Build an AI Visibility Prompt Set That Measures Buyers

Before we decided to build our own AI visibility tracking system, we signed up for more than a few AI visibility tracking tools and tested them for months on various clients to select the right tools for our agency and clients.

In every AI visibility tracking system we tried, we kept running into the same problem: prompt generation.

Every tool we tested either asked for a set of topics or a set of keywords, or simply asked you to supply your own prompts. And when one did generate its own prompts, it would only do it either from keywords and topics that you provided or from a cursory scan of the website to identify some topics and generate some prompts. In all cases, we found it either creates a lot of work or results in sub-par measurement.

In every case, we found that prompts generated from keywords or topics were far too generic, and it has had us questioning and refining the best way to generate prompts ever since. Prompt generation is the first, arguably hardest, and most important input to AI visibility measurement. It's very different than keyword tracking, where it's relatively easy to understand the intent of a keyword, and it's relatively easy to identify the search volume of a keyword.

Generating useful AI visibility tracking prompts is harder than selecting keywords

Being in SEO for over 25 years, I know first hand the care and attention that goes into selecting the right keywords. You need to balance volume and intent. Higher volume is not always better. The whole game is ensuring an accurate understanding of the intent of the keywords and matching it to the business objectives. Getting prompts right is harder.

We quickly realized that the first-party data we had for our clients was the first part of the solution.

There are no prompt volumes for B2B and never will be

Prompt visibility tracking is particularly difficult because prompts are very long, which means almost every prompt is a one of one (SparkToro).¹ No AI company publishes how often a prompt is asked. A few tracking tools sell estimates built from browsing panels, and even they call the numbers directional. Those estimates are more useful for broad, high-volume questions than for niche B2B categories, where a panel may capture few or none of the prompts your buyers ask. And even if the numbers were reliable, it would be next to useless because of the problem of every prompt almost being unique. Volume proxies, however, do exist one level up. A cluster of similar prompts carries the demand of the keywords it's built around, so the keyword volume can weight a cluster of prompts, just not the individual prompt.

Prompt results are responses to questions more specific than they appear

Then there's also the additional issue that every chat is informed by the entire knowledge and memory within your project and your chat. AI tools know a lot about you and what you care about before you ask a question like, "What's the best software for a company like me?" So you have to infer all of the information that your ideal customer might have already put into a chatbot to generate questions that would be a relatively accurate representation of your potential buyer asking a question.

This is a relatively new field, but an increasingly hot topic and therefore there are an increasing number of studies on this. It comes back to general marketing research and knowledge. So I've put this article together as an attempt to document where we are and provide a basis of best practices for prompt generation.

Here is what the research shows:

• Generic prompts favour category leaders. Add who is asking, and mid-market brands move the most.
• The buying stage changes what an AI answer contains. Early questions cite sources; late questions list vendors.
• Naming your brand in a prompt inflates the result. Keep branded and unbranded prompts apart.
• Single prompts are noise. Read clusters of prompts over weeks.

Two findings sit under all four. Adding who is asking to a prompt changes which brands an assistant recommends,² and adding one constraint, such as "for a SaaS startup," moves the list more than rewording the question does.³

Generic prompts measure the category leader

The prompts you track decide the number you get. A low visibility score may say more about the prompts than about you.

AI answers change when the question says who is asking. In a 2026 study of about 2,000 assistant runs, adding a buyer persona to "best CRM software" changed which brands were recommended (arXiv preprint).² Category leaders kept about 80% of their slots no matter who asked. Mid-market brands swapped up to 75% of theirs.

That gap matters if you are not the leader. A generic prompt asks the assistant for its default list, and the default list belongs to the brands everyone already knows. A challenger wins on specific questions: a given company size, a given integration, a given industry. Track only generic prompts and you measure the one game you were never going to win.

The researchers work at a commercial AI company, and their papers are preprints, so treat the exact figures as early evidence. The direction matches independent work on how user context shifts chatbot recommendations (arXiv preprint).⁴

What makes a good prompt

A good prompt has four parts: who is asking, the buyer's situation in their own words, one constraint, and no brand name.

Who is asking. A B2B purchase is a group decision. A 2026 survey of about 18,000 buyers puts the average at 13 internal participants (Forrester).⁵ Each of them asks AI assistants different questions. A CMO asks how to explain marketing's contribution to the board. A marketing operations manager asks whether a tool will keep attribution intact through a CRM migration. A CEO asks which channels to cut. Adding that context changes which brands come back.²

Write the context the way a buyer would say it. "I run marketing ops at a 200-person SaaS company on HubSpot" is how people talk to assistants. "Persona: Marketing Ops Manager" is a label, and research on persona prompting shows the way a persona is phrased changes the result (arXiv preprint).⁶

The buyer's situation. Buyers describe situations. Product teams describe categories. A prompt set written by the product team asks for the best AI visibility platform. A buyer asks why they rank on page one in Google but never appear when they ask ChatGPT for recommendations.

