provenance labels:
- Customer-derived: recorded in an interview, support exchange or discovery conversation, with personal information removed.
- Sales-derived: supplied by sales or customer success from recurring objections.
- Search-derived: informed by available search or on-site search evidence.
- Analyst-inferred: proposed for testing; not evidence of actual demand.
Do not assign prompt search volumes unless you have a defensible source and understand exactly what it measures.
What does a 24-prompt B2B SaaS test pack look like?
The following is an editable proposed template, not a validated demand dataset. Replace bracketed fields with your actual category and buying requirements.
| Family | Four proposed prompts |
|---|---|
| Category discovery | 1. Which [category] platforms suit a US company with [team size]? 2. What software helps [role] solve [problem]? 3. Which [category] vendors serve [industry]? 4. What options should we consider when replacing [manual process]? |
| Requirements | 5. Which [category] products integrate with [system]? 6. Which support [required capability]? 7. Which publish evidence about [security requirement]? 8. Which suit a budget of [realistic budget and pricing basis]? |
| Implementation | 9. Which [category] tools can our [team] administer without a dedicated specialist? 10. What should we evaluate when migrating from [system type]? 11. Which vendors explain implementation responsibilities clearly? 12. Which products support [specific deployment constraint]? |
| Competitor-seeded comparisons | 13. What are alternatives to [competitor A] for [use case]? 14. Compare [competitor A] and [competitor B] for [requirement]. 15. When is [competitor A] unsuitable for [buyer type]? 16. What should we investigate before replacing [competitor B]? |
| Shortlisting | 17. Shortlist three [category] vendors for [specific brief]. 18. Which providers should we invite to demonstrate [workflow]? 19. Which vendors merit an RFP for [requirements]? 20. Recommend options for [use case], explaining important limitations. |
| Branded accuracy | 21. What does [our product] do, and who is it for? 22. How is [our product] priced? 23. Does [our product] support [integration/capability]? 24. What limitations should a buyer check before choosing [our product]? |
Report these as three cohorts: 16 unbranded prompts, four competitor-seeded comparisons and four branded checks. Do not let the branded questions inflate your unaided discovery score.
For each prompt, store its exact wording, family, provenance, buyer segment, priority, approving stakeholder and version. Freeze the baseline version. Add new ideas to a separate exploratory panel rather than quietly replacing difficult prompts.
How many prompts and repetitions are enough?
There is no validated universal minimum established by the sources reviewed here. A 2026 methods paper found material citation variability, but studied three consumer-product topics and explicitly identified B2B generalisation and minimum-sample guidance as areas needing further work. Its findings support caution, not a universal SaaS sample-size prescription. (arxiv.org)
For a manageable pilot, this guide proposes:
- 24 prompts.
- Three selected surfaces.
- Three repetitions per prompt and surface, on separate days within each wave.
- Four waves: baseline and approximately days 45, 65 and 85.
That is 216 observations per wave and 864 across the programme. These are repeated system observations, not 864 independent prospective customers.
Keep a separate exploratory set for paraphrases and follow-up conversations. Changing wording can be useful research, but it should not silently change the main benchmark.
Sources cited on this page
- arxiv.org arxiv.org/abs/2603.08924