Prompt research is the process of identifying and tracking the questions that cause AI systems to compare options and recommend specific brands.
AI prompt research serves the same foundational role for AI visibility that keyword research serves for SEO and PPC, but the unit of measurement is different. Instead of pages and queries, prompt research focuses on how AI systems form and present recommendations.
In AI SEO, visibility only matters when AI is evaluating choices. That’s when it weighs alternatives, applies constraints, and points someone toward a solution. If your brand isn’t present in those moments, it won’t factor into the decision.
Most prompts never reach that stage. They generate explanations, summaries, or general guidance. Prompt research filters those out and focuses on middle- and bottom of the funnel (BOFU) prompts: comparisons, evaluations, and “best” queries where AI weighs alternatives and recommends solutions.
To show how prompt research works in practice, I’ll walk through the exact process I use to track Semrush’s own LLM visibility growth.
My prompt research process follows four steps:
- Identify target audiences and buyer personas
- Describe solutions and how they help those audiences
- Use keyword research as supportive language input
- Turn those keywords into prompts
Once you have that set, step five is tracking it over time.
The process scales in both directions: A small team can run it on a focused prompt set; a large organization can run it across many products, markets, and competitors.
By the end of this guide, you’ll have a repeatable way to build a prompt set that shows where your brand competes and where it doesn’t. But first, let’s clarify how prompt research differs from the keyword research you already know.
How prompt research differs from keyword research
Prompt research targets conversational queries where AI weighs options, while keyword research targets the short phrases people type into a search engine.
For search marketers, prompt research introduces a familiar concept with new challenges. Unlike keyword data, we don’t have years of historical search volume, CPC, or trend data for AI prompts.
Because of that, prompt tracking doesn’t behave like keyword tracking.
SEO rankings tend to be relatively predictable. AI-generated answers are volatile and personalized. Prompt research focuses on direction and pattern recognition, not fixed positions or precise counts.
The contrast becomes clearer when you look at how the two approaches differ in practice.
Even with these differences, the objective hasn’t changed. You’re still defining a set of target questions, improving your visibility around them, and measuring performance over time.
What has changed is how visibility is discovered and evaluated.
Semrush has built a prompt database informed by real clickstream data from ChatGPT and other AI platforms, allowing you to estimate topic volume as it happens on LLMs.
Is keyword research still relevant?
Keyword research is still relevant and plays an important supporting role because it reveals how people describe problems and what intent sits behind their searches.
Those signals help you decide which prompts are worth targeting. The difference is that keywords are no longer the endpoint; they’re a language input that gets rewritten into natural, conversational prompts.
The larger shift is what you optimize for.
Instead of tracking “wins” the way you would in SEO, prompt research looks at which topics, constraints, and personas consistently recommend your brand, and where it fails to appear. That’s why prompt research prioritizes the ideal customer profile (ICP), or type of customer a product is built for, over cost-per-click.
The guiding question changes from which terms are cheapest or highest volume to whether your brand appears for the types of intent that reflect real buying situations.
Tracking AI responses over time makes that visibility observable. Daily snapshots of AI answers create a running record of how your brand is framed, compared, or omitted across decision-oriented prompts.
With that foundation in place, the next step is building a prompt set that reflects how your buyers actually make decisions.
1. Identify your target audience
Identify your target audience so you can define the specific questions they might be asking.
A generic persona like “dog owners” will surface generic prompts like “what’s a good dog food?” AI answers this question with well-known, big-name brands, the safe broadly-known options, not products matched to any particular need.
However, a detailed persona like “dog owners of large dogs (60+ lbs) with allergies to major proteins” might reveal prompts like “what’s a good dog food for large breeds who are allergic to chicken, lamb, and beef?” AI answers this by naming specific products that fit those constraints, since generic brands won’t satisfy them.
Constraints are what push AI systems out of explanation mode and into recommendation mode.
Before generating prompts for LLM tracking, focus on the persona traits that change how AI evaluates options:
- Context & experience level: who’s asking and in what situation
- Primary risk or pressure: what they’re trying to avoid or resolve
- Language & expertise: casual vs. technical phrasing
- Budget expectations: affordable, mid-range, or premium
Where to tap into persona characteristics
Tap into persona characteristics in open, unfiltered spaces like message boards, reviews, and support discussions where people talk about products in their own words.
Personas that consistently uncover risk management, trade-offs, and uncertainty reduction create the strongest foundation for prompt research. Their constraints naturally force AI systems to compare options and make recommendations.
2. Connect your product’s solutions to your persona’s problems
Connect your product’s solutions to your persona’s problems so AI can understand why your product is the right fit for a buyer’s specific situation, concerns, and decision-making needs, not just what features it offers.