One constraint. Add an integration, an industry, a team size or a compliance requirement. One constraint moved the vendor list more than rewording the question did.³ Constraints are where specific vendors win, and they are the reason a buyer would choose you over the default.

No brand name. A prompt that names your company measures recognition. Discovery needs prompts that leave your name out. In one 2026 study of 150 brands, assistants described 96% of brands accurately when asked about them by name. Yet 89% of those brands never appeared in answers to unbranded category questions (Victorious).⁷ Track branded prompts as their own series and use them for a different job: checking that assistants describe your pricing, integrations and positioning correctly.

Give that series a deliberate mix: prompts that check whether you appear, prompts that test sentiment, such as how you compare and why buyers leave, and prompts that check facts like pricing and integrations. Include a few designed to surface criticism, so you see it before a buyer does.

Map prompts to the buying jobs

The questions buyers ask change as they move from noticing a problem to choosing a vendor. B2B buying moves through six jobs: problem identification, solution exploration, requirements building, supplier selection, validation and consensus creation (Gartner).⁸ AI answers change with them.

Early questions rarely produce vendor names. Answers to problem-awareness questions named a brand 0.10% of the time, against 1.29% for category-research questions (Victorious).⁷ The early answers cite educational articles, institutions and video. The later ones cite directories, comparison sites and listicles.

This is the part most prompt sets miss. An early-stage prompt that never names a vendor is still worth tracking, because the sources it cites shape how the buyer frames the problem. In one survey, 95% of B2B buyers chose a vendor from the shortlist they had on day one (6sense).⁹ The shortlist forms in those early answers.

So each stage needs its own measure:

Weight the middle. Buyers use AI assistants most in the middle of the purchase, and comparing vendors is their top task (6sense).⁹ A starting mix of about one third early-stage prompts and two thirds requirements, shortlist and alternatives prompts fits most mid-market software categories. Shift toward early-stage prompts if your category is new and buyers do not yet know what to call it.

Assistants also steer the conversation. For example, Gemini often ends an answer with two or three suggested next questions, and a one-turn prompt set misses that. The solution is to write later-stage prompts for where the assistant sends buyers, not only where they start.

Build prompts for every seat, too. Purchases of software with generative AI features average 14-member buying groups, against 7 for other purchases.⁵ One persona's prompts cover one seat at that table.

How many prompts to run

Run one cluster per content topic, about 24 prompts each. If your content plan has six topic areas, your prompt set has six clusters, so the visibility for a topic, the sources cited for it and the content you publish about it sit on the same line of the report.

A starting cluster: five problem prompts, three exploring, five requirements, six shortlist, three alternatives and two consensus. Track branded prompts separately. Across six topics plus a few category-level prompts, that is about 150 prompts, each categorized or tagged to document which persona, problem or competitor it came from.

Clusters matter because single prompts are noise. Run the same prompt repeatedly and the identical brand list comes back less than once in 100 runs.¹ The standard error on a brand's result from one run of one prompt is about 0.37, falling to about 0.08 at seven runs.¹⁰ Repeating a prompt helps, and pooling a cluster of prompts over several weeks builds the sample further.

Hold a stable core. Keep most of each cluster fixed for at least a quarter, because rewording a prompt starts a new series.

How to read the results

Report a rate, not a rank. The useful number is the share of responses in a cluster that name or cite you, on one AI platform, over a rolling four weeks. Rank position is the least stable measure of all.

Know the margin. This is arithmetic, not a study: at a 20% mention rate, 100 responses carry a margin of about plus or minus 8 points. Detecting a move from 20% to 30% takes roughly 300 responses in each period. One run of one prompt, with a standard error of about 0.37, tells you almost nothing on its own.¹⁰

Report each AI platform separately. ChatGPT's cited domains turn over about 74% week to week, while Google AI Overviews turn over about 5% (SISTRIX).¹¹ Blending them hides both.

Mark model changes. When ChatGPT switched models in May 2026, citation drift reached 47%, against a normal day-to-day change of 1 to 2% (SISTRIX).¹² A trend line that crosses a model change without a marker will mislead you.

Remember what any tracker can see. No tracking method sees exactly what a logged-in buyer sees, since assistants personalize answers with location, memory and history. Every tracker samples. The question is whether the sample is designed.

What not to do

Five mistakes will undermine a prompt set.

Naming your own brand in discovery prompts. It tells you whether an assistant recognizes you. It can't tell you whether a buyer who has never heard of you will find you.

Running head-to-head prompts that leave you out. An assistant answering "Competitor A vs Competitor B" has no reason to mention you, so the prompt spends budget and returns nothing. Ask "alternatives to [competitor] for [use case]" instead, where the answer is a list you are trying to join.

