Chapter 6: From Keywords to Prompts

Last updated: November 1, 2026.

Imagine starting SEO work for a filling machine manufacturer. A few years ago the first step was obvious: open a keyword tool, pull terms like “filling machine,” “liquid filling machine” and “filling machine prices” with their search volumes, group them, and plan a page for each group.

Today that manufacturer’s customers don’t type keywords into an AI assistant. They describe their situation, list their constraints and ask for a recommendation. “Which keywords should we show up for?” is still a fair question, but now it comes with a second one: “Which questions should we show up for?” I call the work of building that list of questions Prompt Research. The goal is a list of 30–50 commercial prompts that we’ll measure in Chapter 7.

1. Topic vs. situation

Remember the purchasing manager from Chapter 1. In classic search, they typed “liquid filling machine” into Google. Today the same person might type this into ChatGPT: “We’re going to fill 3,000 bottles of liquid soap a day. The product foams, and our budget is around €25,000. Which manufacturers with service in Turkey should I look at?”

Both come from the same need. But “liquid filling machine” is the name of a product category, while the long question describes a business: its capacity, its product, its budget, its location and what it expects from service. A keyword describes a topic; a prompt describes a situation.

Prompts are long. Google says searches in AI Mode are two to three times longer than classic searches; we covered that in Chapter 1. Semrush’s US clickstream data shows the same trend: the average length of ChatGPT questions that triggered a search went from 4.7 words in early 2025 to 8.7 words in early 2026 (Semrush).

The odds of two people typing a sentence that long word for word are tiny. So the keyword question “how many times a month is this phrase searched?” has no meaningful answer for a prompt.

A prompt doesn’t stand alone, either. After the first answer, the user follows up: “Which of these has the wider service network?” The decision often happens on the third or fourth question, not the first.

Keywords don’t disappear. They change hands. As we saw in Chapter 3, the model splits a long question into several short searches through query fan-out, and now it’s the model, not the user, writing those short searches. If the model searches for “how to fill foaming liquids” or “filling machine manufacturers,” you still need to show up for those searches. Prompt research doesn’t replace keyword research. It adds a layer on top of it.

2. What people ask AI

The most thorough data on this comes from “How People Use ChatGPT,” a study published in September 2025 by researchers from OpenAI and Harvard. It looks at a representative sample of about 1.1 million conversations from May 2024 to June 2025 (Search Engine Journal):

• About 49% of messages are asking for something (Asking), 40% are getting something done (Doing) and 11% express feelings or opinions (Expressing).
• The share of conversations about seeking information rose from 14% to 24% in one year.
• Practical guidance is the biggest topic, at about 29%.
• Non-work use went from 53% in June 2024 to 73% in June 2025.

The authors conclude that ChatGPT delivers much of its economic value through decision support (NBER).

Most of what people ask AI isn’t commercial. They have it write text, fix code, explain a subject. But a fair share of questions are asked to make a decision. Those are the questions prompt research is after: the ones where decisions about your business get made.

3. What a prompt list is for

The first misreading that comes to mind is opening a page for every prompt. In Chapter 3 we saw Google’s warning: producing separate content for query variations can count as scaled content abuse. The same guide also says you don’t need to worry about not having enough long-tail keywords, or about missing some variation someone might search for (Google Search Central). Two separate pages for “filling machine for 3,000 bottles of foaming product a day” and “filling machine for 5,000 bottles of shampoo a day” give the user nothing.

Turning a keyword list into questions isn’t prompt research either. Rewriting “filling machine prices” as “How much do filling machines cost?” adds nothing new to the list. You have to think about how a real user would ask, and with what details. And the list isn’t something you write once and forget; it changes as your products, your competitors and the way people ask questions change.

The list does three jobs. It decides which questions you’ll measure your AI visibility on; every metric in Chapter 7 runs on it. It shows which of the user’s constraints your site answers and which it doesn’t. And it shows you the competition: which brands get recommended for these questions, and which sources get cited.

