Taobao's AI Universal Search: A Business Teardown

Project positioning Taobao’s AI Universal Search is built to capture the complex, fuzzy search intent that content platforms like Xiaohongshu (RED) and Douyin have been winning — blending e-commerce and local-services data to break out of the traditional “clearly-specified product” search model, and bundling search, price comparison, and product selection to increase time-on-platform and reclaim both traffic distribution power and ad revenue. Longer-term, AI Universal Search could grow into a full Taobao-ecosystem consumption agent, unifying in-app functions while connecting far-field goods, near-field retail, local services, travel/hospitality, and ride-hailing into one-stop service.

Summary and Key Findings

• Current status: gray-testing began in June 2025 on a single-digit-percent slice of users, with plans to expand to roughly 10–12% by mid-July and go fully live by the end of September at the latest. Gray-test metrics are currently coming in below target — usage rate and adoption rate are both meaningfully under plan — driven mainly by weak user awareness, limited promotion, imperfect intent-matching, and response-quality gaps.
• Query distribution: Taobao internally uses two query taxonomies — a traditional one (clear-intent, fuzzy-intent, needs-intent) and an aspirational one (single-category, complex-planning, cross-platform-needs). AI Universal Search routes clear-vs-fuzzy at the intent-recognition stage, falling back to the traditional search/recommendation stack for clear-intent queries (fuzzy-intent plus needs-intent queries make up a meaningfully larger share of traffic here than in traditional search).
• Iteration cadence and focus: the product iterates on a three-week cycle, focused on intent recognition, response latency, response quality, and user adoption. Intent recognition is largely meeting user needs already; response-latency issues are expected to improve soon; response quality urgently needs Taobao’s search team’s data to improve further; user adoption has no clear fix yet and is being tracked long-term.
• Metrics framework: the North Star metric is total Taobao/Tmall search volume (with an ambitious year-over-year growth target for 2026). Relative value is tracked via: 1) search volume and the share of fuzzy searches; 2) user time-on-feature, adoption rate, and average decision time; 3) the share of GMV attributable to AI search.

Related interviews:

• An operations expert on Taobao’s AI Universal Search
• An AI product expert on Taobao’s AI Universal Search

---

I. Background and Operating Results

1.1 AI Search vs. Traditional Search

Traditional Taobao search struggles to serve complex, fuzzy user intent (fuzzy queries make up the large majority of real-world shopping needs, but a much smaller share of Taobao’s own search traffic) — pushing users, when their shopping need is unclear, to turn to Xiaohongshu and Douyin for inspiration first, and pulling brand marketing budgets toward those content platforms for “seeding.” The end result: the value of on-platform search erodes, hurting both platform revenue and Taobao’s grip on traffic distribution. AI search exists to recapture value on queries with high order value and long decision chains.

1.2 Taobao AI Universal Search

As 2C agents keep advancing, the pressure on vertical e-commerce platforms will keep building. In 2025, Alibaba again consolidated its consumer entry points — Shanguo (flash sale) folded into the main Taobao app, and Eleme and Fliggy merged into Alibaba’s China e-commerce business group. AI Universal Search extends “Taobao Wenwen” (Taobao’s existing Q&A assistant):

1. Unify data: blend e-commerce and local-services/social data, breaking past the category-based search model to boost time-on-platform through integrated search, price comparison, and product selection.
2. Become a consumption agent: has a real shot at becoming Taobao’s future consumption agent, offering one-stop service that unifies in-app functions (direct customer-service handoff, order-status lookup, profile/nickname customization) and connects far-field goods, near-field retail, local services, hotel/travel, and ride-hailing. AI Universal Search is embedded directly in the primary search module and carries high internal priority.

Project timeline and gray-test status:

• Prep phase: requirements gathering and development began in March 2025.
• Gray-test phase: launched in mid-June on a single-digit-percent slice (randomly sampled across roughly 10 dimensions including age, spending power, and phone model), running for 1.5–2 months, with expansion to roughly 10–12% planned for mid-July (no further expansion planned beyond that).
• Gray-test data: metrics are currently below target. Impression rate is high (measured as a 2-second-plus dwell); adoption rate is meaningfully below its target range (currently overstated due to a mismatch between recommendations and actual need); usage rate is low (weak promotion, weak user awareness).
• Full rollout: planned for no later than the end of September 2025, driven partly by competitive pressure from AI search features launching on community platforms like Xiaohongshu.

