Category entry points in AI search, and how to choose the prompts your brand should manage

Category entry points (CEPs) are the moments when people start thinking about a category, and in AI search they are where the prompts come from. To choose which prompts your brand should manage, start from those moments and work backwards to the questions people will put to an AI when they are thinking of buying. Seyong Park, founder and CEO of the Korean search data company Ascent AI, put it this way on episode 23.

So I kept arguing that a CEP is a prompt and a prompt is a CEP, which is why it is so important to reverse-engineer prompts from the CEP and reproduce them.

Seyong Park

Park also answered who should start first, and he answered with conditions: complex products whose specs and features consumers do not fully understand, products with a purchase journey of at least a week and up to two or three months, and products bought together with other products. This post covers the part of episode 23 about prompts (12:40 to 20:35 and 27:32 to 32:30). The episode is in Korean. This is the English version of a Korean post, and every quote below is translated from Korean. Each section links to its part of the video.

✦A CEP is a prompt

Watch this part of the episode 27:32

The term CEP comes up many times in Park's Korean book 브랜드, AI의 언어로 말하라 (Brand, Speak the Language of AI). I asked him what it means.

CEP is just short for category entry point. And category entry point is not my term. It is something a professor called Byron Sharp has been talking about for quite a long time.

This is how he summed it up.

It means the moment you think you need a particular thing or a product from a particular brand.

It is called a category entry point, he explained, because when people solve a problem they think of the whole category the brand belongs to before they think of a single brand. He gave two examples.

If you are thirsty, you think of the drinks category: "I need to drink something."

The moment you think, the family has grown, I need to trade in my small car for a family car, you think of the category called SUV.

Park said the idea looks a bit like what marketers have long called the scene in which a product is used. He stressed CEPs in his book anyway, because they connect directly to the first job in GEO.

GEO ultimately starts with deciding which prompts to manage, and in the end people also write their prompts at a particular CEP.

✦Which prompts to manage, and which to leave out

Watch this part of the episode 27:32

It also means you do not have to manage every prompt people write. Park drew the line like this.

You do not make prompts used for research into managed prompts. In the end, the prompts we usually choose to manage are the ones people write when they are considering buying our product or our brand, and those managed prompts are stuck right to CEPs.

So the CEP comes first. Write down the moments when people think of your category, then work backwards to the questions they will ask an AI at that moment. That is the reverse engineering Park is talking about.

✦Will there be too many prompts to manage?

Watch this part of the episode 30:14

In SEO, categories have been set with keyword maps. I asked whether moving from keywords to prompts means the list of things to manage will grow without end.

If you think about the number of prompts, it can feel like a lot. But I do not think that things people never asked before suddenly appeared just because AI search came along.

It is not that intents that never existed before have appeared because AI arrived. The things people always asked are simply coming out in the form of prompts, so I think it basically has to be the same.

He explained it through the way a product spreads through a market.

A brand-new thing that nobody knew about basically starts from "What is that?"

As the product reaches the market, people use it, come to prefer certain brands and compare them, and what they ask becomes more varied.

If a product already has a wide range of prompts, it means a great many people know it, use it or compare it, and that is why there are many prompts.

The number of search queries you handled in the search engine era and the number of prompts cannot be very different. But because the form has changed, the know-how of how to put them together is really needed now. That is what I would say.

By Park's account, the number of prompts worth managing is not far from the range of search queries the product already received in the search engine era. What is new is less the count than the form the questions take.

✦Are people really asking longer questions?

Watch this part of the episode 12:40

Park's book has an example of someone asking an AI to find an air purifier with affordable filters for a small studio flat shared with two cats. I asked Park, who built ListeningMind, whether people really ask in that much detail.

It is true that when consumers searched, two words or fewer than five words was the usual pattern. But now searches have become much longer than you would expect.

They are almost the same, but if just a few words differ it becomes a different query, so it is true that the long tail with very thin search volume has grown a lot.

He had his own example, from choosing schools with his child, who is preparing for university entrance.

My child is preparing for university entrance, and when we were choosing which schools to apply to, I think I used LLMs very heavily.

I just talked through all of these concerns at length, and it recommended suitable schools, took the application links, downloaded them as PDFs, sorted them out and handed them to me.

Education decisions need a lot of searching around, in many different ways. And I thought, oh, this kind of searching could all happen in one window, in one go.

He did not take the AI's answer on trust.

There is still a little uncertainty about whether everything the LLM gave me is right, so I would go back out to Google and check the school again, even whether the page really exists.

On the episode I said I had done something similar: I asked an LLM first for places to go with my child, then checked them again on Google and in map apps. Both are the same pattern, asking an AI at length and checking with search. Park described the difference between a query and a prompt this way.

Basically, with queries you pass your intent to the search engine by firing off several queries one after another, whereas a prompt can split that up or bundle it together.

The longer the questions and the more situation they carry, the more varied the CEPs become.

In practice, CEPs can exist in a huge variety.

✦Who should start first

Watch this part of the episode 16:19

I asked who needs GEO most, and whether anyone can still ignore it. Park said he would need to think more about who can ignore it, and started with the one thing he was sure of.

The more complex the product, say home appliances, the less consumers, apart from a few who are especially interested, fully understand the specs or the individual features of that product.

The worse the information asymmetry in an area, the more anxious consumers are, so I think they will lean more on an AI they believe has more information than they do.

He said home appliances are a category where his company has seen this happen a lot: consumers ask about all sorts of things. When that happens, brands find it harder to meet customers directly through traditional search engines, and Park sees it turning into three problems for brands.

The reasons brands need to care are, broadly, that acquiring leads and customers gets harder, second, the brand experience disappears, and third, it is getting harder to build up first-party data.

So every brand needs to think about it, he said, and he sized the impact by the length of the purchase journey.

For very cheap products, where consumers barely take a journey online in the first place, where they do not search around and just buy on the spot, the impact is smaller than you might think. But for things that have always had a purchase journey of at least a week and up to two or three months, or things that have always been bought together with other products as a bundle, I think it will become very important.

✦How to shortlist the prompts to manage

Watch this part of the episode 16:19

Put into working order, Park's points come to four steps.

1. Check which row of the table above your product falls in. If it is complex, has a long purchase journey or is bought with other products, start now.
2. Write down the moments when people think of your category (the CEPs), as situations: "I am thirsty", "the family has grown and we need a family car".
3. Work backwards to the questions people would ask an AI at that moment while considering a purchase. Leave out pure research questions.
4. Lay them next to the search queries you already received in the search engine era. By Park's account the range will not be very different; only the form has changed.

Once you know which questions to manage, how to put the answers on your pages is covered in How to get recommended by ChatGPT: why AI picks the brands it does, with Park's running tumbler example. How Korean search share has shifted between Naver and Google is in Seyong Park on Naver market share in Korea: 51 to 49 with Google. Asking a fixed set of questions like these every month is part of the AI search work on a Korean SEO project.

✦Common questions