How to generate (good) AI prompts to track

Sorry for the delay in delivering this newsletter — I was dealing with a health issue that made writing difficult.

Back on track now! Expect newsletters every two weeks or so.

In this post, I’m covering something that’s fundamental to effective AI visibility tracking, but is really easy to get wrong or incomplete: identifying the best prompts to track for your business.

There are three big mistakes I see B2B marketers make when they’re figuring out which prompts to track in an AI visibility tool.

There are three main types of prompts, each of which corresponds to a different part of the buyer funnel. Only one type is useful for tracking AI visibility, and most businesses are wasting effort on the wrong one.

Informational prompts cover every sort of non-commercial question someone might have about a given industry. They tend to reflect questions about definitions (e.g., “What’s a CRM?”) or process (”How to motivate a sales team?”).

There are two big reasons the majority of brands shouldn’t track — let alone optimize for — informational prompts:

The exception is if your brand is operating in a new product or service category that has low awareness rates and isn’t already saturated with content. In this case, influencing AI responses to informational prompts means you can influence future prospective customers on the differentiated way you think about solving problems / adding value in your category.

A commercial prompt reflects a prospect with buying intent. The prospect wants to be guided to a paid solution (i.e., product or service) that fits their needs.

Most AI responses to commercial prompts will include a list of recommended brands. Most of your tracked prompts should be commercial.

There are three main flavors of commercial prompts:

As the name suggests, brand-aware prompts reflect someone who knows a brand exists and wants to learn more before purchasing.

Brand-aware prompts are important for assessing the accuracy of AI responses about your brand.

I understand the inclination to seek out search volume data in the AI prompt context — it could help prioritize your work. But AI search volume is just not a thing that exists in the way we think about it in the traditional Google search context.

The queries that people submit to AIs are much longer than the ones submitted to Google. We’re talking about ~60 words on average for the former vs. 3.4 words for the latter.

Consequently, most AI prompts tend to be nuanced and unique.

For example, last week I wanted to see if an online transaction for one of my clients was protected if we paid a vendor via Paypal.

On Google, I probably would have searched “paypal payment protection,” which has an average monthly search volume of 150 in the U.S., according to Ahrefs.

But here’s what I entered into Gemini:

What do you think the average monthly search volume is for that prompt? Safe to say it’s not much more than one search per month.

Because there’s so much nuance and variability in LLM queries, the search volume of most prompts is naturally very low.

That’s why any AI visibility tools presenting traditional Google search volume as a proxy for LLM “search volume” are lying to you.

But there can be tens, hundreds, or even thousands of variants of a given prompt that all share the same search intent, even if the exact words of the prompts are different. I’m sure there are a number of people asking AI chatbots about PayPal’s payment protection policies every month, even if they don’t ask the question in the same way I did.

We just don’t have any good way today of knowing what those patterns are. (Profound offers “search volume” data based on a large corpus of real-world prompts, but to my knowledge it’s only for short-tail concepts like “real estate” and “project management tools.”)

In lieu of search volume data, the way to think about prioritizing your AI optimization efforts is based on relevance to your business.

More on that in the next sections.

Very simply: There’s no scalable way to find out exactly what commercial prompts your prospects are using.

But here’s the good news: It doesn’t matter.

You don’t need to know exactly what prompts your prospects are submitting to AI chatbots to know whether your brand is appearing in the right AI responses.

And that’s because of the way AIs (specifically, LLMs) work:

LLMs represent words and phrases as points in a massive, high-dimensional mathematical space. They determine the relationships between these concepts by calculating the distance and direction between these points, allowing them to simulate a deep understanding of how entities connect.

For example, your favorite LLM knows that “Hubspot” is closely related to entities like “Cambridge, Massachusetts,” “CRM,” and “sales team efficiency.”

So the core question you should be asking when it comes to AI search optimization is: “Do AIs associate my brand with the entities that matter to our business?”

And you don’t need to know the exact wording of your prospects’ prompts to answer this question. To an AI, two different prompts with the same search intent occupy very similar positions in mathematical space.

This is why I don’t worry too much about AI personalization. Because no amount of personalization will change the relationships between entities in the AI’s “brain.”

When we talk about entities relevant to a B2B brand, we’re talking about categories like:

The most efficient way to create a list of each of these entities for a given business is to have an AI take a first stab at it.

If you sell more than one type of product or service, each one likely has different target personas, features, value props, etc. So you need to go through this exercise for each category.

For most of these categories, I find it’s easy for an AI to come up with at least 10-20 entities that are at least somewhat relevant to the business at hand.

But once you start tracking everything that’s “somewhat relevant,” you’ll find that it creates a distracting noise in your data that makes it difficult to know where to focus. (Especially given the lack of “search volume” data.)

So prune the lists from step 1 down to the entities that are core to the messaging of your business. The ones where if someone asked about that thing in an AI prompt, you’d feel angsty about your brand not appearing in the response.

For example:

Those **are the truly relevant entities that should survive your ruthless prioritization.

Now that you have lists of truly relevant entities for each of the categories cited in step 1, you can combine them into natural-language prompts. Again, I’d speed up this process using AI.

Note that it’s easy to go overboard in this phase combining multiple entities in a prompt. E.g., I’m looking for an alternative to [competitor]. What’s the best [product category] software for [personas] doing [use case] who want to [goal]?

Remember, the goal here is to determine if AIs associate your brand with its relevant entities. You can do that with relatively straightforward prompts:

If you include too many entities in a prompt, you’re muddying the signal. In cases where your brand doesn’t appear in the response, you won’t actually know which entity needs to be shored up with more content. And you’re multiplying the prompts you have to track, making focus and prioritization more difficult.

The fastest way to get a list of 40-50 good prompts based on the process above is to use a free Contender account. You can get lists of prompts for any of the 2,100+ different B2B product and service categories in the database.

Alternatively, you can search for a leading competitor and see what categories they show up in. Then click on the category that’s most relevant to your business.

As you know, AI responses are probabilistic. Submit the same prompt two times and you’ll get two different responses.

For each prompt I track for a brand, I like to submit the prompt at least three times. This allows me to see if the brand tends to appear all of the time, some of the time, very little of the time, or never.

If that’s cost-prohibitive in your tracking tool, at least consider doing it for your most important (i.e., relevant) prompts.

Until next time,

Mike