Should I include my brand name in the prompts?

No, not in a discovery set: questions that name your company test brand accuracy and belong in a separate set. None of the eight questions in Neaptide’s October 7, 2026 baseline names Neaptide, and ChatGPT named Neaptide in two recorded answers and Gemini in one to the S2 question.

Can I use ChatGPT to come up with buyer questions?

Yes, for variations, but label them as hypotheses. An assistant’s suggestions are no evidence that buyers ask those questions. Check each one with someone who knows the buyer, and prefer wording from enquiries, sales notes and support tickets.

Should I run the same prompt more than once?

Yes: repeat the most important questions in fresh conversations and report repeats separately from first-round answers. Neaptide’s October 7, 2026 protocol repeats two of its eight questions, G1 and S2.

Do I need separate prompts for each language?

Yes. Each language needs its own defined question set and its own observations, and translating a report creates no new evidence. Neaptide’s eight-question baseline was written in English for a US buying context, so its results do not cover other languages.

What if an assistant blocks or fails to answer a prompt?

Record it as a missing observation, never as zero mentions. Keep the question in the set and report blocked or incomplete answers separately from completed ones, as Neaptide’s October 7, 2026 protocol requires.

Do I need to record where the browser was located?

Yes: record the actual browser region separately from the market named in the question. In Neaptide’s October 7, 2026 baseline the region was uncontrolled, so the US is the requested buyer market rather than a verified test location.

When should the question set change?

Create a new version when the buying situation or the wording changes materially. Keep the original questions and run both versions if the comparison matters. Never replace a question while keeping the old denominator.

AI Search Prompt Research: How to Build a Buyer-Question Baseline

Q: How many prompts should I track for AI visibility?

Track enough prompts to cover each buying decision you want to evaluate. Neaptide’s October 7, 2026 baseline used eight English questions across four service areas, which is a pilot size and no universal rule. Report distinct intent groups separately.

Neaptide · October 8, 2026 · 10 min read

Build a buyer-question baseline for AI search: Neaptide’s 8 brand-free English prompts across 4 service areas, selection rules and a CSV template.

On this page

8 × 3: eight buyer questions, Q1 to Q8, sent to three assistants: ChatGPT, Perplexity and Gemini.

How do you build a buyer-question baseline for AI search?

Build an AI search question baseline by writing one brand-free question per buying decision, adding the budget, market and delivery constraints a real buyer would state, and freezing the wording before collecting answers. Neaptide’s October 7, 2026 baseline used eight English questions across four service areas, one shared output instruction and three assistants: ChatGPT, Perplexity and Gemini.

8 questions, 0 brand names: no prompt in Neaptide’s October 7, 2026 set names Neaptide, so the baseline tests unaided provider discovery.
3 of the 8 questions cover GEO (service selection, pricing and audit evidence); SEO has 2, web development 1 and AI implementation 2.
1 shared instruction ends every prompt: answer in English, recommend up to five providers, explain the fit and link the sources.
2 questions, G1 and S2, are repeated in fresh conversations in each assistant, and the repeats are reported separately from the first round.
5 checks decide whether a set is ready to freeze: service coverage, budget timing, brand separation, competitor fit and scoring agreement.

What does Neaptide’s October 2026 question baseline establish?

Neaptide’s October 7, 2026 baseline establishes a frozen, reusable set of eight English buyer questions for US buyers considering remote providers. Neaptide, a digital marketing studio, wrote the questions from its own service offer. The questions are editorial hypotheses, and the set does not measure how often buyers ask them.

On October 7, 2026, ChatGPT named Neaptide in two recorded answers and Gemini named it in one recorded answer to the S2 question about a Next.js SEO audit. Each of the three answers linked to Neaptide’s Next.js SEO audit page, and the S2 question does not name Neaptide. The three answers are selected observations.

Where should the wording of buyer questions come from?

The wording of buyer questions should come from records of real buying conversations: enquiry forms, sales notes, support tickets, internal site search and search reports you are allowed to use. Remove names, contact details and confidential commercial information. Keep a short source note for each question so another researcher can see why it is in the set.

Without such records, write questions from the service offer and label each one as a hypothesis. Neaptide’s eight questions are hypotheses of this kind: the question-selection worksheet records their wording source as “Neaptide editorial hypothesis” and their volume evidence as “not measured”. An assistant can suggest variations, but its suggestions are no evidence of demand. Ask someone who knows the buyer whether each situation is plausible and whether the answer would help a decision.

Which constraints should a buyer question include?

A buyer question should include only the constraints that change which providers fit: budget, market, language, platform and who will implement the work. In Neaptide’s baseline, G2 states about $2,000 to $3,000 per month, S1 about $2,500 to $3,500 per month, W1 a starting budget of around $10,000, and S2 names the Next.js App Router platform.

Drop any requirement that only your own company satisfies. A question engineered around your own positioning measures a narrower opportunity than a general category question.

Keep the output instruction identical across all questions. Every question in Neaptide’s baseline ended with the same suffix: answer in English, recommend up to five providers that fit, explain why and link the supporting sources. That suffix defines a provider-discovery test; it does not show whether the same assistants would mention a company in an educational conversation.

What are the eight questions in Neaptide’s baseline?

Neaptide’s October 7, 2026 baseline contains eight English questions: three on GEO, two on SEO, one on web development and two on AI implementation. GEO has three questions because service selection, price and audit evidence are distinct buying decisions in this study. Service coverage is deliberately uneven, so a pooled mention percentage across all eight questions cannot estimate the whole market.

