AI Search Prompts and How Buyers Actually Ask
Key takeaways
|
For twenty years, we've studied how buyers search by studying keywords. Short phrases, monthly volumes, difficulty scores, carefully grouped into clusters. It worked, because that's genuinely how people searched, with a few words typed into a box.
Watch someone use ChatGPT and you'll see something very different. They write a paragraph. They explain their situation, their company size, their constraints, what they've already tried and what they're worried about. They ask a follow-up, then another. It reads less like a search and more like a briefing to a trusted advisor.
That changes how you research demand, how you write content and how you measure visibility. Your keyword list still matters. But it no longer captures the whole picture of how buyers look for help, and in AI search, it may not even capture the most important part.
What are AI search prompts?
AI search prompts are the questions, requests and instructions users type into AI assistants such as ChatGPT, Claude, Gemini, Perplexity or Google AI Mode. Compared with traditional search queries, prompts are typically longer and more conversational, include personal context and constraints, and are often broken by the AI into several underlying searches. |
Prompts aren't queries
The clearest large-scale data comes from Semrush's analysis of 80 million clickstream records. The average ChatGPT prompt was 23 words long. When users switched on search, prompts dropped to an average of 4.2 words, much closer to classic Google behavior. That gap tells you a lot. When people treat ChatGPT as a conversation, they give it context. When they treat it as a search engine, they revert to keywords.
The more important difference is what happens after the prompt is written. A search query is matched and ranked. A prompt is interpreted, and often broken into several hidden searches. Research on ChatGPT's behavior found it searched the web on roughly a third of prompts and ran about 2.17 searches per prompt when it did, with those internal queries averaging around five and a half words. And Peec's analysis of more than 20 million ChatGPT fan-out queries found their average length roughly doubled between October 2025 and January 2026, from about six words to about twelve. The system is getting more specific about what it goes looking for.
Why keyword research alone now misses demand
Keyword tools measure what people type into search engines. They can't see what people type into ChatGPT, and they can't see the searches ChatGPT runs on the user's behalf. That second gap turns out to be enormous.
Kevin Indig's research found that a third of the pages ChatGPT cited came from those behind-the-scenes fan-out searches, and 95% of the fan-out queries had no measurable search volume at all. In other words, some of the searches that decide your AI visibility simply don't exist in any keyword database. You can't find them by looking at volume, because there isn't any. The only way to cover them is to understand the full set of questions a buyer brings to a topic and answer those questions well, which is the idea behind query fan-out.
Six kinds of B2B prompt
When I build a prompt map for a client, I sort prompts into six types that loosely follow the buying journey, because each type is won by a different kind of content.
Problem prompts come first, such as our organic traffic dropped thirty percent but rankings are stable, what's going on? They're won by honest diagnostic explainers. Category prompts follow, such as what is generative engine optimization and do B2B SaaS companies need it? Those are won by clear definitions and frameworks. Then come shortlist prompts, such as what are the best B2B content agencies for Series B SaaS companies, which are won largely by third-party lists, reviews and clear positioning.
Comparison prompts, like agency versus in-house content team for a fifty-person SaaS company, are won by honest comparison pages with real criteria. Validation prompts, is this vendor any good and what do customers say, are won by reviews, case studies and community discussion. And implementation prompts, like how do I set up an AI traffic channel in Google Analytics, are won by step-by-step guides, templates and tools.
Most B2B content programs over-invest in category prompts, because definitions are easy to write and used to rank well. They tend to ignore shortlist and validation prompts, which is exactly where vendor decisions form. Those two types are also where your own site matters least and third-party sources matter most, a point developed in brand mentions in AI answers.
Building a prompt map
The richest source of real buyer language is already inside your company. Sales call notes, discovery questions, support tickets and objections are full of the exact words buyers use, and almost nobody mines them systematically. Start there. Then go where buyers talk to each other. Reddit, LinkedIn comments, community Slack groups and review sites show how people describe the problem in their own words, without a vendor in the room.