Treating one run as signal. A single run of a single prompt carries a standard error of about 37 percentage points.¹⁰ If a prompt cited you this week and not last week, that is noise until the whole cluster moves.

Choosing prompts because you already appear in them. Picking the questions you win inflates the baseline and hides the gaps. Note where each prompt came from, and let the buyer's questions decide what goes in.

Tracking prompts where the answer is something else. A platform for booking event venues gets nothing from "best event venues," because the answer lists venues. "Best platform for booking event venues" is the prompt it can win. Before you track a prompt, check that its answer could name a company like yours.

Where prompts come from

The best prompts come from your buyers' own words, and from what your business contains.

Buyer prompts look different from keywords. For most of late 2024 through early 2026, between 65% and 85% of ChatGPT prompts matched no keyword in a clickstream database of more than 27 billion keywords (Semrush).¹³ The share of keyword-like prompts has been rising, so include both shapes: some short category prompts and many written as situations.

Keyword volume still has a job. No one can tell you how often a single prompt is asked, so read results across a category of prompts. Build each cluster from a high-volume primary keyword and let that volume weight the cluster.

Persona pain questions. If your persona documents list the questions each buyer asks, you already have a draft of your best prompts. Add the persona context and one constraint.

Sales and support transcripts. Pull the most common questions from the last quarter of early sales calls and support tickets, and map each to a buying job.

Search Console. Long-tail queries with impressions but few clicks are questions buyers ask that your content doesn't answer well yet.

Bing Webmaster Tools. Its AI Performance report lists grounding queries, the phrases Copilot used to retrieve the pages it cited.¹⁴ They are the assistant's own searches, and they show what it needed to find before it answered.

New customers. When a deal closes, ask what the buyer searched for or asked an assistant when they started looking.

A map of your market. Keep a structured list of your products, competitors, integrations and the industries you serve. Competitors become anchors for alternatives prompts. Integrations become requirements prompts, such as tools that report AI visibility inside HubSpot. Markets become shortlist prompts, such as the best option for government software vendors.

What a designed cluster looks like

Here is part of one cluster, using our own topic "AI search visibility" and our own buyer personas.

How PureSEM builds your prompt set. Your prompt set is generated from your Content Hubs and your buyer personas. It never names your own brand, and it asks alternatives questions anchored on the competitors you track rather than head-to-head comparisons. New prompts can be added for a new topic or product line without disturbing the set already in place.

On a self-serve plan, the set is generated for you and runs without a manual review.

With managed services, we dig deep with you before anything is measured. We work through your buyer personas until they're right, make sure you have robust product messaging, and check that your competitors are tagged correctly. Then subject matter experts review the prompt set so everything is right before we start measuring.

The method above is a starting point. Your first prompt set will miss questions, overweight some stages and underweight others. Run it for eight weeks, look at which clusters move and which stay flat, and adjust. Bad prompts produce a number that tells you nothing. Good prompts tell you where to write, where to strengthen and where you are already winning.

If you want to see what your AI search visibility looks like today, we can show you.

Sources

1. SparkToro, AIs are highly inconsistent when recommending brands or products, https://sparktoro.com/blog/new-research-ais-are-highly-inconsistent-when-recommending-brands-or-products-marketers-should-take-care-when-tracking-ai-visibility/
2. arXiv preprint 2605.30207, Persona Conditioning of Brand Recommendations in Retrieval-Augmented Commercial Chat (2026), https://arxiv.org/abs/2605.30207
3. arXiv preprint 2605.27440 (2026), https://arxiv.org/abs/2605.27440
4. arXiv preprint 2410.05613, https://arxiv.org/abs/2410.05613
5. Forrester, The State of Business Buying, 2026, https://www.forrester.com/blogs/state-of-business-buying-2026
6. arXiv preprint 2601.18572 (2026), https://arxiv.org/abs/2601.18572
7. Victorious, Quarterly Search Report, https://victorious.com/quarterly-search-report/
8. Gartner, The B2B Buying Journey, https://www.gartner.com/en/sales/insights/b2b-buying-journey
9. 6sense, 2025 B2B Buyer Experience Report, https://6sense.com/science-of-b2b/buyer-experience-report-2025/
10. arXiv preprint 2604.07585 (2026), https://arxiv.org/abs/2604.07585
11. SISTRIX, AI citation drift, https://www.sistrix.com/blog/ai-citation-drift/
12. SISTRIX, citation drift across ChatGPT's May 2026 model change, https://www.sistrix.com/?p=179022
13. Semrush, ChatGPT search insights, https://www.semrush.com/blog/chatgpt-search-insights/
14. Bing Webmaster Blog, Introducing AI Performance in Bing Webmaster Tools (February 2026), https://blogs.bing.com/webmaster/February-2026/Introducing-AI-Performance-in-Bing-Webmaster-Tools-Public-Preview