How to map generative AI prompts across each stage of the B2B sales funnel

Keyword research has always required thinking about intent: someone searching "what is CRM software" is in a very different place than someone searching "HubSpot vs. Salesforce." With generative AI search, these intent categories still hold, but the universe of questions within them has exploded. In an article published on Search Engine Land, Casey Nifong proposes an answer to this problem: prompt mapping, a method for deciding which prompt variations deserve to be measured and how to organize them across the sales funnel.

The starting point is recognizing that "best CRM software" today becomes something like "which CRM is best for a 50-person B2B team that uses HubSpot for marketing and needs better pipeline reporting?" Change the company size, the stack, the pain point, or the priority, and you have another plausible prompt. Since it's not possible to track every variation, the tracking set needs to capture how the buyer's questions and priorities change as they move toward a decision.

What prompt mapping solves that a single score doesn't

For those managing channel and budget, Nifong's central argument is the most actionable one: an aggregate AI visibility index hides where the brand actually shows up.

Say you track 100 prompts and show up in 40% of the responses. Break that 40% down by funnel stage and you might find the brand appears in 70% of brand and comparison prompts, but in only 10% of problem-discovery and solution-research prompts.

>

-- Casey Nifong, Search Engine Land

This breakdown changes the diagnosis. "We show up in 42% of prompts" says much less than knowing where that visibility is concentrated and where it collapses. Prompt mapping is, in essence, the extension of keyword mapping to the generative era: it identifies the questions the audience asks AI platforms and organizes them by topic, intent, persona, and journey stage.

What "showing up" means at each stage

The insight of the piece is that visibility means different things depending on the prompt's intent, and the interpretation of the result changes along with it. The source breaks this down into four stages:

| Stage | Question the brand needs to answer | Example prompt | |---|---|---| | Awareness | Are you part of the problem space? | "Why is churn increasing among our SaaS customers?" | | Consideration | Are you associated with the solution? | "What tools help predict customer churn?" | | Evaluation | Do you make the shortlist? | "Which churn prediction tools integrate with Salesforce?" | | Decision | How is your brand represented? | "How much does [Brand] cost? Is it worth it?" |

The detail that changes how you read the data: at the decision stage, the inclusion rate barely matters. If someone asks about the brand by name, the focus shifts from "did I show up or not" to how the answer represents the company. Is the price up to date? Do the described capabilities match reality? What objections and limitations does the AI raise? Which competitors show up alongside it? Here the enemy isn't absence, it's the wrong answer, something a visibility score never captures.

In evaluation, the warning is about the drop in visibility as criteria get more specific. The brand may appear in "best customer retention software" and disappear as soon as the buyer adds industry, company size, integration, or use case. It's exactly at that erosion point that the map shows where the shortlist drops you.

How to build the map without bloating the list

Nifong is emphatic on a point that matters to anyone with limited time and budget: thousands of prompts aren't necessary. A smaller, deliberate set tends to be more useful than a huge collection of loose questions. The proposed roadmap:

1. Define your business's actual stages. Awareness/consideration/evaluation/decision are the baseline, but a complex B2B purchase may include technical validation, security review, and internal stakeholder approval. The stages should reflect how the customer actually researches and buys.
2. Collect the questions where they already exist. Here's the strongest MarTech insight in the piece: don't build the map in a marketing brainstorm. The sources are sales calls, customer interviews, search data, on-site search, support tickets, reviews, Reddit, industry forums, and People Also Ask. The sales team is especially valuable, because the questions prospects repeat in discovery and demo calls turn into strong consideration and decision prompts.
3. Turn themes into natural language, without mechanically converting everything. From "email security software" come prompts with distinct problems, requirements, and competitors, like "which alternatives to Proofpoint are best for preventing phishing?", not twenty ways of asking the same thing.
4. Add real customer context: industry, company size, persona, geography, budget, use case, integration, business model, pain point. "Best CRM" and "best CRM for a 20-person B2B team that uses Gmail" produce very different recommendations.
5. Tag everything, at minimum by stage and topic, and by intent, persona, product, competitor, and branded vs. non-branded when it makes sense. It's the tagging that lets you discover that a five-percentage-point drop is concentrated in non-branded evaluation prompts for a specific category, and not spread out.
6. Prioritize a tracking core by relevance to the business, commercial intent, and the likelihood a real buyer would ask that question, and then run the same prompts consistently. Since the AI's answer varies from one run to the next, repeatable testing over time is worth more than constantly swapping prompts.

The mistake that undoes the effort, and where it doesn't pay off

The piece lists common pitfalls, and the costliest one is the same one that motivates the method: tracking only category and brand prompts. If the set begins when the buyer already knows the category or the company name, the entire problem-discovery phase stays invisible, exactly where the AI decides whether you enter the conversation. Another one: giving equal weight to every prompt, mixing 50 broad awareness prompts with 10 high-intent ones into a single rate.

There's a counterpoint the source doesn't make explicit, but that follows from its own logic: prompt mapping has a real operational cost (qualitative collection, tagging, repeated runs) and only pays off where AI search already influences the purchase. For operations with very low ticket size, short decision cycles, or an audience that doesn't turn to generative assistants during research, building and maintaining this map can be more effort than it's worth. The article itself acknowledges that the practice is still emerging and that it's easy to put together a set that looks complete without generating useful data.

What changes for those deciding channel and content in Brazil

The practical message is to measure in order to direct content production based on real gaps, not guesswork. Missing from awareness prompts? Content about the audience's problems is lacking. Rarely showing up in consideration? The content doesn't clearly connect the brand to a capability or use case. Competitors winning in evaluation? Comparisons, alternatives, features, integrations, and proof are missing. Gap in decision? Then it's time to investigate the AI's answer itself: outdated information, an important capability missing, a third-party source telling half the story.

The gain is trading "visibility dropped" for "visibility dropped in non-branded evaluation prompts for category X," which is an investment decision, not a lament. The map isn't a final deliverable: buying behavior, products, competitors, and terminology change, and the set needs to evolve while keeping a stable core to measure trends. Starting small, with the question that moves the customer from problem to purchase, is cheaper and more honest than chasing a pretty score that doesn't tell you where to act.