How to Build Prompts to Track AI Visibility

When I talk to teams about how they're tracking AI visibility, the conversation almost always goes the same way. 

They've connected a tool, auto generated prompts from the tool or AI, but when I ask which prompts they're tracking and why, the answer is usually some version of: "We pulled a list of category queries that made sense for us."

That list is the problem.

Generic prompts measure category presence, whether brands in your space are showing up in AI-generated answers. Teams spend weeks setting up tracking infrastructure, generate a 200-prompt list, run it for a month, and the output tells them something like: "three competitors show up more than us." 

Fine. But now what? 

There's no direction in that data because the prompts are rarely built around a specific product, a specific ICP, or a specific funnel position.

A prompt set that produces useful data should be derived from the insights from the product and ICP information. These prompts will give you the accurate benchmark whether you are showing up, for the buyers you are actually trying to reach, on the queries that map to how those buyers actually think and search.

Here's how we go about creating such prompts.

Why we create targeted prompts and not spread wide 

The instinct most teams have when building a prompt set is to go broad. Cover more topics. Get more data points. That instinct comes from keyword research, where surface area matters and volume is a useful signal.

Going wide with generic prompts means you're measuring everything loosely instead of measuring the queries that matter precisely. Keep in mind that LLMs are probabilistic.

Consider this scenario, for Scale Theory few of the themes to focus on AI search are

• AI Search agency
• SaaS SEO agency
• B2B SEO agency
• AI SEO agency for SaaS
• SaaS GEO agency
• SaaS AEO agency

Now if we go about creating 200 prompts covering all the three topics we would end up with around 30 queries per topic. The insights from for each topic would be based out of a lower dataset leaving decision makings to guess and gut feeling. 

So instead of doing that we select one or a maximum of two topics and build prompts around it, this helps us in two ways

• Bigger data set in terms of presence and citation information giving solid direction for content or outreach strategy
• Targeted strategy focussing on the top theme which helps us move fast and see quick results within a couple of months

How to identify the core themes or topics relevant for your business?

Once we have a clear understanding of the product and ICP, we work to identify the core topics that sit at the intersection of the two: topics that are relevant to the product and relevant to the ICP. 

From that filtering, we narrow to the top three themes. Think of these as you would primary keyword clusters in Google search, broad territories that cover a range of specific queries. 

For Scale Theory, one theme might be AI SEO agency. Another might be SaaS SEO agency for startups.

How to pick the top theme to start out with?

Three themes is a starting list. We don't build all three at once.

To decide where to begin, we look at two factors in combination: 

• How much demand exists for that topic across search and AI platforms. The search volume data in Google can be a good pointer to judge the demand of the topic in the internet
• How relevant that topic is to both the product and the ICP.

We select the theme with the best combination of demand and fit. 

A lot of teams make the mistake of trying to track everything at once and never getting clean enough data on any single theme to know what's working. Starting focused means the first two months of tracking actually tell you something.

How we map out the keyword landscape

Different buyers in the same category search in different ways. A Head of Sales searching for a CRM might use entirely different language than a Revenue Operations lead searching for the same product. 

Once we've locked in the priority theme, we use Google Keyword Planner to pull all the related keywords and phrases that fall under it. This helps us to understand the full keyword universe under the topic.

From the full list, we filter down based on relevance to the product and the ICP. High-volume terms that pull in the wrong buyer come out. Terms that reflect genuine purchase intent stay in.

How we build the persona matrix

This is the part most teams skip, and it's the reason most prompt sets end up generic.

A prompt set built with one is built around buyers. These are prompts that reflect how a specific buyer, with a specific problem, in a specific context actually searches that produce tracking data you can do something with.

The persona matrix is a table with six columns: core topic, persona, use case, USP, industry, and location. Every row is a combination you'll use to build prompts from.

Here's what that looks like for an AI CRM product:

The prompts we build from this table are grounded in real buyer context. 

How do we build and distribute the prompts across BOFU, MOFU, and TOFU funnels?

With the persona matrix built, we split prompts across three intent types.

1. Bottom of funnel or high intent solution seeking queries — 70%

BOFU prompts are queries where the buyer is actively looking for a product. They know the category. They are evaluating options and shortlisting providers.

“Best AI CRM software for a fintech company to automate lead generation” is a BOFU query. The buyer is looking for a list of options to compare and find the best tool that fit his needs

Seventy percent of our prompts fall here because this is where commercial impact is highest. A brand that shows up consistently in BOFU responses is on shortlists. One that doesn't is invisible at the moment buyers are ready to make a decision.

