From a Provider's First Quote to Expanding Revenue:
Awakening Soomgo's Growth Engine

Breaking down funnel data by cohort, a significant portion of the decline originated from new provider churn. Only 15% of new providers completed their first payment within 7 days of sign-up, and the 1-month payment rate dropped 6.3pp year-over-year. Providers who failed to convert to a first payment during this period almost always churned.

In other words, the overall efficiency decline was not only a pre-send problem but also the cumulative result of new providers failing to get established right from the engine's starting point.

Quote send conversion rate after viewing a request dropped from 35.7% to 32.7%, and the median time to send a quote increased from 8 hours 36 minutes to 10 hours 28 minutes. Conversion rate was declining while send decision speed was also slowing simultaneously. Following the same funnel further, cash purchase and store visit conversion rates within the quote-sending funnel were also declining. Drop-off was largest at the cash shortage notification modal in particular. Both the declining conversion rate and slower send speed, along with drop-off concentrated at a specific point, signaled the existence of a barrier causing providers to hesitate right before sending.

I redesigned the initial friction stage and structured the experience to bring the activation benchmark forward, helping new providers reach their first payment faster.

AS-IS

Cash shortage was only surfaced after the quote had been written, meaning drop-off could occur after the provider had already invested time and effort.

TO-BE

Changed to surface the cash shortage state before writing, creating a natural path into the store and cash top-up flow.

Next, I optimized the timing of state-based reminder messages within the request response stage for both post-first-payment and existing providers.

By redesigning the reminder timing and message context within the request response stage, I connected them into a conversion flow that improved both OR and CR together.

Experiment Result

OR +24.55% Quote Open Rate

CR +6.19% Quote Sent After Request Review

QR +32% Quote Conversion Rate

Across all providers who received requests, the conversion rate from reviewing a request to actually sending a quote was gradually declining. Providers were only receiving individual request notifications, and information about how much time had passed since receiving a request was barely accessible outside the product. Since many providers were checking requests intermittently while juggling other work, missing the initial notification often led to unread requests going unchecked.

Data showed that most providers sent their quote within 1 day of opening the request. This meant that even the same reminder could vary greatly in effectiveness depending on who received it and when. I therefore designed an experiment that branched reminder timing based on the request review status.

Experiment results showed that for providers who had opened but not yet sent a quote, 24 hours after opening was most effective; for providers who had not yet opened, 1 day before the request deadline was most effective. The former aligned with the pattern of providers checking requests in batches during specific off-work hours, while the latter was well-suited to trigger loss-aversion psychology — being the point after which responding becomes impossible. Overall, this improved quote open rate by 24.55%, quote send conversion after request review by 6.19%, and overall quote conversion rate by 32%.

By simultaneously improving new provider onboarding and the response conversion stage across all providers, conversion efficiency was recovered at both the starting point and the repeat stage of the growth engine. Redefining the first payment as the activation benchmark, reducing pre-send friction, and delivering messages aligned with the moment of action improved both the engine's starting point and its final stage.

Ultimately, this project was not merely a notification enhancement — it was a redesign of the conversion structure to help providers quickly move into real action after their first payment and repeatedly engage thereafter.

If quantitative data couldn't provide the answer, asking providers directly was the fastest way to identify the cause. Without additional product development, I implemented a survey modal directly using Braze in-app message custom code, collecting zero-party data through button click responses to understand the actual pain points providers were experiencing.

User interviews and surveys confirmed that what was blocking conversion was not a lack of features, but uncertainty about outcomes and behavioral barriers.

The goal was to create a state where new providers could 'just try sending once', and for existing providers, a structure that allowed faster pricing judgment and send decisions immediately after reviewing requests. The core of the solution strategy was not simply increasing exposure but more precisely removing the barriers right before sending.

Although the modal was deployed for survey purposes, an unexpected result emerged. Providers who clicked the CTA to access the guide content showed approximately 9% higher quote send conversion within 1 hour compared to those who closed the modal. It was concluded that the bigger problem was not an absence of information, but the absence of a delivery mechanism connecting necessary information right before the moment of action.

Rather than solving the problem with a single message, I concluded that different approaches were needed based on barrier type and redesigned the structure accordingly.

AS-IS

The existing system modal had information, but was not sufficient in its guiding role toward action.