1.3 Comparison with “Taobao Wenwen”

Taobao Wenwen and AI Universal Search sit under different teams, with similar technology and function but a much deeper entry point for Wenwen.

• Team ownership: AI Universal Search is led by Taobao’s core-framework team; Taobao Wenwen is led by the search team.
• Entry point and interaction: Taobao Wenwen is a consumer-facing chatbot that predicts needs from a user’s order history and cart data, with built-in themed templates (e.g. a “hard at work” mode, a “compliment machine” mode). But it only triggers when a user explicitly types “Taobao Wenwen” into the search box — a fairly deep, easy-to-miss entry point.
• History: Taobao Wenwen entered beta in September 2023 and rolled out fully during Double 11 that year. During beta it experimented with going beyond a pure shopping assistant (a wedding planner, a trip planner), but ultimately narrowed back to product recommendations due to technical constraints and strategic focus.
• Performance: underwhelming. DAU sits in the high hundreds of thousands; transaction-conversion rate (converting sessions with a purchase, ad traffic excluded) is low for precise-keyword queries and even lower for broad-keyword queries; next-day retention is very low; responses run too long, and most users exit before finishing reading (dwelling only a couple of seconds); the product never built real user habit.

---

II. Primary Search Types and Core Features

Taobao currently runs two internal query taxonomies — a traditional one and an aspirational one. AI Universal Search still uses the traditional taxonomy for now, and may migrate toward the aspirational one over time. A single AI Universal Search query runs meaningfully longer than a traditional search query.

2.1 Traditional Query Taxonomy (Currently in Use)

At the intent-recognition stage, AI Universal Search classifies roughly half of queries as “clearly-specified” and reuses the traditional search stack directly for them; the remaining fuzzy/needs-intent share is meaningfully larger here than it is in traditional search.

Search Type	Meaning	Share in Traditional Search	Share in AI Universal Search
Clear intent	The user knows exactly what to buy, and the product is fairly standardized. E.g. “buy an Estée Lauder Advance Night Repair bottle.”	The large majority	Around half
Fuzzy intent	The user’s need is relatively vague, with many possible product specs — typically resolved by browsing and comparing a set of results. E.g. “buy a refrigerator.”	A minority	Around half (combined)
Needs intent	The user can describe the problem clearly but can’t translate it into a product category, so search can’t surface a match. E.g. “what do I do about a red bump on my scalp?”	(folded into fuzzy-intent share)	(folded into fuzzy-intent share)

2.2 Aspirational Query Taxonomy (Future Target)

Search Type	Meaning	Example	Expected Query Share
Fuzzy-need mining	AI performs logical and intent analysis, explaining its reasoning rather than just listing products.	E.g.: what shampoo should I use for a red bump on my scalp. Fits: health/personal care, household goods	The large majority
Open-ended decision support	Infers relationship, gender, age, and interests, generates a shortlist against a budget, and pulls in livestream/social content to justify the picks.	E.g.: what should I get my best friend for her birthday. Fits: gifting (beauty, electronics, apparel)	(folded into the above)
Complex scenario planning	Pulls in travel-booking components and computes a bundled set of Taobao product recommendations (not just storefronts).	E.g.: planning a 5-day Yunnan trip with my parents (flights, hotel, sun gear, travel gear). Fits: travel, family activities, gatherings	A meaningful minority
Cross-platform needs integration	Connects inside and outside the Alibaba ecosystem, reading livestream history and linking to products, supplementing with third-party test reports.	E.g.: find the whitening serum a specific livestream host sold. Fits: livestream commerce, brand exclusives, cross-border goods	(folded into the above)

2.3 Core Features of AI Universal Search

Current features: user Q&A, question answering, product recommendation, guided search

Planned features:

Planned Feature	Description
Emotionally-aware interaction	Infers user mood from chat history — e.g. detecting stress and recommending a stress-relief diffuser.
Multimodal generation	E.g. given “camping gear,” AI generates a camping guide with images and text.
Decentralized search	Adds AI scanning (image, voice), using OCR and ASR to understand image and voice content.
Unified functionality	Handles customer-service escalation, product-feature configuration (avatar/nickname design), and more directly within the search box.

