제품

요금제

솔루션

개발자 센터

연구

회사 소개

블로그

Select Language

플레이그라운드

엔터프라이즈 문의하기

고객 사례

When Video Became the Signal: How GS SHOP Drove a 57.5% Lift in Ordering Customers

Korea's leading home shopping platform had a video recommendation engine in name only — the system couldn't read what customers were actually watching.

By integrating TwelveLabs video understanding via Amazon Bedrock, GS SHOP turned 30-second clips into rich intent signals across four contextual dimensions.

The result: a 57.5% lift in ordering customers, a 29.4% conversion rate increase, and broadcast search time cut from hours to seconds.

고객사례 문서 확인하기

The result: a 57.5% lift in ordering customers, a 29.4% conversion rate increase, and broadcast search time cut from hours to seconds.

고객사례 문서 확인하기

Key Outcomes

The Challenge

Four-Axis Framework

Solution

Conclusion

지금 시작할 준비가 되셨나요?

무료로 시작하기

데모 예약하기

About GS SHOP

GS SHOP is one of South Korea's leading omnichannel commerce platforms, spanning TV live channels, data home shopping, and mobile live commerce. Its Short Pick experience presents product highlights in approximately 30-second clips, bringing content and commerce together in a personalized mobile shopping journey. Part of GS Retail. gsshop.com

Key Outcomes at a Glance

+57.5%

Lift in ordering customers

+29.4%

Conversion rate lift

+21.7%

More unique clicks

1-2hrs → seconds

Video search time per broadcast

The Challenge

Most "video recommendation" systems are simply traditional product recommendation engines with a video clip attached. GS SHOP's Search & Recommendation team realized their existing system had no way to interpret what a video actually meant to a viewer.

The platform relies heavily on "Short Pick" clips—30-second highlights manually curated from one-hour live broadcasts. For producers, finding specific visual moments (such as a host placing kimchi on boiled pork) meant manually scrubbing through hours of footage, routinely consuming an entire afternoon.

Furthermore, when a customer stopped scrolling to watch a clip, the legacy collaborative filtering engine remained blind to the visual context. If a customer watched a video featuring running shoes, the system simply surfaced more running shoes based on product category tags. It couldn't detect the styling, the lifestyle setting, or the specific aesthetic context that might indicate a latent interest in cross-category items like windbreakers, leggings, or outdoor gear. The video was functioning as a passive thumbnail rather than an active data signal.

Why TwelveLabs

Before committing to an infrastructural overhaul, GS SHOP initiated a focused proof-of-concept using TwelveLabs Marengo on Amazon Bedrock.

They stress-tested the model in Korean using a highly specific, multi-layered action query:

"A shopping host placing a piece of kimchi on boiled pork and eating it with chopsticks."

The query required the model to simultaneously process complex cultural vocabulary, identify the subjects, and track continuous action across frames. Marengo successfully located and returned the exact scene from a raw, one-hour broadcast tape in seconds.

This successful validation shifted the project from a cautious experiment to a full-scale production rollout. Building proprietary models in-house was ruled out due to the high costs of data labeling, GPU infrastructure, and specialized machine learning headcount. Because TwelveLabs models were natively accessible as fully managed APIs via Amazon Bedrock, GS SHOP could deploy advanced video intelligence instantly within their existing AWS environment without changing their vendor contract infrastructure.

Solution: What GS SHOP Built

Rather than replacing their validated product-based recommendation engine, GS SHOP extended it by layering a new video-context signal on top.

The Four-Axis Appeal Point Framework

To ensure the video signal captured context rather than just repeating product tags, Claude filters out brand and product names, mapping the video descriptive data across four distinct axes.

Functionality

Core product properties (e.g., waterproof, comfort fit).

Style & Sentiment

Aesthetic and emotional tones (e.g., casual, premium, relaxed).

Use Situation

When and where the item is featured (e.g., commuting, camping, outdoor activity).

Practicality

Real-life utility adjustments (e.g., easy storage, machine washable).

At recommendation time, a customer's real-time intent signals are combined with these video vectors and existing product data. A customer who clicks on a running shoe video is no longer trapped in a repetitive loop of more running shoes. Instead, the system recognizes the context of outdoor activity and lightweight practicality, seamlessly surfacing relevant items across entirely different categories—like athletic socks, windbreakers, or leggings.

"When we asked Marengo, in Korean, for a scene of a shopping host placing kimchi on boiled pork and eating it with chopsticks, and it came back in seconds with the right clip, we stopped treating video understanding as something we would have to build ourselves. TwelveLabs already had it ready on Bedrock. That let us focus on the real problem: how to turn that understanding into a recommendation signal."

Lee Chang-min,

Part Lead, GS SHOP Search & Recommendation Part

The Results

Metric

Performance Lift / Baseline Shift

Total Ordering Customers

+57.5%

Conversion Rate

+29.4%

Unique Clicks

+21.7%

Average Video Watch Time

6.3s → 8.0s

Broadcast Search Time

Reduced from 1–2 hours to mere seconds

Beyond the immediate metric lift, the new system protects the user experience by automatically flagging and filtering expired promotions or out-of-stock items after a live broadcast ends. The architecture also features built-in graceful degradation: if the video pipeline ever encounters a delay, the system seamlessly defaults back to standard product data without disrupting the user interface.

Conclusion

With production stabilized, GS SHOP is moving into the next phase: a data-driven content flywheel. By identifying which specific visual "Appeal Points" consistently drive customer clicks and conversions, the team plans to feed these successful parameters directly back into the creative process. Using generative video models on Amazon Bedrock, future Short Pick clips will be shaped directly by real-time audience performance data, fully closing the loop between video production and viewer intent.