TO-BE

The improved modal was redesigned to more directly address the core barriers of pricing a quote and writing the quote itself.

Based on these results, I further examined whether there was a mechanism that could actually alleviate the difficulties of 'pricing a quote' and 'writing a quote' — the main causes of drop-off for new providers at first payment and first quote. It was found that the existing AS-IS system modal displayed before quote writing was not sufficiently delivering the necessary information and action guidance, leading to the improved guide modal campaign. In the improvement experiment, quote send conversion rate improved significantly by 9.26%.

Challenges remained even after the activation improvement. Even when first payment was achieved, it did not lead to repeat behavior. The repurchase conversion rate within 7 days of sign-up was 7%, and repurchase rate among providers who made a first payment was only 30%.

Research identified customer non-response after quote sending as the primary cause of repurchase drop-off. The customer response rate was 18.8% against quotes sent, and only 26.8% against quotes opened. Non-response was a structural problem that was actually weakening repeat conversion.

Further analysis with the data team also confirmed that what differentiated repurchasers was not the speed of first payment but the experience that followed. Quote send volume, request open rate, and whether the provider experienced customer responses were the key variables.

Why Don't Customers Respond?

There were two reasons customers didn't respond. For those who didn't open the quote, price dissatisfaction was common; for those who opened but didn't respond, many didn't know what or how to reply. The gap between the information customers wanted (price comparison range, quote details) and the information providers supplied was blocking responses.

So What Information Needed to Be Reinforced to Generate Customer Responses?

Provider Behavior Nudge

For providers without customer replies, reconnecting them to pricing guidance and quote content reinforcement guide

Customer Behavior Nudge

App push designed to help customers recall the quote they received and initiate a response

There were two solution directions. One was to directly prompt customers to respond; the other was to increase provider appeal and strengthen their hiring competitiveness when exposed to customers. For customers, I designed an app push to help them recall the quote they received and initiate a response. Accounting for the two barriers of price dissatisfaction and the burden of responding, the approach focused on communicating the reasons to respond — such as price comparison and quote detail review — more clearly rather than simply offering benefits. Meanwhile, providers didn't know what information customers used to make hiring decisions after receiving a quote. I reconnected providers to pricing guidance and quote content reinforcement guides so they could better reflect the key information (experience, reviews, detailed description, quote content) that influences customer hiring decisions. Providers who sufficiently entered this information did receive more customer responses, and response experience was also connected to repurchase likelihood. Compared to the control group, customer response rate within 1 day of quote open increased by 17.8%, and the guide application rate among TG providers was 13.93% higher. However, the higher guide application rate also implied that, until then, necessary information was not being sufficiently delivered to both new and existing providers.

While reminder notification and guide modal improvements each addressed timing and information delivery, providers needed messages that delivered necessary information together at the right moment to actually take action. A channel was needed to rapidly test the optimal combination of timing and content, and the guide chat room was designed for that purpose.

However, the guide chat room experiment results differed from expectations. Across the entire TG, first payment rate, first quote send rate, repurchase rate, and hire rate all declined slightly compared to CG. But when the segment was broken down, a different picture emerged. TG users who did not click were pulling down the overall average, while users who did click showed much stronger follow-up behavior: first payment rate +226%, repurchase rate +285%, and 21-day hire rate +243% compared to CG.

This showed that the guide chat room was less a strong performance channel in itself, and more a channel that could capture or amplify users with high conversion potential when a click occurred. Therefore, the core challenge was not whether the channel existed, but what content and structure would increase click-through rate.

BigQuery data mart managing prediction scores at the provider·service level

Structure connecting BigQuery prediction values to Braze user properties

Operational structure branching segment-specific messages and send timing in Braze Canvas

However, the current segmentation structure had limitations in deriving the optimal segment × USP combination. Accordingly, I conducted a quote send prediction model experiment with data scientists and analysts. The model predicted quote sending probability at the provider × registered service level, using RFM variables and recent behavior data as inputs.

To design a structure connecting the derived prediction scores into actual CRM experiments, I collaborated with the DataOps team to load model extraction results into BigQuery, linked the necessary experiment variables to Braze user attributes, and built a pipeline enabling message experiments in Braze Canvas by segmenting based on the probability change score compared to service average (`score_14_diff`). While the effect was diluted in the overall average, uplift by segment was repeatedly confirmed when broken down by score_14_diff and behavioral experience. This confirmed that even for the same provider, the required intervention purpose and message tone differ based on state and experience — and we are currently exploring extending this into an operational structure that automatically branches messages and timing based on score and behavioral context.

