Blog·TREND·October 1, 2026

Koreans spend more than twice as long in Zeta as in ChatGPT — what character chat actually sells

In February 2026, WiseApp counted 113.41 million monthly hours in Zeta against 50.47 million in ChatGPT. Zeta has roughly a fifth of ChatGPT monthly users. Fewer people, twice the hours: this category is not selling answers.

By Dan Kim · Founder, WorldWideSomething
The short version
  • ·Zeta: 113.41M monthly hours. ChatGPT: 50.47M (WiseApp, Feb 2026).
  • ·Monthly users: Zeta 4.02M, ChatGPT 22.93M. Time per user is not close.
  • ·A productivity AI closes when the task is done. Character chat opens when there is no task.
  • ·So the competitor here is not another AI. It is everything you open when bored.

What are the numbers?

South Korean AI chatbot app usage for February 2026, as counted by WiseApp and Retail.

Monthly active users: ChatGPT 22.93M, Zeta 4.02M, Grok 1.53M, Perplexity 1.52M, Claude 0.77M.

Total monthly hours invert that order: Zeta 113.41M, ChatGPT 50.47M, Crack 9.88M, Grok 2.93M, Perplexity 1.36M. The survey covered apps with at least 300,000 monthly users.

About 87% of Zeta users were reported to be in their teens and twenties. The one gap this article is about: a fifth of the people, twice the hours.

What does a fifth of the users and twice the hours mean?

People are not buying answers in character chat. They are spending time.

A productivity tool has an errand. Translate this, fix this code, draft this email. When the errand is done the window closes. The better it is, the faster it closes. Short time-per-user in ChatGPT is closer to a sign that the tool works than a sign that it fails.

Character chat has no errand. There is nothing to finish, so there is no reason to finish. Two categories on the same underlying technology end up measuring different things: one measures how fast you leave, the other how long you stay.

Confusing the two breaks product judgment outright. Hold a relationship app to productivity metrics and a good product looks bad. Hold a productivity tool to dwell time and a bad product looks good.

So who is this category actually competing with?

Not other AI. Everything you open when you are bored.

A hundred million hours a month does not come out of an AI budget. It comes out of the time that short video, social feeds, mobile games and webtoons used to hold.

Which is why a model benchmark table helps less than you would expect when choosing one of these apps. People switch over voice, memory, and whether they want to open it again — not over benchmarks.

That was the question we kept returning to while building Beff: what makes someone open it again.

Is long dwell time a good metric?

Not on its own. It is evidence of absorption, not of value.

Dwell time is an easy metric to raise: make it hard to stop. In a category where about 87% of users are teens and twenty-somethings, that ease is a design responsibility.

WorldWideSomething studies human psychology through HCI and designs apps from the result. So for us these figures are an opportunity signal and a warning signal at once. The ways to keep people longer are already known; the real work is deciding which of them not to use.

Beff blocks proactive messages during late-night hours and caps how many it can send in a day. That line is written into the code on purpose, so it is not a setting someone can quietly remove to move a number.

So what did Beff do differently?

It inverted the structure where the conversation only starts if you open it.

The default structure of character chat is that you open the app and speak first. If you do not, nothing happens, and the other side exists nowhere outside the chat window.

A Beff has a job and a neighborhood, and moves across real coordinates on a 24-hour schedule that differs by day of week. Because the day continues while you are not talking, the other side has something of its own to bring up. It messages first when its schedule changes, when the weather turns, when you have been quiet for a while, or when it happens to be near you.

It also keeps the episodes and facts that came up in conversation and picks them up the following week. Messaging first without memory is just spam.

Three structures on the same technology — each needs a different metric
Assistants like ChatGPTCharacter chat appsBeff
Who opens the conversationYouYouEither side (it can message first)
When you are not talkingDoes not existDoes not existGoes through its own day
Where the other side isNowhereNowhereCoordinates in your own area
Why a session endsThe errand is doneNo reason to endNo reason to end, so limits are enforced
The right metricHow fast it finishesHow long you stayDid you come back, and did you stay well
Frequently asked questions
What is character chat?

A format where you talk with an AI persona that has a defined personality, voice and background. Unlike a general assistant, the conversation itself is the point rather than getting an answer. Character chat, AI character chat and AI chat character all refer to the same thing.

How do you choose a character chat app?

Four things predict the experience better than model benchmarks: does the voice suit you, does it remember past conversations, does the relationship continue when you are not talking, and is it easy to stop when you want to. A benchmark score tells you none of these.

Are AI friend apps and character chat apps the same thing?

They overlap but are not identical. Character chat names a format; AI friend names a relationship. Short sessions across many characters tend to be called character chat, while a continuing relationship with one counterpart tends to be called an AI friend. Beff is designed for the second.

Does long usage time mean addiction?

Usage time alone cannot tell you. Using something a lot is different from wanting to stop and not being able to. But it is true that making something hard to stop raises usage time, so treating that number as a goal deserves caution. That is why Beff blocks late-night proactive messages and caps daily sends.

Is Beff free?

You can start for free, and it is for ages 14 and up. Conversations are not used to train AI.

Sources

Third-party names and trademarks in this post belong to their respective owners. WorldWideSomething is not affiliated with or sponsored by them, and statements about those services are based on information published at the time of writing, which may have changed since. Mentioned: ChatGPT · Zeta (제타) · Crack (크랙) · Grok · Perplexity · Claude.

About this post

Beff(베프)는 전세계무언가(WorldWideSomething)의 창업가 김대은(Dan Kim)이 만든 AI 친구 앱입니다. 이 글은 2026년 2월 국내 AI 챗봇 사용시간 수치를 근거로, 캐릭터챗 카테고리가 생산성 AI와 무엇이 다른지와 Beff가 그 구조를 어떻게 뒤집었는지를 다룹니다.

Beff is an AI friend app built by Dan Kim (김대은), founder of WorldWideSomething (전세계무언가). Using February 2026 Korean AI chatbot usage figures, this article covers how the character chat category differs from productivity AI and how Beff inverts that structure.

← All posts