How do I know which prompts trigger AI recommendations in my category?

Start with this answer: the prompts that trigger AI recommendations in your category are rarely the literal question your customer types. AI engines rewrite that question into several sub-queries, then retrieve sources for each one. Your job is to find the underlying buyer intent and test a working sample of prompts across the major chat engines.

If you spent the last decade pulling keyword lists from Google Keyword Planner and Ahrefs, this will feel upside down. The customer types one thing. The engine searches three or four others. The recommendation that comes back is built from sources the customer never asked for.

I know how that lands. It looks like the rules just changed under your feet.

They did. But the new rules are findable, and you can do the work yourself in an afternoon.

Why the literal prompt is the wrong target

Old keyword work was a guessing game about exact phrases. AI search is a guessing game about intent shapes. One thing drives that shift. It is now built into every major engine.

Google calls it query fan-out. When you ask AI Overviews a question, the system issues several related searches. It pulls answers across subtopics and data sources. Then it folds them into one reply. ChatGPT search works the same way. Per OpenAI's documentation, the system rewrites your prompt into targeted queries. It retrieves results, then sends more specific queries based on what came back.

Academic work backs this up. The Princeton research on Generative Engine Optimization was published in 2024. Their headline finding: the right content patterns can lift visibility in AI engines by up to 40 percent. The less-quoted finding is that the queries those engines actually retrieve against are reformulations of the user prompt, not the prompt itself.

So if you build your blog, your service page, or your bio around the literal prompt your customer would type, you are aiming at a query the engine never sends.

This is the part that stings. Most of the keyword tools you have been using were aimed at the wrong target.

The five shapes that prompts in your category will fall into

When you sit with how customers actually talk to AI chatbots, the prompts cluster into a small number of recurring shapes. These five cover most of what you will see in a small business category.

1. Comparison prompts. "What is the difference between X and Y in [city]?" The customer is choosing between two known options.
2. Qualified location prompts. "Best [service] in [suburb] for [specific need]." The customer wants a recommendation, narrowed by place and qualifier.
3. Problem-statement prompts. "I have [problem], who should I see?" The customer leads with a symptom, not a service category.
4. Criteria-stacked prompts. "I need a [service] that does [thing one] and also [thing two] and is [budget tier]." The customer has stacked filters.
5. "Best of" list prompts. "Top five [service] for [audience] in 2026." The customer wants a curated set, not a single answer.

Each shape pulls a different mix of sources. Comparison prompts pull review sites and trade press. Problem-statement prompts pull editorial pages and forum posts. "Best of" list prompts pull listicles, often from outside publishers. The shape tells you what kind of source the engine reaches for. That tells you where you need to be cited.

The prompt is the question. The shape is the strategy.

How AI engines actually pick what to recommend

The shape of the prompt drives source selection. A 2024 study on search engines post-ChatGPT shows a clear pattern. AI search leans toward earned media (third-party trusted sources) over brand-owned content and social posts. That lean is much stronger than what Google's classic web index shows.

Two takeaways for a small business.

Your own site is rarely the source the engine reaches for. It is one input. You get cited by being named on someone else's page in the right context.

Intent matters more than authority alone. Research on user intent recognition with large language models shows the engine infers what kind of answer the user wants. Then it picks sources that match. A "where can I get this fixed" prompt returns different sources from a "should I get this fixed" prompt. Same topic, different sources.

Your literal keyword and your customer's intent rarely line up.

A quick scene. A bookkeeper in a regional town wants to know which prompts trigger her. She types her own service into ChatGPT. Nothing useful comes back. She tries "how do I find a good bookkeeper for a small trades business." Now an answer shows up, with three businesses listed. None of them are her. Two of them are local rivals. The shape of the prompt that surfaces her rivals is the shape she now needs to be cited inside.

The prompt she would have picked for keyword research was not the one that pulled the pick.

A simple method to find your category's trigger prompts

You do not need a paid tool. You need a notebook, a couple of hours, and the will to ask people how they would phrase the question.

Here is a method that works for most small business categories.

Step one. Interview five customers in plain language. Ask each person: "If you had to find a [your service] using ChatGPT or Gemini today, what would you type into it?" Write down their exact words. You are not asking what they would Google. You are asking what they would say to a chatbot. That is a different prompt.

Step two. Add three to five competitor-style prompts. For each of your top competitors, write the prompt that would be most likely to surface them by category. This covers the comparison shape and the qualified location shape.

Step three. Run the prompts on the major AI chat engines. ChatGPT, Claude, Gemini, Perplexity, Grok, and DeepSeek are the surfaces most people use. Google AI Overviews and Microsoft Copilot are worth a check too. Run each prompt at least twice on each engine. Answers vary.

Step four. Score each result on five lenses. For each prompt and engine, note five things. Did your business appear (visibility). How was it described (tone). Was the description correct (factuality). Did the answer give a customer something to do (next step). Did the answer hold up across engines (match).

Step five. Look for shapes, not perfect prompts. The point is to spot the prompt shape that triggers picks in your category. It is not to find one magic phrase.

Around 10 well-picked prompts on the major chat engines is enough to see real patterns for a small business. More is fine. Fewer leaves you guessing.

If I were you, I would start with the customer interviews and treat the rest as scaffolding around what real people say.

What to do with the prompt patterns once you have them

This is where most people lose the thread. They run the test. They see the gap. Then they freeze. The obvious answer is "write more content" and that feels endless.

Pick one prompt shape. Just one.

If your category surfaces best on comparison prompts, look at the outside comparison pages where your rivals are named. Get yourself onto the right ones. Ask for an add or write a guest piece.

If your category surfaces best on qualified location prompts, the work is local trust signals. Schema markup, the structured data patterns Google publishes, business name match across directories, and a clean Google Business Profile.

If your category surfaces best on problem-statement prompts, the work is editorial. Plain content that names the problem, then names you as the kind of business that solves it. Forum threads count. Podcasts count.

You do not have to do all five shapes. You have to do the one your customers actually use.

Common traps to avoid

Three traps show up almost every week.

The first is testing only on ChatGPT and thinking the rest will match. They do not. ChatGPT's source mix is its own. Perplexity weighs citation count. Gemini leans on Google's index. Claude is picky. Grok and DeepSeek are different again. A pattern that holds on ChatGPT can be invisible elsewhere.

The second is running each prompt once. AI answers are not the same every time. The same prompt at 9am and 11am can return different businesses. Sample each prompt at least twice. Only what shows up across runs is real.

The third is over-reading a single recommendation. One result in a sample of 20 is noise. Eight results in 20 is a pattern.

A few tools do this kind of sampling at scale. HubSpot's AEO Grader, Otterly, SE Ranking, and Get Recommended are some. Pick by which engines and which signals you want covered. The method above works by hand for the cost of an afternoon.

A note on how often this changes

The major AI chat engines update their picks faster than Google ever did. Recent Pew Research data on how Americans use AI shows fast growth in use, mostly in younger adults. Prompt patterns shift with use.

A quarterly re-test is enough for most categories. Re-test sooner if you launch a new service, change your name, or notice a shift in the kind of work coming in. Treat the prompt list as living. Not as something you build once.

Run the work, see what shows, fix one thing

If you have read this far, you have the method. The work that pays is the work you do tomorrow morning.

Five customer chats. A short list of rival-style prompts. A run on the major chat engines. A score on the five lenses. Then pick one prompt shape and do the next small thing inside it.

The reader who closes this tab and writes the first prompt on a sticky note has moved further than the reader who saves the post for later.