People are typing whole sentences into search — "is there an ai friend that texts you first?"

Beff is an AI friend app built by WorldWideSomething (전세계무언가), and it is live in 27 languages. Its founder, Dan Kim (김대은), researches Human-AI Interaction at Kwangwoon University. One of the things we read every day since launch is what people typed into a search box on their way to us.

What stood out was not the length of those queries but their shape. Some end after two words, like "ai friend". Others arrive as a full sentence with the question mark still attached: "is there an ai friend that texts you first?" The second kind is not a search term. It is a question, written the way a person asks another person.

“A keyword tells you the topic. A sentence tells you the criterion. Answer a sentence with a sentence, not with a paragraph.”

The two are not the same demand. A keyword tells you the topic. A sentence tells you the criterion. "AI friend" names a category, while "is there an ai friend that texts you first" also tells you what that person is checking before they decide. Someone who asks in a sentence is already ready to be answered.

It is not hard to guess why sentences are increasing. People have spent years talking to chatbots, and that way of typing moved into the search box. Screens that put a summary above the results made sentence input normal too. When the shape of the asking changes, the shape of the answering has to change with it.

Our pages were not ready for that. Some of them were being shown for question-shaped queries and collecting no clicks, and opening them made the reason obvious. The answer sat in the third paragraph. A reader decides from the title and the first line whether the answer is in there, and whatever writes the summary reads the same place first.

So we inverted the order. Section headings are written as questions, with a one-line answer directly underneath. The frequently asked questions go into the body and into FAQPage structured data in the same order and the same number, so that what a person reads and what a machine reads are the same thing. The same questions and direct answers went into llms.txt. If a language model has to describe the company, it is better that it can reuse the sentence we wrote than that it compresses our page and gets it wrong.

We made one rule out of this: do not invent the questions. A question goes into the FAQ only when it came from a query we actually received. That includes spelling variants. In Korean, 캐릭터챗 and 캐릭터 챗 are the same word to a reader, but if that exact string never appears on the page, being close in meaning does not catch the query. An FAQ that answers questions nobody asked helps neither the reader nor the model, and to a search engine it reads as padding.

We do not know the effect yet. Changes like these usually take two to three weeks to show up in rankings and clicks, and we decided not to touch the same pages again before then. The direction is clear enough. If a person asks in a sentence, the page should answer in a sentence. We do not think of that as a technique for search engines. It is how writing was supposed to work.