11st Street's AI Search Doubled Purchase Conversion: What Narrow-Then-Search Means for E-Commerce
South Korea's 11st Street more than doubled purchase conversion with generative AI search. Here is how disambiguating intent before retrieval works, and what operators should implement in their own search and product data for the AI agent era.
Key Takeaways
- South Korea's 11st Street rolled out its generative AI search across all major categories, reporting that purchase conversion during the July pilot was more than double that of its standard integrated search.
- The core of the design is that results are not listed immediately after a query. The AI first surfaces buying conditions such as "lightweight and portable" or "optimized for gaming," and retrieval starts only after the shopper picks one.
- Structuring intent before retrieval mirrors how AI agents discover products, which leaves e-commerce operators with the task of embedding that condition vocabulary into their own product data.
A Generative AI Quietly Placed in the Search Bar Doubled Purchase Conversion
July pilot recorded AI-search conversion rate more than double standard keyword search
www.techtimes.comOn August 18, 2026, South Korean e-commerce major 11st Street announced the full rollout of its generative AI Search across all major categories of its mobile app. No new chat window was added. The existing search bar stayed exactly where it was, and what changed was the few seconds that follow after a query is typed.
The number that came out of it was from a one-month pilot in July with limited availability. According to 11st Street, purchase conversion for AI Search was more than twice that of the platform's standard integrated search during that period (Bridge Economy). Purchase conversion measures the share of search sessions that end in a completed transaction, and in most e-commerce environments that figure sits in the low single digits. At that baseline, a doubling reads less like a feature improvement and more like a change in behavior.
"Search is the critical starting point where shopping begins," said Seol Geon-ho, head of 11st Street's development group, adding that the team focused on narrowing the exploration range in an online environment with an overwhelming number of options.
Flipping "Search Then Narrow" into "Narrow Then Search"
The problem keyword search never solved is a simple one. Words match, intent does not. If three shoppers type "swimwear" wanting body coverage, UV protection, and the cheapest item in stock respectively, the system returns the same results from the same string. Sorting it out is left to filters, facets, and repeated searches.
11st Street's AI Search inserts a disambiguation step ahead of retrieval. When a shopper enters a product category, a language model trained on product domain knowledge infers the criteria that consumers in that area actually apply and presents them as granular search keywords. For laptops, that means "lightweight and portable" and "optimized for gaming." For swimwear, "covers body shape" and "UV-blocking." For eggs, "animal welfare certified" and "antibiotic-free certified."
Only after the shopper picks one does a second AI layer take over. It applies a multi-criteria analysis to the narrowed pool, weighting price competitiveness, delivery cost, delivery speed, and purchase review scores, and surfaces roughly five products that best fit the condition. Instead of a ranked list of thousands, the result is a set small enough for a human to actually compare. Related items and additional search terms sit alongside them, leaving an exit for shoppers who want to keep exploring.
A second entry point is what the company calls situational search. Everyday phrasing such as "gifts for a housewarming," "late-night snack ideas," "toys a young boy would enjoy," or "summer dresses for women in their thirties" is read directly as purchase intent, and the AI returns products against it. Shoppers who do not know retail terminology can reach products in their own words. Technically this requires mapping colloquial expressions onto structured product-category intent, which is the most computationally demanding layer of the system.
The difference from conventional filters lies here. A filter is a pre-built taxonomy that the shopper has to learn before using it. AI Search generates the relevant decision axes dynamically, per category. It works even when the shopper does not know the correct product vocabulary, and that is the foundation of the conversion gap.
How Far Can the Two-Times Figure Be Taken at Face Value
Some caution is warranted. The figure is company-reported, covers only one month of limited availability in July, and has not been independently audited. Shoppers who found a new feature during a pilot period skew toward users who are inclined to try new things in the first place. Whether the same gap survives a full rollout will only be answered by post-launch numbers.
There is a comparison point. Adobe Analytics reported that shoppers arriving at retail sites from generative AI sources during the 2025 holiday season converted 31 percent higher than those from other channels. 11st Street's number is far larger, but the two measure different things. Adobe's 31 percent reflects high-intent traffic sent in from external AI tools, while 11st Street's doubling reflects an intent filter inserted directly into the moment of search on its own platform. The latter is a more direct intervention, and structurally more likely to produce a wide gap.
The feature is also currently available on the mobile app only, with web and desktop unsupported, and the user base behind the measurement is skewed toward South Korea's mobile-first shopping behavior.
Why Korea's Number Three Bet on the Search Bar Instead of Warehouses
11st Street sits under SK Planet, part of the SK Telecom group, and its ownership was tidied up in 2025 when SK Square transferred its stake to SK Planet. That closed a long-running discussion with financial investors over an exit and freed the company to concentrate on operational moves.
