China's Liangpei charges $275 for an AI match — and refunds it if there is no marriage within 3 years

Traditional dating apps have long operated according to a strange logic: the longer users search, the more they pay. Endless feeds, paid Super Likes, subscriptions for the sake of subscriptions. Successful outcomes are bad for the product because happy couples leave and stop generating revenue. Liangpei ("good match"), a Chinese startup from Shenzhen, decided to turn that model on its head.

How it works

Onboarding is built around a voice conversation. The user spends about 20 minutes speaking with an AI matchmaker and answering 25 questions: relationship history, marriage plans, income, whether they own a home and, if not, when they expect to. The other person goes through the same process. The neural network builds profiles, calculates a compatibility score, and recommends a match.

Then come restrictions rarely seen in Western apps. Only one match can be active at a time. Until either person closes it permanently, neither can communicate with anyone else. This deliberately kills the dopamine-driven swiping mechanic: there is no reason to keep scrolling once you already have a specific person to focus on.

A similar shift is also underway in the West. After a match, Breeze does not open a chat; it immediately books a table at a bar. Known puts users through a half-hour voice interview with a neural network and charges only for a confirmed date.

Pricing that aligns incentives

On the paid plan, a user makes a one-time payment of 2,000 yuan (about $275). If they do not marry within 3 years, the money is refunded. If they do marry, there is no refund. The timeframe was not chosen arbitrarily: the service surveyed young people at a Shenzhen marriage registration office, and about 70% of them had married within 3 years of meeting their partner.

The important point is not the novelty but the change in metrics. Conventional dating services make money from retention and LTV. Liangpei makes money from delivering a fast result and effectively pays for its own failure through refunds. The company needs users to leave happy as quickly as possible. It is a rare case in which the business model and the customer's goal are aligned.

Matching as a search problem

Founder Zeng Xinxun came from search, not dating. He built search engines and recommendation systems at WeChat and TikTok and was recently a tech lead at Kimi. His thesis is that partner matching is a search problem, except the targets are people rather than documents. The team is now training a model that takes two profiles and estimates the probability that a relationship will work out.

A caveat is necessary: this is still an early-stage product, launched in July, and there are no public data on its actual marriage rate. The 70% claim is based on a survey conducted at a marriage registration office, not a controlled study, so the sample is biased toward people who were getting married anyway. Questions about income and housing also reflect a specific cultural context and cannot be transferred one-to-one to other markets.

What this means in practice

For product teams, the main lesson is that constraints can become features. One active match, voice onboarding, and outcome-based payment all act as filters for serious intent. For developers, it is notable that a voice interview with an LLM replaces questionnaires: it produces a richer signal than checkboxes while still imposing a 25-question structure that is convenient for scoring.

For founders, the conclusion is even more direct: if your revenue grows because customers fail to solve their problem, sooner or later the market will produce a product that profits from the opposite outcome. A money-back model requires confidence in matching quality, and AI makes that confidence at least somewhat justifiable.