Content decisions come out of this list, but not every row means a new page. What you have is less a content calendar than a measurement and diagnosis list.

4. Constraints and where your site answers them

What makes a prompt valuable is the constraints inside it. In the Chapter 1 example the user gave five: daily capacity, a product property, budget, country and service needs. AI builds its answer with all of them and leaves out brands that don’t fit.

For a B2B machinery manufacturer, constraints can be grouped roughly like this:

The practical part is the last column. Every constraint maps to a piece of information that has to be stated clearly on your site. If you never write your capacity as a number, it gets harder to be recommended to a user who says “3,000 bottles a day.” If you don’t describe your service regions, it gets harder to make the list of a user who says “with service in Turkey.”

In classic search, a user could visit your site, pick up the phone and ask about your service area. If AI can’t find that information, it can simply skip you. AI doesn’t give you the benefit of the doubt on a constraint it can’t verify.

5. Where to find prompts

For keywords, you could open a volume tool and pull a list. There’s no ready-made source like that for prompts, so you have to collect them from several places.

Start with your sales team. Ask them: “What do customers ask on the first call? What objections do you hear at the quote stage? Why do they pick us, and why do they pick a competitor?” Most of those questions are the same ones the customer asks an AI assistant that same day.

The free-text fields in quote request forms are where users describe their constraints in their own words. Going through the last few years of quote requests is the fastest way to build a list of real constraints. Questions from WhatsApp, live chat and phone logs help too; after-sales questions about spare parts, service and maintenance are a good source for brand verification prompts.

You can also pull long queries out of Search Console. In the Queries report, a regex filter like ^(\S+\s+){7,}\S+$ leaves only queries of eight words or more. That’s real data showing that people are typing longer, more constrained questions into Google as well.

Other places worth checking:

• Site search: what people type into your search box shows what they couldn’t find on your site.
• Google’s “People also ask” box and the follow-up questions suggested under answers in AI Mode.
• Forums, Reddit and complaint sites: you see how people phrase questions to each other and what gets asked about your brand.
• Competitors’ comparison pages and FAQ sections.

You can ask AI itself, too: “What questions would a cosmetics manufacturer’s purchasing manager ask when choosing a filling machine?” That’s not a bad way to widen the list. But AI produces plausible questions, not the questions real users ask. Don’t add them to the list until they’ve been checked against your sales team and real data.

Some AI visibility tools also offer prompt suggestions by industry or estimated prompt volumes. I’ll come back to these in section 12. The short version: they point you in a direction, but don’t read them like Search Console data.

6. Five people asking about the same machine

Different people ask about the same product in different ways. The people asking about a filling machine fall roughly into these groups:

• The purchasing manager asks about price, lead time, payment terms, warranty and service: “Which filling machine manufacturers have service in Turkey and can deliver within two months?”
• The production or maintenance manager asks technical questions about foaming, viscosity, fill accuracy, cleaning and changeover time: “What filling system should we use to prevent overflow with foaming liquid soap?”
• The entrepreneur starting a new business asks about budget, starting capacity, semi-automatic vs. fully automatic and installation support: “I’m starting a small cosmetics brand. What should I look for when buying my first filling machine?”
• The buyer abroad asks the same questions in English or their own language, with certification, shipping and remote service coming to the front: “Which Turkish filling machine manufacturers offer CE certified machines with remote support?”
• The existing customer asks about spare parts, service, maintenance and upgrades: “Where can I find spare parts for company X’s filling machine?”

I call these persona clusters. Each one asks at a different stage of the decision, with different constraints and different words. Each also needs something different from your site: the purchasing manager’s question leads to the service and delivery page, the production manager’s to technical content, the entrepreneur’s to guide-style content.

Build the clusters from sales and customer data, not imagination. The answer to “Who actually buys from us?” is in your CRM and in your sales team’s heads. Keep the number small, too. Three to five clusters are enough for most businesses; more just inflates the measurement work.