When people ask AI to help them choose between options, they’re rarely comparing feature lists. They’re trying to decide whether a product fits their situation, reduces risk, and feels like a safe choice.
AI recommendations tend to reflect that behavior. Brands are suggested more often when their products clearly resolve the specific hesitation a buyer feels at the moment of decision.
So, make sure your content includes these product details:
- Features: What the product delivers in concrete, factual terms AI can reference directly (e.g., “single-protein formulas,” “SOC 2 compliant,” “native Shopify integration”)
- Benefits: The outcomes those features produce for the persona, translating features into results that reduce concern (e.g., “easier digestion,” “faster onboarding,” “lower implementation risk”)
- Use cases: The specific situations where the product fits cleanly, helping AI match solutions to scenarios (e.g., “for sensitive stomachs,” “for small teams,” “for regulated industries”)
- Problems resolved: The specific risk, friction, or uncertainty the product removes, often the strongest recommendation trigger (e.g., avoiding allergic reactions, preventing costly mistakes, reducing vendor lock-in)
- Fit factors: The signals that make the option feel safer or smarter than alternatives, such as clarity, simplicity, consistency, or alignment with buyer expectations
The details also need to extend beyond the content on your own site. Your product needs to be described consistently across the web, including review platforms and comparison content, as AI pulls from multiple sources when comparing options.
Validating which attributes matter in AI comparisons
Validating which attributes matter in AI comparisons matters because it tells you which features and benefits actually drive AI’s brand recommendations, not just which ones you assume do.
If you need help determining which attributes are driving persona preferences, use Brand Performance in the Semrush AI Visibility Toolkit. This tool shows which features AI already emphasizes when comparing brands in your category.
For example, for the business Dover Saddlery, AI consistently explains its recommendations using operational fit indications, like “one-stop assortment breadth” when buyers need multiple items at once and “expert fitting and consultative support.”
These are the reasons AI gives when justifying why Dover is a viable choice in a specific decision context. Collectively, they position the brand as a dependable, expert outfitter which is the signal AI needs to recommend a retailer when the buyer’s priority is reliability over exploration.
These attributes become the building blocks for prompt generation. When you feed persona constraints and product fit factors into an LLM, you give it the context it needs to generate decision-stage prompts, not generic questions.
3. Use keyword research to support prompt discovery
Keyword research supports prompt discovery by confirming how your audience naturally frames problems rather than estimating demand.
Tools like our Keyword Magic Tool reveal patterns in language, including:
- Which constraints appear repeatedly
- Which modifiers feel natural versus technical
- Which brand-plus-ingredient combinations show up consistently
Start with a topic tied to a constraint. In this case, “dog food ingredients” reflects how ingredient-sensitive buyers might frame the problem.
Phrases like “limited ingredient dog food,” “best limited ingredient dog food,” and “limited ingredient dog food for allergies” recur across commercial and mixed-intent searches.
This consistency indicates how buyers in this niche phrase their options and modifiers and gives you a starting point for prompt research.
4. Turn your keywords into prompts
Turn your keywords into prompts by using Semrush’s Prompt Research tool to see which specific prompts users use with your above keywords.
For example, I entered “limited ingredient dog food” in the Prompt Research tool.
In the “Topics” view, AI clusters the category around formulations and brands, including hypoallergenic diets, limited ingredient products, and brand-specific variants. That structure indicates the “limited ingredient” topic already aids decisions, making it a strong candidate for a BOFU prompt.
Review the “Intent” column to verify if a topic leans towards BOFU prompts. Hover over the intent bar to view a topic’s breakdown and look for topics with commercial and transactional intents.
Expand any potential BOFU topics to get specific prompts. Look for prompts with a high number of brands mentioned.
When brand mentions appear consistently, and the questions reflect a real choice being made, you’ve reached a prompt worth tracking.
How smaller teams can approach prompt research
Smaller teams can approach prompt research by narrowing the process to a handful of priority products, one or two personas, and the problems that create real buying pressure — then growing the set once it proves out.
Prompt research doesn’t require full market coverage to be useful. A five-person team selling scheduling software gets more value from tracking twenty prompts tied to its actual buyers than from trying to map every conceivable AI query in its category.
Narrow scope also keeps the work sustainable. A small prompt set is one a team can review weekly, update as messaging changes, and act on — a sprawling one gets built once and ignored.
Small teams can follow this four-step workflow:
- Pick two to three priority products or services — the ones driving the most revenue or growth right now
- Document one or two personas in detail, focused on the constraints that push AI from explaining to recommending
- Use Prompt Research in the AI Visibility Toolkit to shortlist candidate prompts, checking AI volume, topic difficulty, and intent for each
- Save the shortlisted prompts as the team’s tracked set, and revisit them monthly. Skip the temptation to expand into every possible prompt in the category before this set has proven its value.
