Prompt research: what clients actually ask AI

In the foundational article of the series, we explained why it is now crucial for a brand to be a source cited by AI, and in the tools article, we discussed how to measure this visibility. But the main practical question remains: where do you actually get the list of queries whose visibility should be tracked?

The answer lies within a separate discipline — prompt research. This is the analysis of exactly how your clients formulate their tasks in ChatGPT, Gemini, or Perplexity. And it is completely different from traditional keyword research.

Why a keyword and a prompt are different things

Classic SEO operated with short keywords. A person would type "buy CRM Kyiv" into Google and get ten links. In a conversation with AI, the same client formulates it differently: "Which CRMs are suitable for a startup of up to 10 people in Kyiv that works with European clients?"

The difference is fundamental. A prompt carries context: team size, industry, constraints, and market language. It is not just two words, but an entire task with specific conditions.

This shift is confirmed by numbers. According to Profound, which analyzed over 50 million real queries, the share of purely navigational searches (when a person looks for a specific website) dropped from 32% in classic Google search to about 2% in ChatGPT. Instead, a whole new category has emerged — "generative intent," which accounts for about 37% of queries: the user is not looking for a page but is asking the AI to gather, compare, or recommend something.

The scale of the difference was also demonstrated by Google itself: at I/O 2026, the company reported that queries in AI Mode are, on average, three times longer than classic ones, and every sixth query is multimodal (involving a photo or a file). Furthermore, Semrush, based on over a billion clickstream rows, revealed something even more telling: from 65% to 85% of prompts cannot be matched with any traditional keyword — meaning that a significant portion of AI queries is simply invisible to a standard semantic core.

Four levels of the prompt funnel

Prompts can be conveniently broken down by stages of the sales funnel:

• Top of the funnel (TOFU) — awareness. "What is end-to-end analytics?", "why does a business need a CRM". The person is not choosing yet — they are figuring it out.

• Middle of the funnel (MOFU) — comparison. "CRM A vs CRM B", "alternatives to [well-known service]", "best tool for an online store". This is where the choice begins.

• Bottom of the funnel (BOFU) — decision. "Reviews of a specific brand", "how much does the service cost", "is it worth the money". The client is almost ready to pay.

• Post-purchase. "How to set up", "how to integrate with another service". This affects retention and repeat sales.

Key takeaway for executives: classic SEO chased query volume at the top of the funnel. However, the money in AI search is concentrated lower down — on comparisons and decisions, where the purchase intent is strongest.

Commercial intent signals

Among all prompts, there is a category where being cited is most valuable — queries with a clear intent to buy. They are easy to recognize by their patterns:

• "best [tool] for [role or niche]"

• "alternatives to [service name]"

• "[product A] vs [product B]"

• "is [product] worth the money"

• "pricing overview for [product]"

A person using such a prompt is no longer exploring the market — they are narrowing down their choices to two or three options. If the AI names a competitor instead of you in its response, the deal is practically lost before the first contact.

At the same time, purely product-specific queries are few: according to Profound's analysis, only about 6% of prompts consistently trigger product cards in ChatGPT. However, this category is "sticky" — if a query worked today, there is about an 83% probability it will work tomorrow as well. In other words, commercial prompts are few, but every won position lasts long.

Where to get prompt ideas

The biggest mistake is inventing queries out of thin air: real formulations already exist, you just need to collect them. Here are five reliable sources:

1 Sales department notes. Call and chat records (Gong, Zendesk, Intercom) are literally clients' questions in their own words.

2 Top queries from Google Search Console and GA4, reformulated into a conversational style.

3 The "People Also Ask" block in Google search results — a ready-made list of related questions.

4 Reviews on G2, Capterra, Trustpilot — the language clients use to describe their pain points and selection criteria.

5 Industry communities and Reddit threads — these are exactly what AI cites most frequently.

And why this is relevant right now: AI Mode has been working in Ukrainian since autumn 2025, and free personalization (Personal Intelligence) supports nearly a hundred languages. This means Ukrainian-language prompts are no longer a theory, but a real channel. Ukrainian specificity: a significant portion of our sales happens through messengers. The archive of requests in Telegram, support chats, and call recordings is an underestimated but highly accurate source of real phrasing. Monitoring tools with their own query databases also provide great ideas.

How to turn prompts into a content plan

The collected list is not a strategy yet. Next, it must be turned into an action plan:

1 Group prompts by funnel stages — it immediately shows where you already have content and where the gaps are.

2 For every BOFU prompt, create a separate page with a comprehensive answer in formats that AI "extracts" best: a direct 40–60 word answer, an FAQ block, or a table.

3 Create dedicated pages for comparison queries like "your brand vs alternatives" — with honest criteria and numbers.

And a warning: the set of prompts changes rapidly — the prompt map should be reviewed at least quarterly.

What's next

With a query map, metrics, and tools in place, it is logical to assemble everything into a step-by-step plan. In the next article of the series, we have prepared a practical checklist: "AI Visibility in 180 Days" — what to do in the first 30, 90, and 180 days.

Ultimately, a prompt map is the semantic core of the new era: not a set of keywords for search rankings, but a list of real intents by which AI decides which brand to name. Building such a core manually and keeping it relevant is meticulous work that makes sense to delegate: this is exactly what the semantic core collection service from Mas Agency is built upon, fully adapted to prompt intents for AI.

Sources used