How Many Buyer Questions Should Your Business Track in AI Search? A Practical Sizing Method

Short answer: most businesses should start by tracking roughly four to eight real buyer questions for each distinct buying situation they serve, plus a small, separate set of branded questions. For a single-service, single-market firm with a few well-defined situations, that can land somewhere in the 20 to 40 range. Multi-service, multi-market, or multi-segment businesses scale by unit rather than by ambition, and the practical ceiling is whatever you can re-run on a regular cadence and actually act on. Treat those numbers as a working heuristic informed by published practitioner guidance, not a validated industry standard, because we have not found one.

This guide is for CMOs, SEO directors, owners, and marketing operations leaders who are deciding how to scope an AI search visibility baseline, whether in-house or through an audit and prompt setup service. It covers the three things that matter more than the headline number: the unit you count against, how the questions get chosen, and how the baseline turns into published content instead of a PDF that gathers dust.

Why the number is the wrong place to start

Ask five people how many prompts to track and you will get five confident answers. Publicly available guidance from companies that sell AI visibility tools and services ranges from about 15 to 100 starting questions, and those counts are defined against different units: some mean a grand total, some mean per persona, some mean per topic. None of the sources we reviewed cites a validated threshold, and we are not aware of one.

That is not a reason to shrug. It is a reason to reframe the question. The useful version is not “how many prompts?” It is “how many buying situations do we serve, how many ways do real buyers phrase them, and how much of that can we observe reliably and act on every month?” Answer those three and the count falls out on its own.

A few terms worth defining before you size anything

• Buyer question (or prompt): a phrasing a real prospect might type into an AI assistant or AI search tool, such as ChatGPT or Google’s AI features, while trying to solve a problem or choose a provider. “Who repairs heat pumps in Walnut Creek?” is a buyer question. “Best HVAC” is a keyword.
• Prompt panel: the full set of questions you track, organized into a core set (stable, re-run every cycle, used for trends) and an exploratory set (rotated in and out to test new territory).
• Unbranded vs. branded: unbranded questions do not mention your company name and show whether you are surfaced at all. Branded questions do mention you and show how AI systems describe you when asked directly. Report them separately; they answer different questions.
• Baseline: a point-in-time record of what AI systems were observed saying in response to your panel. It is evidence of a moment, not a prediction of future behavior.
• Visibility gap: a tracked buyer question where competitors, directories, or nobody at all gets surfaced, and your brand does not, or is described inaccurately.

Step 1: Decide the unit you are counting against

This is the single biggest source of confusion in public answers on this topic, and it changes scope materially. Some guidance states a total (“track 40 prompts”). Other guidance states a per-unit figure (“30 per product or buyer segment”). A business with six service lines across four metros reads those two sentences very differently.

Our recommendation is to make the unit explicit before you write a single question. A unit is one distinct buying situation: a service or product, in a market where it matters, for a buyer who has a specific problem. Useful ways to define units by business type:

• B2C Home Services: service line × metro or service area × situation (emergency, replacement, maintenance, financing).
• Professional Services: practice area or specialty × client situation × geography where licensing or local presence matters.
• Growth-stage B2B technology: product or category × buyer segment or persona × decision lens (use case, comparison, integration, security, pricing model).
• Private Equity portfolios: a shared unit template applied across every portfolio company, so results are comparable brand to brand and roll up cleanly.

Then prune. Not every combination deserves tracking. A hypothetical roofing company that does 80 percent of its revenue in two counties should not build units for seven. Pick the units that drive revenue or that you are actively trying to grow.

Step 2: Build the mix inside each unit

Within each unit, the questions should reflect how buyers actually move toward a decision. A simple, defensible mix for unbranded questions looks like this:

1. Problem and discovery questions (the largest share): “Why is my AC running but not cooling?” “What should a small manufacturer look for in cyber insurance?”
2. Use-case and situation questions: “Can a tankless water heater handle a four-bathroom house?” “Which CRM works for a 12-person field sales team?”
3. Comparison and alternative questions: “Heat pump vs. furnace for the Bay Area?” “What are alternatives to [category leader] for mid-market HR?”
4. Local or provider-selection questions (where geography is relevant): “Who are reputable estate planning attorneys near Pasadena?”

Then add a small branded set outside the unit structure, typically five to ten questions total, not per unit: “What does [Company] do?” “Is [Company] a good choice for commercial roofing?” “What do customers say about [Company]?” Keep these separate in reporting so they never inflate your unbranded visibility picture.

For professional services in regulated fields, keep tracked questions informational. You are observing whether AI systems surface a firm when someone asks how a process works, not producing legal, medical, or financial advice. Any content that comes out of the program needs the disclaimers those industries require.