People Also Ask boxes and autocomplete are imperfect but useful for spotting question patterns. You can also ask an AI directly, with something like, I'm a VP of marketing at a 200-person SaaS company trying to improve our pipeline from content, what questions would I ask an AI assistant? Then edit the results hard, because AI-suggested prompts lean generic.
Add context variants for the personas that matter to you, because the same question asked by a CMO at a large company and a founder at a small startup can produce different answers. Then sort everything by type and trim it to somewhere between thirty and a hundred prompts that genuinely reflect demand. Throughout, write prompts the way buyers would, not the way your product marketing team would. If buyers say "content agency" and you say "demand orchestration partner," track the buyer's words.
Putting the map to work
A good prompt map earns its keep in three ways. For content planning, each prompt type points toward a format, so gaps in the map are gaps in your library, and prompts where you're absent but competitors are named become your priority list. For measurement, running the same prompt set across several assistants on a schedule, several times each, gives you your share of model. Because answers vary between runs, the point is to measure rates rather than react to single answers, as explained in AI non-determinism.
The third use is the one teams forget, which is messaging. Prompts show you how buyers frame the problem. If your positioning doesn't match the language in your prompt map, that's worth fixing before you publish another page. Sometimes the most valuable finding from a prompt map isn't a content gap at all. It's a vocabulary gap.
Common mistakes
The most common mistake is putting your own brand name into visibility prompts. Asking whether your company is any good measures reputation, not discovery, and both deserve tracking, separately. Another is only tracking the prompts you already win. Pick the ones you'd want to win. Changing the prompt set every month destroys the trend, so add new prompts by all means but keep a stable core. Testing from a personalized account skews results, the issue covered in AI memory and personalization. And ignoring follow-ups misses how real conversations actually unfold. For your most important prompts, test the next question too.
AI search prompts in the Engine Optimization Matrix™
Your prompt map is the input to the Citation column of the Engine Optimization Matrix™, especially the LLMO row, which asks whether ChatGPT, Claude or Perplexity name you specifically when prompted. Without a defined prompt set, that cell can't be measured. With one, every other cell has a clear target.
For related reading, see AI search visibility tracking, ChatGPT prompts for B2B marketers, anatomy of an AI response and what is RAG.
A model can't recommend a brand it doesn't know exists. The Demand Doctrine™ for LLMO covers how to get your brand into what the models remember, not just what they retrieve. Out November 18. Want all four engines? Mastering AI Visibility.
Frequently asked questions
What are AI search prompts?
They're the questions and instructions people type into AI assistants, which tend to be longer and more conversational than search queries.
How long are ChatGPT prompts?
Semrush's analysis of 80 million clickstream records found the average ChatGPT prompt was 23 words, and 4.2 words when search was enabled.
How are prompts different from keywords?
Prompts are full sentences with context and constraints, and AI assistants often turn one prompt into several underlying searches.
Can keyword tools show AI prompt volume?
Not directly. Keyword tools measure search engine queries. Many of the searches AI assistants run on a user's behalf have no measurable volume.
What are the main types of B2B AI prompts?
Problem, Category, Shortlist, Comparison, Validation and Implementation prompts, which roughly follow the buying journey.
How do I find the prompts my buyers use?
Mine sales calls, support tickets and discovery questions, read communities where buyers talk and use AI to suggest variations, then edit for real buyer language.
How many prompts should I track?
Thirty to 100 prompts, run several times each across several AI assistants, gives a useful picture.
Should I include my brand name in tracked prompts?
Track branded prompts separately. Visibility prompts should not include your brand name, since the goal is to see whether AI suggests you unprompted.
About the author. Harold Bell is the founder of MQL Magnet and has spent about 16 years in enterprise technology marketing, working with brands like Cisco, Rubrik and Nutanix in different capacities. He's a member of the Forbes Communications Council, the creator of the Engine Optimization Matrix™ and the author of Partner Over Product. |