Prompt Research shows which candidate prompts are worth tracking before a team commits resources to them, since it surfaces AI volume, relevance, and intent for each query. That reduces the guesswork of picking prompts based on instinct alone.
How larger organizations can scale prompt research
Larger organizations need to scale prompt research because they’re tracking visibility across more products, personas, markets, and competitors than any manual process can cover.
A five-product SaaS company selling into three regions might need separate prompt sets for each product line, each buyer persona, and each market’s competitive set. Manually brainstorming that many prompts isn’t realistic.
Enterprise AIO‘s Prompt Generation automation solves that volume problem directly. It builds persona-tailored prompt lists from your brand name and topics, the same way keyword research tools build keyword lists from a seed term.
Here’s how:
- Enter your brand name and up to 10 topics — specific phrases, product categories, or comparison terms like “Mailchimp alternatives” work best; each topic generates roughly 20 prompts, so a full 10-topic setup produces 200+
- Layer in persona attributes for each segment worth tracking separately — goal, budget sensitivity, experience level, and decision factors — since a budget-conscious beginner and an expert premium-focused buyer ask AI very different questions about the same topic
- Review the Generated Prompts table and select the ones that reflect real buying situations, then use the “Start tracking selected prompts” button to move them straight into your AIO project
- Tag prompts by product line, persona, or market as you import, so segmented reporting is built in from the start
That segmentation is what makes the dataset usable at scale. A regional marketing lead can filter to their market’s tagged prompts; a product team can filter to theirs — without either one re-running research from scratch.
Once a prompt set is tracked, Query Fan-Out Analysis adds a second layer of visibility. This tool reveals the background searches AI systems run to build their answers.
When ChatGPT or another AI platform responds to a prompt, it typically breaks that prompt into several related queries, retrieves results for each one, and synthesizes them into a single answer — a process known as query fan-out.
For example, when someone asks “best limited-ingredient dog food for allergies,” AI systems like ChatGPT and Google AI Mode break that question into multiple sub-queries, which could be:
- Hypoallergenic dog food recommendations
- Single-protein dog food brands
- Grain-free dog food for sensitive stomachs
- Dog foods without common allergens
The AI then retrieves information for each sub-query and merges it into a single response. This process allows AI to provide richer, more specific answers, even when no single source directly addresses the original query.
Track these variations to see how well you appear for all queries related to intents. If your brand appears across variations, you’ll have a better chance of being recommended.
This approach mirrors how AI systems actually process queries, helping you build a prompt set that captures the full range of sub-queries AI might generate when evaluating your category.
The goal is the same as it is for smaller teams: use the broader, tagged dataset to see where different audiences are already asking about the brand, category, and competitors before deciding which prompts are worth tracking closely.
5. Track your prompts and measure visibility over time
Tracking your prompts and measuring visibility over time shows you which prompts you appear in and which ones you don’t, so you can adjust your strategy accordingly.
Start by configuring Semrush’s Prompt Tracking. Once your campaign is running, Semrush checks these prompts daily and records whether your brand appears in the AI-generated response. You’ll see AI Visibility, Mentions, Source Pages, and Average Position from the Landscape tab.
This helps you measure where you’re present, where competitors are winning, and where you’re missing visibility.
It’s also easy to generate a PDF from your tracking campaign to report on your progress to other stakeholders.
How many prompts should you track?
The AI Visibility Toolkit includes a fixed allowance of tracked prompts: small teams can track up to 25, and enterprise setups run from 500 to 15,000 depending on configuration. Our guide to which AI search prompts to track covers how to choose them.
With a smaller prompt allowance, focus on prompts that might recommend your products or services that drive revenue. Based on our internal testing, 10 well-chosen prompts per product are usually enough to see whether AI systems consistently recommend your brand or default to competitors.
With a larger allowance, add prompts only where evaluation criteria change, like persona, industry, or use case rather than using small wording variations that usually produce the same AI behavior and don’t create new signals.
Turn the growth of AI into an actionable signal for your marketing
As AI platforms influence more buying decisions, many brands still don’t know whether they’re being recommended or overlooked. Prompt research addresses that uncertainty by focusing on the moments where AI evaluates options and recommends a solution.
With Semrush, those decision moments become measurable signals you can monitor, interpret, and act on over time.
Start by documenting one persona and generating 10 decision-stage prompts this week with Semrush’s Prompt Research.
Then add them to Prompt Tracking in Semrush’s AI Visibility Toolkit to monitor where your brand appears, where it doesn’t, and how AI frames your category.
From there, AI visibility becomes something you can work with, not guess at.
