Every AI visibility tool asks you the same question on day one: which prompts do you want to track?
It sounds simple. It isn't. Nobody outside OpenAI, Google and Anthropic knows what people actually type into AI tools, and real conversations rarely look like a neat keyword. As Linda Sallai, who co-leads LLM strategy at BrokerChooser, put it, a search used to be "best stock brokers." Now it's something like "what is the best broker for me that also gives fractional shares in India?"
At Ahrefs Evolve, we asked 12 in-house marketers how they handle this. Almost all of them track prompts. Almost none of them fully trust the data.
- Eden Wallace at NRG runs four AI visibility tools: "I don't think anybody has it right."
- James Shaw at The Alliance: prompt tracking is "Directional. It's important. It can tell you where you're going," but it isn't hard data.
- Stefanie Maio at Evans Hotels: "These are just trends."
- Dave Kimball at Invite Networks isn't tracking prompts at all yet.
So if the data is directional anyway, the real question is how you choose prompts that point in the right direction. Here are the five approaches we heard, from the most human to the most technical.
1. Ask your customers how they talk
The strongest answer we heard came from Irina Jordan, VP of Marketing at Checkmate. Her buyers are restaurant executives at brands like Arby's, Wendy's and Buffalo Wild Wings, and they don't phrase questions the way marketers do. So she goes to their conferences and asks what's on their minds, what they're talking about and what triggers a conversation. Then she uses their words as her prompts.
"The curse of knowledge is real. If you make an assumption of what you think they're asking for, you will be so off."
Jonathan Rundle at ChemPoint faces a similar challenge. His buyers are chemical formulators asking very scientific questions: what a chemical will do in a formulation, how to solve a problem, where to source a product. His team builds prompts around those real problems.
Alexandre Teulade at ClearPoint Strategy adds one rule: keep your own brand out of most of your prompts. "Really put yourself into the customer journey, and like, what would someone in our ICP really search for."
2. Start from keywords, then rewrite in a human voice
Several teams start with the keyword research they already have. Linda's team at BrokerChooser takes traditional search queries and uses AI to turn each one into several conversational prompts.
Anton Holm at the International Rescue Committee does the same, building prompts from the IRC's established SEO keywords and priority areas like cash assistance, malnutrition and maternal support. He's fine with AI tools that expand a keyword into a full fan-out of prompts, but he adds a warning:
"Never lose sight of what it takes to create your own prompt and to really put yourself in the mindset of your buyer or your supporter."
Anton also flagged something many teams miss. The IRC runs sites for the US, Korea, Germany, Sweden, the UK and the EU, and AI answers vary by who's asking and where. A prompt that looks great in San Diego may look very different in Seoul.
3. Track topics, not exact prompts
Jordin Frey at U.S. Career Institute stopped chasing individual prompts altogether. "Everyone searches so differently, it's hard to pick a specific prompt and nail it down." Instead, her team tracks coverage of the broader fan-out topics AI pulls from.
Lance Holt at Morning Brew takes a similar approach with a topic matrix he credits to the Growth Memo. One side is the brand's tag taxonomy, or what they cover. The other is personas, or who they're writing for. Together they define the signals Morning Brew wants to send to AI.
4. Prioritize comparison prompts
Some of the most important prompts aren't about what you do. They're about how you stack up.
Dave learned this the hard way. While his company was pitching a project, someone asked an AI tool which of two competing vendors would be better, and it picked the competitor. "So from a prompt standpoint, we're going to be looking at how do we compare against our competitors."
Will Heany at Snowflake sees the same urgency. Head-to-head prompts are often "the final prompt that someone throws into Claude before they decide to make that buying decision."
Eden's reporting shows NRG's strengths and weaknesses in the eyes of AI next to those of competitors X, Y and Z, because "it's no longer about my website. It's about reviews and social and Reddit and everything that is on the internet."
5. Build your own system
The most technical teams are building their own prompt infrastructure.
Eden spent about 100 hours using an AI coding agent to write roughly 80,000 lines of Python. The result is an automated, interactive reporting suite built on Ahrefs Brand Radar data that anyone at NRG can open. Will's team is building prompt tracking inside Snowflake's own data warehouse, which they find "a lot more trustworthy than relying on someone who paid for prompt volume." BrokerChooser built a proprietary tool to measure how often it's cited.
At the other end, Stefanie took a simpler route. Her agency built a small baseline set of generic, brand and location-specific prompts, so the team can see whether its website work moves the needle.
How we'd build a prompt set
Pull the best of each approach together and you get a prompt set that's small, grounded and useful:
- Start with customer language. Mine sales calls, support tickets, reviews and conversations with real buyers. Our team found the same thing Irina did: talking to your audience is still the best prompt research there is.
- Map it to topics and personas. Use Lance's matrix: what you cover on one axis, who you're writing for on the other. Track coverage across the grid, not just a handful of exact phrases.
- Keep most prompts unbranded. You want to know whether you show up when buyers don't already know your name.
- Add your money prompts. Include head-to-head comparisons and "best X for Y" questions for your top competitors. Check them often.
- Test by market. If you sell in more than one country or region, check how answers differ by location.
- Use the data for direction, not promises. Watch trends, gaps and how AI describes you. Report revenue as the outcome.
For more on putting your audience first, read Want to know how to do keyword research in the AI era? Know your audience. For how to track what you find, see how to track brand mentions in AI search.
Better questions beat more prompts
The teams we talked to have tools that can track 25,000 prompts. The ones getting the most out of them aren't tracking the most. They're tracking the right ones: the questions their real buyers ask, in the words those buyers use.
No tool can hand you that list. Your customers can.
Want help building a prompt set grounded in your audience? See our GEO/AEO services or let's talk.
