Bing ShoppingBing Shopping

I led Bing Shopping design for four years, across three shifts in search: result matching, intent-based orchestration, and AI-generated pages. The old SERP pattern was no longer enough, but the replacement was not obvious. My direction came from real shopping behavior: search is not chat. Instead of forcing chat patterns onto search, I used AI to make the layout more useful for decisions. 我在 Bing Shopping 做了四年 design owner，完整经历了搜索的三次变化：结果匹配、按意图编排、 以及 AI 生成页面。旧的 SERP 形态不再够用，但替代方案并不天然清晰。我的判断来自真实购物行为：search 不是 chat。我没有把 search 硬改成对话，而是用 AI 让 layout 更服务于决策。

Bing Shopping is a small slice of search, but a clear lens on how search changed. Each product cycle forced us to redesign not only UI, but the logic of what to show first. Formats changed repeatedly; user goals did not. Bing Shopping 在搜索里只是一个切片，但它很能说明搜索是怎么变的。每一轮产品迭代， 我们重做的不只是 UI，还包括“什么信息该先出现”的逻辑。形态一直在变，用户目标没有变。

In practice, this work moved through three stages: matching, orchestration, and guidance. 落到设计实践上，这段工作经历了三个阶段：匹配、编排、引导。

In Phase 1, shopping answers were distributed by query-product match quality. The same card had to work in different SERP slots, in the Shopping vertical, and later across Microsoft surfaces. So I treated the card as a reusable spec, not a one-off component. 在阶段一，shopping answer 按 query 与商品匹配度分发。同一张卡片要同时适配不同 SERP 点位、 Shopping vertical，以及后续跨微软产品语境。所以我把它当成可复用规范，而不是一次性组件。

drag →

• Hard priority calls in tight space. Which fields must stay, which can be compressed, and what cannot be hidden at all for purchase decisions. 在有限空间里做优先级取舍。哪些字段必须保留，哪些可以压缩，哪些对决策绝不能隐藏。
• Two consistency systems in tension. The card needed to align with Bing's own design language while also fitting a cross-Microsoft shopping experience. 两套一致性同时拉扯。卡片既要遵循 Bing 自身设计语言，也要融入跨微软的 shopping 体验。

I decomposed the card into fields (Image, Title, Price, Ratings, Sellers, SKUs, Hashtags, Shipping, CTA) and prioritized them using flight data plus user research, not visual preference. To resolve the consistency tension, I set one rule: 我把商品卡拆成 Image、Title、Price、Ratings、Sellers、SKUs、Hashtags、Shipping、CTA 等字段， 并用flight 数据与用户研究确定优先级，而不是靠审美拍板。为解决一致性冲突，我定下一条规则：

If intent is the same, content priority stays the same; visual language can adapt by product (radius, spacing, type). Across products, the stable anchor is user intent. 只要 intent 相同，内容优先级就保持一致；视觉语言（圆角、间距、字号）可以随产品调整。 跨产品不变的锚点是用户意图。

replay

This principle became the basis for a body of shipped work across Bing's shopping surfaces, moving traffic and revenue at scale (rich nav +43k DAU and ~$1.2M revenue; image collage +50k DAU and $3–4M) and beginning to close the WPSBS gap against competition. 这条原则成了一批真实上线工作的基础，铺开到 Bing 的各个 shopping 面上， 带来规模化的流量与营收增长（rich nav +43k DAU、约 $1.2M 营收； image collage +50k DAU、$3–4M），并开始收窄对竞品的 WPSBS gap。

A single card pattern stopped being enough. Different queries imply different shopping tasks, so answer composition needed to vary by intent. My first attempt split by query type; research showed where that model failed. 单一卡片范式已经不够。不同 query 对应不同购物任务，answer 组合必须随 intent 变化。 我最初按 query type 切分，但研究很快暴露了这个模型的边界。

• No universal template per intent. We had to build hypotheses, then test them against UX and live signals. 不同 intent 没有现成标准模板。只能先建假设，再用 UX 与线上信号验证。
• Scale pressure on SERP. The taxonomy had to be understandable, actionable, and broad enough for massive query diversity. SERP 的规模压力。分类体系既要好理解、可执行，又要覆盖海量 query 多样性。