Step 3: Pull questions from real buyer language, not a brainstorm

The quality of a panel depends on where the questions come from. A conference room produces questions marketers would ask. Buyers phrase things differently, and the research we have on prompting suggests they phrase things very differently from one another. In a SparkToro study of AI brand recommendations, 142 survey responses to the same prompting task contained barely two prompts that resembled each other. The author’s conclusion was that any tracked prompt set can only approximate real prompting. Build your panel from the best approximation you have:

• Sales call notes and recorded discovery calls, especially the first question a prospect asks.
• Support tickets, intake forms, and chat transcripts.
• Review text, including the reasons people give for choosing or leaving a provider.
• Objections your team hears repeatedly, phrased as questions.
• Funnel stage: early “what is” questions, mid-funnel “how do I choose” questions, late “who should I call” questions.
• Competitor-comparison language your prospects actually use, including nicknames and misspellings.

Write each question the way a tired homeowner or a time-pressed VP would type it, including the context they would naturally include. “Need a roofer this week, Concord CA, hail damage, insurance claim” is closer to reality than “best roofer Concord.”

Step 4: Size the panel, with the arithmetic in the open

Once units and mix are set, the count is simple multiplication. The table below shows illustrative sizing for four common business shapes. These are worked examples, not benchmarks, and the per-unit figure of roughly four to eight questions is a working convention, not a statistically validated number.

Notice that none of these land at “hundreds” for a single business. There is a reason to resist bigger panels, and it has nothing to do with software limits. Illustrative arithmetic: 40 questions checked across five AI platforms is 200 answer checks per pass. If you repeat each check seven times to get past run-to-run noise (more on that below), you are at 1,400 observations per cycle. Every observation needs to be recorded, read, and compared. A panel you cannot read is a panel you will not act on.

Why a single run is weak evidence

This is where many first attempts at a baseline go wrong. Someone runs 30 prompts once, builds a slide, and declares the company “invisible in AI search” or, worse, “doing fine.” Either conclusion is on shaky ground, because AI answers change between runs of the same prompt.

Several studies point in the same direction, though each has limits worth knowing:

• A preprint from researchers at the University of St. Gallen (not shown to be peer reviewed; the first author lists an affiliation with a company that supplied the data export) ran identical prompts daily across four AI engines for about 45 days. The set of cited sources overlapped only 34 to 42 percent between consecutive days, and the set of brands mentioned overlapped 45 to 59 percent. The authors conclude that a single observation is misleading and that visibility is better understood as a probability across repeated runs. The study used Swiss-German verticals, eight prompts per vertical, and Swiss servers, so apply the specific figures to a U.S. business with care; the direction of the finding is the useful part.
• The SparkToro study linked above, a company-published study that is not peer reviewed (the author discloses that his co-researcher works at an AI tracking startup that supplied data), had 600 volunteers run 12 prompts a combined 2,961 times across three AI tools. ChatGPT and Google’s AI almost never returned the same brand list twice. The study found appearance rate across many runs more stable than position in any one list, and the author judged ranking position to be of little meaning. The study focused on brand and product recommendation prompts, so it speaks most directly to comparison and provider-selection questions.
• At the model level, a peer-reviewed workshop paper found that hosted large language models rarely produce the same response ten times from the same input, even with settings meant to make output deterministic. That paper studies hosted language models in general, not AI search products specifically, so it supports “answers vary” as a general property rather than measuring any one product.

Context may add more variation. OpenAI’s help documentation for ChatGPT search notes that the system may rewrite a question into one or more search queries, may use approximate location for local results, and may let saved memories influence how the query is rewritten. Google’s documentation on AI features describes a “query fan-out” technique in which AI Overviews and AI Mode may issue multiple related searches across subtopics and data sources, and states that the two features may use different models and techniques, so the responses and links they show will vary. One buyer question, in other words, can trigger several underlying searches, and in ChatGPT’s case location and saved memories can affect how the question is handled.

The practical takeaway is the same across all of these: one pass over a panel is a snapshot, not a verdict. Keep the panel fixed and observe it repeatedly, so that skill can be told apart from luck.

Repeat runs and cadence: a sensible default

There is no agreed standard for how many times to re-run each question. The St. Gallen preprint’s own analysis puts the minimum at about seven runs per prompt per day for brand-level estimates and eight when cited sources matter, derived from their data and prompt set. It also recommends rolling aggregation over two to four weeks rather than reading a few days in isolation. The SparkToro author, looking at list-type recommendation prompts, suggests usually at least 60 to 100 runs before averaging. Those are different studies with different designs; cite them as reference points, not rules.

The same preprint also bears on whether a very small panel is enough. It advises against monitoring on one or two prompts, because similarity between runs varied widely from prompt to prompt in its data (roughly 0.2 to 0.8), so a couple of prompts mostly reflect those prompts’ quirks. It recommends a diverse prompt portfolio but does not set a size for one. That is why the sizing in this guide rests on units and on what your team can read and act on, not on a magic count. A panel of ten can be a reasonable pilot, but treat anything it shows as a first look, not a baseline.

A workable default for a business program:

• Run the core panel on a recurring schedule and aggregate to a monthly view. Monthly reporting is a practical rhythm for most leadership teams; the right cadence ultimately depends on stakes and resources.
• Repeat each question several times per cycle rather than once, and report appearance rate (how often you were surfaced) at the question and topic level, not position in a single answer.
• Log run context for every observation: platform, date, location setting, whether memory or personalization was active, and whether the account was logged in. Without that, you cannot tell a real change from a context change.

