北京智谱Zhipu AI

AI 产品经理AI Product Manager

我负责的主产品是 AI 搜索

如今,AI 搜索已经成为原子化能力,内化到了模型、Harness 等部分。

而当时的策略,都来源于一个具体的时代背景:2024 Q3,AI 搜索正在寻找自己的形态。

My primary product was AI Search. Today, search is an atomic capability embedded across models, harnesses, and related systems. In Q3 2024, AI search was still discovering its product form.

01 调研 Research

2024 Q3 AI 搜索百家争鸣A market in rapid formation

各家 AI 搜索在同一时期密集崛起。独立搜索产品、对话产品、传统搜索厂商都在寻找入口、过程和结果的最优形态;用户也在第一次形成“AI 可以替我搜索”的认知。

Search-first products, chat products, and incumbent search companies were simultaneously exploring the best form for entry, process, and results.

路线 01 Route 01

独立搜索产品2024 · 月访问量 Monthly visits

独立首页 · 过程展示 · 来源追溯

  • 360 AI 搜索287M
  • Perplexity94M
  • 秘塔 AI 搜索7M
  • 天工 AI 搜索5M
路线 02 Route 02

对话形态 AI 搜索产品2024 · 月访问量 Monthly visits

主对话入口 · 搜索调用 · 对话内呈现

  • Kimi29M
  • 豆包17M
  • 通义10M
  • 海螺 AI4M
  • 智谱清言3.6M
市场速度 Market
行业月访问量从 3 月不到 1,000 万,增长至 6 月约 1 亿、下半年近 7 亿。Aggregate monthly visits: <10M in March, ≈100M in June, and ≈700M in H2.
清言当时 Baseline
独立 AI 搜索约 2 万 DAU;主 Chat 为 60—90 万 DAU,其中搜索调用约占 10—20%。AI Search ≈20K DAU; main Chat 600–900K DAU; search use ≈10–20%.
共同形态 Convergence
市场逐渐在搜索入口、过程、引用和主对话协同上形成共识。The market converged around entry, process, citations, and coordination with the main chat.
产品洞察 Product insight

独立入口会产生独立预期。只要入口叫“AI 搜索”,用户就会对搜索产生预期。反过来,为了强化用户的“搜索”心智,产品本身也应该与 Chatbot 形态区分开。

A dedicated search entry creates clear expectations. Strengthening that search mindset also requires a product form distinct from a chatbot.

02 设计策略 Product strategy

让 AI 搜索“像搜索”Make AI search feel like search

接手时,产品仅具备 LLM + Web Search 能力,但产品形态仍延续 Chatbot 形态,没有拉开差异。改版沿着三件事展开——打开即搜索、运行中看见检索、搜索结果可溯源。

The product had LLM and web-search capabilities, while its experience still followed the chatbot form. The redesign focused on entry, visible retrieval, and traceable results.

01
首页 UI Entry

打开即进入搜索心智Search from the first glance

当时的问题 Problem
独立入口已经存在,页面仍沿用对话产品结构,用户体感难以理解。The dedicated entry still inherited a conversational structure, making the product difficult to understand through use.
设计动作 Design
强化 AI 搜索标识,以居中的搜索框建立首要动作;用热搜与建议词降低第一次输入成本。Center the search box, strengthen product identity, and use trending queries to reduce first-input friction.
产品判断 Insight
首页先完成品类教育,让“输入—搜索—获得答案”成为无需解释的主路径。The homepage teaches the category through a self-evident input, search, and answer path.

The homepage establishes category recognition before the first query: a focused search box, clear identity, and lightweight prompts.

02
搜索链路 Process

让用户看见它正在搜Make retrieval visible

当时的问题 Problem
新的搜索链路加入副模型筛选信息,检索网页更多、等待更长;新增的后台工作对用户保持隐形。A new filtering model searched more pages and took longer, while the added work stayed invisible.
设计动作 Design
把 Query 拆解、智能阅读、来源卡片和结果数量送到前台,用连续反馈解释等待。Expose query decomposition, source reading, cards, and result counts as continuous feedback.
产品判断 Insight
过程展示承担两项价值:管理等待预期,也让用户感知模型确实访问了外部世界。Visible process manages waiting and proves that the model is retrieving from the web.

Query decomposition, source reading, and progress feedback turn waiting time into visible product work.

03
可信体验 Trust

让引用承载“可信”的作用Make citations carry trust

当时的问题 Problem
AI 搜索的回答结果与 Chatbot 无异,用户难以区分幻觉和真实信息。AI-search answers felt similar to chatbot responses, leaving users unable to distinguish hallucinations from grounded information.
设计动作 Design
把全部来源放进持续存在的侧边栏,保留网站、标题和摘要;关键词与“全部来源”都能打开对应网页集合。Keep sites, titles, and summaries in a persistent source rail, grouped by query and available as one complete set.
产品判断 Insight
可信感来自可核验。引用需要成为产品结构的一部分,让用户拥有继续检查原网页的路径。Trust grows through verification. Citations need to be part of the product structure, with a clear path back to the original web pages.