The competitive numbers are unforgiving. Coupang's shopping app carries roughly 34.9 million monthly active users, while Naver Plus Store grew to 8.75 million in May 2026 and overtook 11st Street's 8.21 million to consolidate second place, according to Korean media. The Tech Times article this piece is based on places 11st Street second, but recent Korean tallies put it in third.
What 11st Street chose in response was profitability rather than scale. Its operating loss for January to March 2026 was 7.8 billion won, a 19 percent improvement year over year, and its core open-market business has held 25 consecutive months of profitability since first turning positive in March 2024. Marketing spend is being trimmed and fixed costs reduced through AI-based automation of product management and price response. Locking in existing customers around the free 11st Plus membership follows the same line.
Placed in that context, AI Search is not a feature addition but a strategic instrument. It increases transactions per search without building a single warehouse or same-day delivery lane. The search bar is one of the few conversion surfaces that can be improved without capital.
Korean E-Commerce Has Become a Testbed for Intent Resolution
The same movement is visible across the market. Naver deployed a conversational shopping agent into the Naver Plus Store app in the first quarter of 2026, using HyperCLOVA X to recommend products based on preferences, budget, and review data. On August 7, Coupang expanded a set of features including AI product comparison on detail pages and sentiment summaries of reviews, while Kakao outlined a plan to complete ordering and payment inside KakaoTalk conversations (Bridge Economy). We have tracked the broader Korean shift in an earlier article.
The direction is the same abroad. Amazon retired the Rufus brand in May 2026 and folded its capabilities into Alexa for Shopping in the search bar. Shopify reported that AI-driven traffic and orders tripled year over year in the second quarter of 2026. Naver's conversational agent is covered in more detail here.
What they share is a reversal of order: clarify intent first, then retrieve against it, rather than retrieving first and filtering later.
Structuring Intent First Has the Same Shape as Agent-Driven Discovery
This is where it gets relevant for e-commerce operators outside Korea. What 11st Street built inside its own app is a pipeline that converts vague shopper language into structured purchase conditions and narrows candidates to five through multi-criteria analysis. That is essentially the same pipeline an external AI agent runs when it looks for products.
The decisive difference is the raw material. 11st Street generates its condition vocabulary with full access to its own product database, purchase reviews, and delivery data. An external agent has only the product data a merchant publishes. If nothing equivalent to "animal welfare certified" or "antibiotic-free certified" exists in the product record, the item simply cannot enter consideration on that axis.
Read that way, the 11st Street case is both an internal search improvement and a specification for what product data needs to carry to be picked up by agents. Whether use cases, situations, certifications, and constraints exist as structured data matters identically inside and outside the storefront.
What Operators Should Implement in Search and Product Data
The first step is to identify the decision axes shoppers actually use in your main categories. Query reformulation patterns in search logs, meaning how a shopper rewrites a query within one session, expose axes that no filter currently covers. Words that are not catalog terminology, such as "quiet," "light," or "works in a rental," tend to be sitting there.
The second step is to carry that vocabulary into product data. Adding attributes is unglamorous work, but it pays off identically in on-site AI search and in external agent discovery. The mapping by stage looks like this.
| Stage | 11st Street's implementation | What your store needs |
|---|---|---|
| Intent interpretation | A language model trained on product domains infers decision axes | A list of decision axes per category, extracted from query reformulation logs |
| Condition surfacing | Granular search keywords presented for the shopper to pick | Structured attributes for use case, situation, and certification |
| Narrowing | Multi-criteria analysis on price, delivery cost, delivery speed, reviews | Machine-readable inventory, delivery terms, and review scores |
| Presentation | About five best-fit items plus related products and extra keywords | Grounds for cutting the list, and a path back to exploration when it misses |
| Natural-language entry | Situational search handles everyday phrasing | Life-context vocabulary in product copy such as gifting, season, and recipient |
Third is the willingness to return fewer results. 11st Street narrowed to about five items in order to take the comparison burden off the shopper. Narrowing requires grounds, and those grounds come from structured data on price, delivery, and reviews. Put the other way, an operator without that data cannot reduce the count.
Last is measurement design. The right metrics for AI search are not impressions or click-through rate but purchase conversion per search session and the rate of query reformulation. Whether reformulation fell is the most direct signal that intent resolution actually moved upstream.
Conclusion
11st Street's announcement offers one concrete answer to the question of where generative AI belongs in e-commerce. Rather than adding another chat window, it swaps the order of operations inside the search bar everyone already uses. Within the limits of a one-month, company-reported pilot, a doubling supports the direction of that design.
What matters next is whether the gap holds after full deployment. In a market where Naver's shopping agent is live and Coupang keeps thickening its comparison features, an advantage from any single feature is unlikely to last long. For operators elsewhere, though, the more important point is not who wins that race but the fact that the principle of structuring intent first lands squarely on their own product data.