7. Prompt types

Prompts also need to be sorted by type, because each type tells you something different when you measure:

• Category discovery: “Which companies in Turkey make filling machines for foaming products?” No brand is named, and this is where the shortlist gets built. Visibility is worth the most here.
• Problem and solution: “How do we stop foam overflow when filling liquid soap?” The user isn’t shopping yet; they’re trying to solve a problem. This is where you see whether your guides get cited.
• Comparison: “Piston filling machine or submerged-nozzle filling machine?” or “X or Y?” This measures your technical content and how you stack up against competitors.
• Price and terms: “How much does a semi-automatic filling machine cost?” Shows whether your pricing information is passed along correctly.
• Brand verification: “Is X Machinery reliable?”, “What’s X Machinery’s service network like?” The user already has you on the shortlist and is doing a final check. This is where you find out whether your brand is described accurately.
• Alternatives: “Filling machine manufacturers similar to X Machinery.” A question asked with a competitor’s name, where you’d want to be on the list too.

Most of the list should be prompts that don’t name a brand, because that’s where new customers discover you. But include a few prompts with your brand name in them. These measure accuracy rather than visibility: When AI is asked about your brand, does it say the right things?

8. Starting with 30–50 commercial prompts

It’s tempting to make the list big from day one, with hundreds or even thousands of prompts. I think that’s a mistake, for two reasons.

The cost of measuring multiplies. Because the same question gets a different answer each time (Chapter 3), you need to run every prompt several times and measure each platform separately (Chapter 4). Run 40 prompts on 4 platforms, 5 times each, and you have 800 answers. At 500 prompts, that becomes 10,000. Reading 10,000 answers every month, checking them for accuracy and diagnosing problems isn’t realistic.

Non-commercial prompts also water down the measurement. Showing up for “What is a filling machine?” is fine, but it does little for sales. When the list fills up with questions like that, the overall rates move up and down while the changes on questions close to a sale go unnoticed.

So build your first list with 30–50 commercial prompts. By commercial I mean questions close to a sale, where the user is making a decision. Once the list has been measured for a month and you’ve seen what works, you can expand it.

Here’s a sample split for a 40-prompt starting list:

This split isn’t a rule. On an e-commerce site, price and terms questions carry more weight; for a well-known brand, brand verification questions matter more. Make sure each type has enough prompts and choose the split on purpose.

Before a prompt goes on the list, look at three things. How close to buying is the person asking it? How much would being recommended for it be worth to you; is it about your main product line or a minor side product? Does your site answer it? If the answer to the last one is no, don’t drop the prompt. Note it as a content gap.

9. Writing prompts the way real users do

Small details in the prompt text decide how good your measurement is.

Write in the user’s language: Turkish if you sell in Turkey, the target market’s language if you export. As we saw in Chapter 4, language changes which searches the model runs and which sources it reads.

Add constraints, but don’t overdo it. A real user usually gives two to five; a prompt with ten isn’t realistic.

Don’t lead the question. “Is X Machinery the best filling machine manufacturer in Turkey?” isn’t a measurement question; it’s a request for confirmation. Unbranded prompts should never contain a brand name.

Some users write the same intent short (“foaming product filling machine manufacturers”) and some write it long, with constraints. Putting both versions on the list for your priority intents shows how your visibility changes with question length.

For your most valuable prompts, prepare a two- or three-step conversation: the first question, then “Which of these has the wider service network in Turkey?”, then “Compare the first two on price and delivery time.” As I said in Chapter 1, users narrow the list with every question. If you’re in the first answer but drop out on the second question, you’re not on the list when the decision is made.

Finally, lock the wording. Once you’ve written a prompt, use it word for word every month. If you ask with a different sentence each month, you can’t tell whether a change came from you or from the question.

10. The searches behind a prompt

The model splits the user’s long question into several short searches and picks its sources from those results (Chapter 3). It doesn’t show site owners which searches it ran, but since we know how it works, we can make an educated guess.