Trust becomes a product interaction when users can inspect source titles, summaries, and original pages.

上线产品 Shipped product三条设计策略,汇入一次完整搜索。

搜索过程 · 16 秒 · 默认 2× · 用户身份与会话侧栏已裁切16 sec · default 2× · personal identity and chat history removed

这些设计在今天已经非常常见。回到 2024 Q4,它们属于第一批定义 AI 搜索产品形态的尝试;随后,过程可见、来源可核验与搜索侧栏逐渐成为 AI Native 产品的基础部件。Common today, these patterns were still being defined in late 2024. They later became foundational parts of AI-native products.

03 新功能 Experiment

知识体系:一次早于节奏的探索The knowledge map experiment

理念上,它希望 AI 搜索从“搜答案”继续走向“刷知识”,为一轮回答生成可继续探索的知识结构。

The feature extended answer retrieval into a stream of knowledge that users could continue exploring.

  1. 01
    创新意图Intent

    客观问答触发旁路模型,把用户问题扩展成三组知识点,点击后继续搜索。A side model expanded objective questions into three navigable groups of knowledge.

  2. 02
    上线前的担心Concern

    产品的搜索心智仍在建立,用户关注点还停留在“准不准”,对于创新形态的接受度可能不高。The search mindset was still forming, and users remained focused on answer accuracy, which could limit adoption of a new interaction form.

  3. 03
    上线后信号Signal

    上线后的行为数据验证了上线前的担心。当时搜索的核心还是搜得准,用户更关注回答以及来源。Post-launch behavior confirmed the concern: users cared most about accurate answers and their sources.

    6%知识体系点击率
    Knowledge map
    对比 vs.60%来源侧栏点击率
    Source rail

    两项数据来自不同功能埋点,用于观察用户选择方向。Separate feature metrics, compared here to show the direction of user attention.

个人观点:当时做的知识卡片略显急促,只是把对话底部的推荐问题改变了形式,真正的“猜你喜欢”还需要用户画像、推荐系统与搜索上下文共同参与。Personal view: the knowledge-card release moved too quickly. It mainly changed the presentation of follow-up prompts; true “for you” recommendations need user profiles, ranking systems, and search context.

04 产品规划 Planning

2025 AI 搜索产品规划The plan written at the time

2025 年 1 月 14 日,参与了 AI 搜索 2025 年度规划。根据上线数据、用户反馈、搜索技术链路,我把后续规划收敛为两个方向:可信 AI 与搜索原子化。

On January 14, 2025, I contributed to the annual AI-search plan and organized the next stage around trustworthy AI and atomic search.

2025.01.14规划依据:上线数据、用户反馈、搜索技术链路Inputs: launch metrics, user feedback, and the search pipeline.

01 · 数据指向 Data signal

可信 AITrustworthy AI

规划依据:来源侧栏点击率达到 60%,用户会主动核验回答依据。

规划动作:识别权威信源,通过多源交叉验证降低偏差,让引用持续可见、可以追溯,并处理信息冲突与内容时效。

Prioritize authoritative sources, cross-source checks, traceable citations, and recency handling.
02 · 链路指向 Technical signal

搜索原子化Atomic search

规划依据:AI 搜索是 LLM 信息的补充,该能力将会应用在不同的产品中,用户直接进入独立搜索产品只是当下的中间态。

规划动作:支撑清言主 Chat 的搜索调用,形成 API 与浏览器插件能力,并接入第三方产品、Agent 和 Workflow。

AI search supplements the information available to an LLM and can support many products; a standalone search destination was an intermediate product form.

产品最核心是要解决用户需求。当底层模型能力趋同时,Context 决定了模型表现,此时搜得准不准决定了产品是否能准确理解用户需求;因此,AI 搜索也从独立页面进入更底层的模型能力中。

Products begin with user needs. As foundation models converge, context shapes model performance, and retrieval quality determines how accurately a product understands those needs. AI search has therefore moved from a standalone page into deeper model capabilities.

05 其他参与 Also shipped

同期参与的清言产品工作Other work from the same period

三项工作均已上线,这里保留简要记录。

Three additional product contributions shipped during the same period.

01 · Shipped

地理位置获取Location awareness

让“附近美食”等查询获得可用的位置上下文;参与方案并推进上线。Added usable location context for nearby recommendations and supported delivery.

02 · Shipped

4o 数字人4o digital avatar

围绕导游与教师场景完成两轮用户研究、形象方向判断,并参与后续上线。Ran two rounds of research for guide and teacher scenarios, shaped the avatar direction, and supported launch.

03 · Shipped

打字机交互Streaming response

梳理多端生成与滚动体验,推动交互规则统一并上线。Aligned generation and scrolling behavior across clients and brought the unified rules to launch.