Round 1 — Overall Average

Rounds 1 & 2 — Response Pattern by Segment

Round 3 — Score Segment × Quote Experience Status

When the reasons providers stop differ, even the same message works differently. In this project, barriers were classified by type, and the exposure timing and messages tailored to each barrier were designed to improve conversion in the repeat behavior stage.

Even in the prediction model experiment, rather than judging effectiveness from the overall average alone, I first looked at under what conditions differences appeared. Just as a question that started with a survey led to a modal experiment and then a guide chat room, an experiment that discovers the conditions under which differences occur is not a failure but a learning that shapes the design of the next experiment.

This was a project that first identified where users were stopping, then designed more precise messages for those points to improve conversion.

I started by narrowing the target. Analyzing provider revenue and activity data in BigQuery, I defined the core target segment for the ad product in SQL based on monthly revenue range, category, and repurchase patterns, then designed a demand validation and pre-demand capture strategy centered on that group.

※ At the time, Soomgo Plus operated under the name 'Top Fixed Ad', with winning bidders confirmed after a 2-week bidding period. Currently, the bidding mechanism remains but has been updated so providers bid using their held cash, and existing payments are automatically cancelled when a higher bidder appears.

I built an operational structure that minimized engineering resources while rapidly collecting sign-up intent and reconnecting at the launch moment.

Leading the open-alert signup event, I focused on building a structure that could be experimented with quickly from planning. Using Braze custom attributes, I collected sign-up intent without additional engineering and built a structure where automatic notifications were sent at launch.

Open-Alert Signup Operational Structure

Event Page Modal Display

→

'Yes' Click

→

Braze Custom Attribute Set

→

Auto KakaoTalk Notification on Launch

→

Actual Bid

Code for setting custom attribute when 'Yes' is clicked on the open-alert signup modal

{
  "eventType": "SET_CUSTOM_USER_ATTRIBUTE",
  "key": {
    "operator": "LITERAL_VALUE",
    "value": "adsopennoti_p-125_250529"
  },
  "value": {
    "operator": "LITERAL_VALUE",
    "value": "true"
  }
}

Designed using Braze without additional engineering to collect sign-up intent and track subsequent bidding behavior.

Capturing initial demand using event page and in-app modal

Driving actual bidding behavior through in-app and KakaoTalk notifications at launch

Through repeated segment × value proposition experiments, I subsequently built a message pipeline combining service and regional data to automatically reflect market-type USPs. To do this, I directly created a marketing table in SQL combining service, regional, and provider characteristic data, and designed a structure using Braze Connected Content to call that data at message send time.

Representative Automated Message Examples

Growth Opportunity Market

“⏳ $[Region]$ $[Service]$ peak season growth opportunity! With fewer bids right now, secure your first customers for just ₩800/day.”

Direct Request Expansion Market

“🎯 $[Region]$ $[Service]$ is a market top providers are watching! Use peak season ads to grow direct requests and have customers seek you out.”

Competitive Defense Market

“🚀 $[Region]$ $[Service]$ is hitting peak season maximum requests! Competition is fierce. Advertise now to grow direct requests and protect your peak season revenue.”

Efficiency-Focused Market

“☄️ Capture the $[Region]$ $[Service]$ peak season demand surge! Run efficient ads from just ₩800/day minimum to capture all requests!”

The same product was communicated with different USPs based on market type and provider characteristics.

Built the initial communication structure without engineering and strengthened message impact through segment refinement

Using only Braze — without additional engineering — I built the open-alert opt-in and automated send structure, enabling fast and stable initial communication operations. After the payment model shifted to immediate payment, CRM message attribution expanded significantly. Post-refund attribution rose from an average of 16.9% in rounds 1–5 to an average of 58.5% in rounds 6–9, peaking at 73.2%. This demonstrated that messages were functioning as a direct conversion trigger — not just a reminder. Subscription conversion rates and post-bid repurchase changes were analyzed directly in SQL and fed back into subsequent communication and segment design improvements.

※ Detailed figures and analysis data related to ad slots, bidders, and winning bidders are confidential; this document covers only the structure and key insights.