A quick note on why small counts are noisy, labeled as arithmetic rather than a sourced claim: the worst-case 95 percent margin of error for a simple proportion is roughly ±20 points at 25 observations and roughly ±10 points at about 100. Repeated runs of a fixed panel are not a random sample of your buyers, so do not treat that as a confidence interval for your baseline. Treat it as a reminder that “we appeared in 3 of 10 runs” and “we appeared in 5 of 10 runs” may not be different at all.

What to record for every answer

A mention count alone is thin evidence. For each observation, capture enough to make the next decision:

• Mentioned or not, and whether the mention was a recommendation, a passing reference, or a caution.
• Position or context: first option, one of several, or buried in a caveat.
• Competitors and alternatives named, including directories, marketplaces, and review sites that took the slot. For a closer look at this, see our guide on how to track which brands are mentioned alongside yours.
• Cited sources: which pages the system pointed to, and whether any are yours.
• Accuracy of the description: services, locations, specialties, and claims about you. An inaccurate description can be a bigger problem than no mention.
• Run context, as described above.

Managing the panel over time without breaking your trend line

A baseline is only useful if cycle two is comparable to cycle one. Three rules keep the data honest:

1. Freeze the core set. Changes to wording, even small ones, create a new question. If you must change a core question, retire the old one and start the new one’s history from zero.
2. Use the exploratory set for experiments. New service lines, new markets, seasonal questions, and competitor launches go here. Promote an exploratory question to the core set only after it has earned a place by revealing something you acted on.
3. Keep version history. Date every addition and retirement, and record the reason. When a leadership team asks why visibility on a topic moved, “we changed the question in March” should be a one-click answer, not an archaeology project.

Retire questions that have gone stale (a discontinued service, a market you exited) or that never produce actionable variance. Add questions when sales or support surfaces a recurring buyer concern you are not tracking. Aim for a stable core that changes slowly and an exploratory layer that changes deliberately.

From baseline to action: the part most audits skip

A common concern about buying an audit is a fair one: it can feel like a one-time report. A document lands, a few people nod, and months later nothing has been published. The fix is structural. Every tracked question should carry three fields from the day it enters the panel:

• Owner: who is accountable for the gap, whether that is an internal subject-matter expert, a marketing lead, or a managed partner.
• Gap decision: one of “close it,” “watch it,” or “accept it.” Not every gap deserves content. A question with low commercial value, or one where a national directory has consistently been surfaced in your recorded observations, may be a “watch.”
• Content path: for every “close it,” the specific asset that will address the buyer question, the expert who will contribute, the review step, and the publish date.

Prioritize by commercial weight first (does this question sit close to a purchase decision?), observed gap second (were competitors or directories surfaced where you were not?), and feasibility third (do you have genuine expertise to bring?). This is where a 35-question panel is easier to act on than a 300-question panel: you can look at all 35 in one meeting, decide, assign, and move.

How CiteHarbor approaches audit, prompt setup, and what comes after

The method above reflects how CiteHarbor approaches its AI Search Visibility Audit. We research buyer questions, organize them into units, and observe how your brand is surfaced relative to competitors across five major AI search engines, with branded and unbranded questions reported separately. Findings are documented as point-in-time evidence of what those surfaces showed during the observation window, never as a permanent state or a forecast of future AI behavior. A well-built audit should leave you with a prioritized list of visibility gaps, each with an owner, a gap decision, and a content path.

That list is what feeds the Content Authority System: a program of human-reviewed articles and social content designed to address the specific buyer questions where the evidence shows you are absent, inconsistent, or described inaccurately. It is designed to build the foundation for stronger visibility in AI-generated answers; it does not guarantee citations, recommendations, rankings, leads, or any specific outcome. The content is grounded in the business information and subject-matter expertise you provide, and CiteHarbor handles research, drafting, editorial review, and publishing operations.

Monthly tracking measures your citations against your baseline, monitors competitor citation visibility privately, and emails a branded PDF performance snapshot. In the managed-service model, clients are not required to log into dashboards or approve individual updates, and every piece of content is still published subject to verified client authorization, editorial and industry guardrails, and the publishing permissions you configure. For leadership teams that want AI visibility coordinated with their broader marketing priorities, CiteHarbor also offers a Full Service Partnership.

Evaluating an audit and prompt setup service: questions worth asking

Whether you talk to us or anyone else, these questions separate a credible baseline from a slide deck:

• Can we see the full prompt set, and how was each question sourced?
• How is the unit defined, and is the count stated per unit or as a total?
• Are branded and unbranded questions reported separately?
• How many times is each question run per cycle, and is the run context (platform, date, location, personalization) logged for every observation?
• Are findings presented as dated, point-in-time evidence with sources, rather than as permanent conclusions or forecasts?
• Does every visibility gap come with an owner, a gap decision, and a content path?
• Who reviews content before it is published, and how is client authorization verified?
• Does the provider avoid guaranteeing citations, rankings, leads, or other specific outcomes?

A credible baseline shows its work: the questions, the units, the repeated observations, and the limits of what it can tell you. If a provider cannot show those, treat the headline number with caution.