---

III. Technical Implementation

• Base model: a mixture-of-experts architecture, with a shared base model fine-tuned separately per category; each category-specific model ships as soon as its own tuning is complete.
• Data sources: internal Alibaba databases, blending e-commerce data (orders, cart contents, beauty-profile data) with local-services and social data (Fliggy, Eleme), plus external crawled text (Xiaohongshu and other vertical creators, transaction and payment map data).
• Product roadmap: 2025 is focused on data integration and feasibility testing; 2026 shifts to product and feature iteration, with a priority on women’s consumer categories (chosen to fit the platform’s overall positioning, targeting a demographic rather than a category).

3.1 Iteration Cycle (Three Weeks per Loop)

Week	Iteration Stage
W1	Monitor + PRD + reporting (fast validation): monitor already-shipped features; user research supports data collection.
W2	Development + continued monitoring: mostly focused on data infrastructure (building internal labels, integrating with other teams’ data), with limited product feature work (unchanged through Q4).
W3	Testing + launch + iteration: iteration focuses mainly on existing high-volume categories (given data volume), and isn’t always category-specific tuning — sometimes it targets general user pain points instead.

3.2 Iteration Focus

Focus Area	Current State	Status
⭐ Response quality	Currently reads more like a generic assistant than “a shopping platform giving you personalized advice.” The fix is integrating Taobao search team’s data for further training (a meaningful majority of the needed data is still missing).	Long-term tracking (urgent)
Intent recognition	The base model is strong enough to largely meet user needs already.	Largely solved
Response latency	Time-to-first-token is the current focus, expected to resolve within a matter of weeks — not technically hard, mostly a compute-allocation issue.	Short-term fix
User adoption	User trust and awareness remain weak; many assume the AI is just steering them toward a purchase. Needs sustained, long-term habit-building.	Long-term tracking

3.3 Metrics Framework

• North Star metric: total Taobao/Tmall search volume (an ambitious year-over-year growth target for 2026, measured against a trailing-year baseline).
• Value dimensions: 1) is the share of fuzzy search rising; 2) is user time-on-feature increasing; 3) is GMV conversion improving.

Business- and technical-layer tracking metrics:

Dimension	Metric	Definition	Current / Target
Business	Adoption rate	Multi-turn dialogue implies progression (not repeating the first question means the first turn was “adopted”), and post-recommendation dwell time exceeds a short threshold	Currently well below target (inflated somewhat by a mismatch between recommendation and need); meaningful room to reach the target
	Usage rate	At least one query issued (gray-test phase has no promotion; full launch will invest in long-term habit-building)	Currently low
	Impression rate	Counted at 2+ seconds of dwell time	Currently high (likely understated versus reality)
	Average decision time	Time from first search (need arises) to placing an order, excluding searches that don’t convert	Currently several minutes; targeting a large reduction
	AI-attributed share of GMV	Weighted attribution (first exposure/first click and last exposure/last click each carry partial weight)	Currently a small single-digit share; targeting a meaningfully larger share
	Above-the-fold CTR	/	Currently a single-digit share; targeting a modest improvement
	Overall CVR	/	Targeting a meaningful sequential improvement
	User interaction time	/	Currently under a minute; targeting a meaningful increase
	Retention / turn count	/	/
Technical	Model generalization	Coverage of standard sales scenarios	Currently around half; targeting a large majority
	Relevance precision	Manually assessed, weighted by opportunity size per industry	Targeting the low-to-mid 90s percent
	Average response time	/	Targeting sub-second
	Intent-recognition accuracy	/	Targeting 90%+

---

IV. Team Structure

• Algorithm: a meaningful team, with partial support shared from the foundation-model team
• Product: a small dedicated team
• Engineering: a dozen or so, fully shared with the main-site team
• Support teams: small dedicated support from both the user-research and data-analytics teams

Appendix: References

• AI E-Commerce Has Really Arrived (industry commentary)
• The Year of Large Models: Taobao’s Universal AI Has Arrived (industry commentary)
• Baidu Baike entry on Taobao Wenwen