For the 10–15 most valuable prompts on your list, write out by hand the searches the model might run in the background. For the purchasing manager’s prompt, that list might look like this:

• foaming liquid filling machine
• liquid soap filling machine capacity
• submerged nozzle filling machine
• filling machine prices
• filling machine manufacturers Turkey
• filling machine service

This is a guess; the real searches may be different. But these short searches can be measured in classic keyword tools and in Search Console. For each one, check: Do we rank near the top on Google? Which pages show up instead of ours? Do we have a page that answers this search?

The prompt describes the user’s situation and decision. The short searches behind it decide which pages get read on the way to that decision. If you want to show up for a prompt, you need to show up for the searches behind it. The Ahrefs data from Chapter 1 is what that looks like in practice: in its March 2026 study, only 37.9% of pages cited in AI Overviews ranked in the top 10 for the same query, and Ahrefs attributes this to sources also being picked from the sub-queries in query fan-out.

11. Keeping the list in a table

You don’t need a complicated system for a prompt list. A simple table with these columns is enough:

The last column is the most useful one. Asking “Which page on our site answers this question?” for every prompt brings content gaps to the surface on its own. If no page answers it, AI has nothing to base a recommendation of you on for that question.

This table is just the list itself. We’ll keep the measurement results, meaning whether you were mentioned, recommended and linked for each prompt, in a separate table, which we build in Chapter 7.

12. Prompt volume is unknown

In a keyword list, every keyword had a search volume next to it. In a prompt list, that column is very hard to fill. AI platforms don’t publish which questions users ask or how often. As we saw in Chapter 4, ChatGPT, Gemini, Perplexity and Claude don’t give site owners a query report, and Google’s AI report doesn’t include query data either.

Some tools show estimated volumes for prompts. These estimates are usually modeled from user panels or clickstream data; the Semrush data in this chapter also comes from a US clickstream panel. They’re useful for seeing general trends, but don’t read the volume for a single prompt like an impression count in Search Console. For Turkish prompts, the samples behind these estimates may be much smaller.

Without volume, you can set priorities with three indirect signals. If your sales team hears a question every week, people are asking AI too. The Google search volume for the topic behind a prompt gives you a sense of how much interest there is. And then there’s deal value: the value of the sale behind a prompt can matter more than how often it’s asked.

In B2B, that last one often decides it. One of our B2B machinery clients gets 30–40 visits a day, and a single machine sells for about €20,000. For a business like that, being recommended for a prompt asked a few times a month can be worth far more than being mentioned in a generic question asked thousands of times a month. Pick prompts by how close they are to a sale and what the deal is worth, not by volume.

13. Maintaining the list

A prompt list isn’t something you build once and shelve, but you don’t rewrite it every month either. Most of the list, the core list, should stay the same for at least six months. The value of measurement comes from tracking the same prompts the same way over months.

Review it every quarter: new questions your sales team is hearing, new products, new competitors, products you no longer sell. Add or remove a few prompts if needed, and log every change with a date. When your overall visibility rate moves, that log is the only way to tell whether the change came from the list or from your visibility.

Keep an eye on the platforms, too. When a platform launches a new feature or model, the same prompts can produce different results; don’t mistake that for a change on your own site. We’ll come back to this cycle in Chapter 17, as part of a system that runs continuously.

What you should have at the end

By the end of prompt research, you should have a plain table of 30–50 rows. Each row holds a question a real user would ask, who asks it, the constraints it carries and the page on your site that answers it. The rows where that last column is empty are your first to-do list.

The next step is measuring the list. Chapter 7 covers that: five metrics instead of a single score, running each prompt several times and measuring each platform separately.

Checklist

Sources

• Semrush: ChatGPT search insights (April 7, 2026)
• Search Engine Journal: ChatGPT study, 1 in 4 conversations now seek information (September 15, 2025)
• NBER: How People Use ChatGPT, Working Paper 34255 (Chatterji et al., September 2025)
• Google Search Central: Optimizing your website for generative AI features on Google Search (last updated July 10, 2026)

Previous: Chapter 5, GEO, AEO, LLMO: What Has Really Changed? · Table of contents · Next: Chapter 7, How to Measure AI Visibility