For BOFU, we track three metrics:

• Presence (does our brand appear in the response?)
• Owned citations (are our own pages being cited as sources?)
• Brand mention in citations (across all third-party pages cited in the response, how many mention our brand by name?). 

These three numbers are what all BOFU activity maps back to.

2. Middle of funnel or pain point related queries — 20%

MOFU prompts are queries where the buyer is dealing with a pain point and searching for answers, not a product. They may not know the category name yet.

How can I use AI to automate my outbound sequences? is a MOFU query. The buyer has a problem. Our product is a potential answer. But the LLM is more likely to return a workflow or a method than a product recommendation, which means presence here looks different from BOFU.

We track two things in MOFU:

• Presence (are we appearing at all, and if so, how do we compare to competitors?)
• Citations — specifically, the topics that are showing up in cited content across responses. If a topic keeps appearing in MOFU citations that we don't have a page on, that's a content gap with a clear signal behind it.

3. Top of funnel or informative queries — 10%

TOFU prompts are informational queries, but not basic ones. We're not tracking "what is AI CRM?". What we're looking for are the questions our ICP asks when they're trying to understand how to use something or improve something related to their work.

How do companies use AI to streamline lead management? is a TOFU query. The buyer wants to understand a method or approach. They're not looking for a vendor.

At this stage, we're tracking citations. We look at the topics that appear in cited content and use that to guide content strategy. If we consistently see a type of content being cited that we haven't built, that tells us where to go next.

Why do we track prompts at all three funnel stages and not just BOFU?

I've had clients ask whether TOFU and MOFU prompts are worth the effort. The answer is yes, and the reason comes down to how buying decisions actually unfold.

Most buyers don't start with a BOFU query. They start by trying to understand something. A Head of Sales at a SaaS company might search for how to use AI to reduce manual prospecting before they ever search for AI CRM software. Once they understand the approach, they look for whether it applies to their pain point. Once they're convinced it might, they start searching for tools.

The journey moves through TOFU, then MOFU, then BOFU. Users do not search in a linear sequence, they loop back, search in parallel, talk to peers, but that's the general direction.

If you're only visible in BOFU responses, you're showing up at the end of a journey that started without you. The buyer may have already formed a view of the category, already encountered competitors, and already shortlisted options before they got to the query where you appear. MOFU and TOFU are where you build presence before the buying decision is made.

The 70/20/10 split reflects where the commercial impact is highest, not where the entire journey happens. 

BOFU gets 70% because brand mentions there have the most direct impact on whether you end up on a shortlist. MOFU and TOFU get the rest because being absent from those stages has a cost that doesn't show up in BOFU data alone.

How do we act on what the data shows?

Tracking without a clear action framework produces reports, not results. Once the prompt set is running, here's how we translate the data into decisions.

1. Create or optimize content

Across all three funnel stages, we look at which topics appear most often in citations — both our own pages and third-party sources. The question is always the same: do we have a page on this topic?

If we do and it's not getting cited, we optimize it. Better structure, more depth, clearer answers to the specific questions LLMs are surfacing that topic for.

If we don't have a page on it, we create one. The gap between what LLMs are citing and what we've published is a direct content brief.

This applies at every stage. BOFU citation gaps tend to point toward product, comparison, or category pages. MOFU gaps point toward use case and problem-solution content. TOFU gaps point toward informational content tied to the questions your ICP is already asking.

2. Targeted outreach — BOFU only

For BOFU queries specifically, there's a third lever beyond content.

We look at which third-party pages are being cited most often in BOFU responses, these are typically listicles, comparison roundups, or category guides. We check whether our brand is mentioned on those pages. If it isn't, those pages are the priority for outreach.

The outreach is super targeted: we contact the publisher, explain that we want to be included, and offer something in return, like a link from one of our pages, or a paid placement depending on the publication. 

The goal is to get our brand name onto the pages that LLMs are already pulling from when buyers are actively evaluating options.

Why and how does this approach work out?

The reason this approach works where generic tracking doesn't is that every element is connected to a specific buyer making a specific decision. There's no ambiguity about what the data is telling you or what to do next.

Most teams I've seen give up on AI visibility tracking after a few months because the data doesn't translate into direction or the strategy might not have dedicated goals. They're staring at numbers that show presence going up or down without understanding why or what to change.

Building the prompt set the right way is what makes the data interpretable. When your prompts are built around real buyer intent, the gaps they surface point to real actions: a page to optimize, a piece of content to write, a listicle to get onto. The work stops feeling abstract and starts connecting directly to how buyers find you.

That's what the system is built for.