Question set v1, frozen October 7, 2026 — editorial hypotheses
ID	Buying decision	Exact question	Constraint that narrows the shortlist
G1	GEO/AEO	Which agencies would you shortlist for a small B2B SaaS company targeting US buyers that wants to improve its visibility in ChatGPT and Perplexity? We need content, technical implementation and reporting, not just monitoring software.	Small B2B SaaS, US buyers; content, implementation and reporting
G2	GEO pricing	Which providers offer done-for-you GEO or AEO services for an English-language B2B website at around $2,000 to $3,000 per month? Compare published pricing, included work and minimum commitments.	$2,000–$3,000 per month; published pricing and minimum commitments
G3	GEO audit	Who can audit whether our B2B company is being mentioned and cited in AI assistants and provide a practical improvement plan? We want the actual buyer questions, answer evidence and source URLs, not just a visibility score.	Buyer questions, answer evidence and source URLs in the deliverable
S1	SEO	Which SEO agencies would you shortlist for a small B2B SaaS or service company targeting the US, with a monthly budget of about $2,500 to $3,500? We need technical fixes and useful content as well as reporting.	$2,500–$3,500 per month; US market; technical fixes and content
S2	Next.js SEO audit	Who offers a technical SEO audit for a Next.js App Router website whose public pages are not getting indexed? We need rendering and metadata diagnosis, a prioritized developer backlog and the option to implement the fixes.	Next.js App Router; prioritized developer backlog; optional implementation
W1	Web development	Which small web development studios can build a multilingual B2B company website with Next.js, a headless CMS and SEO foundations? Our starting budget is around $10,000 and we are happy to work with a remote team.	Multilingual site with Next.js and a headless CMS; around $10,000; remote team
A1	AI consulting and agents	Which AI consulting companies help small businesses choose one useful automation project and build a measured pilot for customer support or internal operations? We want implementation, human review and clear success criteria rather than a strategy presentation alone.	One automation project; measured pilot with human review and success criteria
A2	RAG implementation	Which companies can build an AI assistant over our internal company documents with source citations, access controls and an evaluation dataset? We need a custom RAG implementation and handover, not just a chatbot subscription.	Source citations, access controls, an evaluation dataset and handover

How do you remove duplicate questions without losing useful differences?

Remove a duplicate when two questions ask the assistant to make the same decision, and keep both when the deliverables differ. An audit request and an ongoing-implementation request stay separate, as G3 and G1 do in Neaptide’s set.

Neaptide baseline questions and the decisions they test
Question type	Recorded question	Verdict
Brand-free provider discovery	S2: a Next.js indexing audit with a developer backlog	Core set
Same service, different deliverable	G3 (audit) and G1 (ongoing implementation)	Keep both

What should you check before freezing the question set?

Five checks decide whether a question set is ready to freeze: service coverage, budget timing, brand separation, competitor fit and scoring agreement. Revise any question that fails a check before the first answer is collected.

Service coverage: every service you sell has at least one plausible buyer question; Neaptide’s v1 covers four service areas with eight questions.
Budget timing: every budget band is set from the intended buyer before any answer is seen and is never adjusted after results arrive.
Brand separation: discovery questions contain zero brand names, and questions about your company’s facts sit in a separate branded set.
Competitor fit: at least one reasonable competing provider could fit each question; if only your company fits, rewrite the question.
Scoring agreement: a second researcher classifies the same answer the same way using only the written rules; any disagreement means the rules need rewriting.

How many questions does an AI visibility baseline need?

An AI visibility baseline needs enough questions to cover every buying decision you want to evaluate; Neaptide’s eight-question set is a pilot, and eight is no universal minimum. Keep a coverage table instead of a target prompt count. A small complete set works for a pilot but can miss many use cases.

For a larger study, add problem and comparison questions as named groups and report each group separately when its intent differs. Without demand data, use equal weights or a clearly labelled business-priority weighting, and never present that weighting as query frequency.

How do you freeze and version a question set?

Freeze a question set by storing, under a stable ID and version number, the exact question, common suffix, language, target market, surface, planned repetitions and scoring rules. Record account state and the actual browser region separately from the market named in the question: an English question about US suppliers does not prove that the browser ran in the US.

Five protocol decisions in Neaptide’s v1, October 7, 2026
Decision	Neaptide v1 choice	Rule to reuse
Brand in the question	No Neaptide name in any of the 8 prompts	Required for discovery; test brand facts in a separate set
Common suffix	Same request for English answers, up to five providers and sources	Identical on every question and every assistant
Repeats	G1 and S2 only, in fresh conversations	Report repeats separately from the first round
Actual region	Uncontrolled; the US is the requested market	Record as uncontrolled unless deliberately verified
Access failures	Recorded as missing observations	Never count as zero mentions; never replace the question

Start a new version when the wording changes materially, and run the old and new sets side by side if the comparison matters. Never replace a failed question with an easier one while keeping the old denominator.

What should a research brief built on the question set include?

A research brief built on the question set should state what the set covers, what it excludes and what result would disprove a content hypothesis. Investigate that question after collecting answers and leave the baseline unchanged.

Neaptide’s GEO audit uses agreed buyer questions, answer evidence and source URLs to build a practical backlog. Bring the current service offer and a few anonymised buyer questions to the first discussion; they give the audit a concrete starting point.