We began with high-traffic head queries, then reframed from "query type split" to "decision stage + intent + category." For each class, we defined expected user needs and validated them in UXLab and live data. This made composition decisions much sharper: 我们先从高流量头部 query 入手，再把框架从“按 query type 切”重构为 “按决策阶段 + 意图 + 品类切”。每一类先定义用户期待，再在 UXLab 和线上数据中验证， 于是组合决策更精准：

drag →

Serious comparison认真比较 Visual categories视觉品类 Broad exploration宽泛探索

Serious comparison认真比较

Not a wall of tables: summary, verbatim quotes, and source trails from social and video. Users accept AI synthesis only when provenance stays visible. 不是堆表格，而是“总结 + 原文引用 + 社交/视频来源路径”。用户接受 AI 总结的前提，是来源可追溯。

Visual · Phase 2

Review-focused answer: summary + quotes + source trails 评价型 answer：总结 + 引用 + 来源路径

assets/shopping/p2-intent-comparison.png

This answer family (Single Product, Gifting, Magazine, Top Products) shipped as one system. Top Products raised CCR by +23%; Magazine won APSAT in 63–69% of comparisons. The Single Task Pane I led on SERP improved WPSBS by +12.3 in its triggered set, and Algo+STP revenue grew ~20% YoY. I owned this stream end-to-end, from framework definition to answer composition. 这套 answer（Single Product、Gifting、Magazine、Top Products）作为同一体系上线。 Top Products 的 CCR 提升 +23%；Magazine 在 63–69% 的对比中赢得 APSAT。我主导上线 SERP 的 Single Task Pane， 在触发集上带来 +12.3 的 WPSBS 提升；Algo+STP 面营收同比增长 约 20%。这条线从框架定义到 answer 组合均由我端到端负责。

Phase 3 moved from curated templates to query-level generation. In GenSERP, page structure is assembled in real time per query. While much of the industry defaulted to chat UI, we found that search behavior is still browse → compare → narrow, so pure chat interaction was a mismatch. 阶段三从“人工配置模板”进入“按 query 实时生成结构”。在 GenSERP 中，页面结构会随每条 query 动态组装。 行业普遍转向 chat 形态，但我们发现搜索行为依旧是 浏览→比较→收敛，纯对话交互并不匹配。

• Define a search-native AI interaction. We needed guidance patterns tailored to search, not a copy of generic chat logic. 定义搜索原生的 AI 交互。需要的是适配搜索的引导机制，而不是照搬通用 chat 逻辑。
• Keep quality under infinite variation. With pages generated dynamically, we needed a durable system for hierarchy, consistency, and experience baseline. 在无限变体中守住质量。当页面动态生成时，必须有稳定机制去约束层级、一致性和体验下限。

Step 1 — AI enhances informationStep 1 — AI 增强信息 Step 2 — AI adapts the experienceStep 2 — AI 自适应体验 Step 3 — AI redefines SearchStep 3 — AI 重新定义搜索

Step 1 — AI enhances informationStep 1 — AI 增强信息

Question: How can AI make information better?

AI summarizes complex product information, surfaces new signals (e.g., price insights, review summaries), and helps users consume information more efficiently.

Design focus: Build trust through transparency and explainability. 问题：AI 如何让信息变得更好？

AI 总结复杂的商品信息，挖掘新的信号（如价格洞察、评价总结），帮助用户更高效地理解信息。

设计重点：通过透明度与可解释性建立信任。

Visual · Phase 3 · Step 1

AI summaries, signals, and cited insights inside search results搜索结果中的 AI 总结、信号与可追溯洞察

assets/shopping/p3-step1-ai-info.png

Across four years and three phases, tools changed and methods changed. The core did not: search exists to present information efficiently and help people decide better. AI did not replace that core; it expanded what we can do with it. 四年、三个阶段，工具在变、方法在变，但内核没变：搜索的价值仍是高效呈现信息，帮助用户更好决策。 AI 不是替代这个内核，而是放大它能做到的事。