Korean Cases: Consultations, Finance, and Public Experiments
Review hospital no-shows, SMS and phone contact, insurance retention and comparison-to-contract, consultation-to-purchase cases, and randomized Korean studies with sample caveats.
Written and reviewed by Eric Han · Co-founder & CTO, reopt
Key takeaways
- University-hospital no-shows are generally 5–7%, mostly follow-up visits, and cannot be transferred to new consultations. Outpatient appointments arranged after emergency visits have 56.4% no-shows.
- Insurance comparison-to-contract is 9.2% for auto and 62.5% for travel, measured over different periods by the Financial Services Commission.
- Online insurance has the lowest first-year retention, 70.0%, but the highest fifth-year retention, 60.8%, excluding “other” channels. Do not rank channels from one year's figures.
- Korean randomized evidence exists: purchase rates within 24 hours of a coupon push notification range from 3.65% to 7.83% across four groups (lower value calculated); Naver Webtoon comment recommendations have control CTR of 1.819%.
- We found no consultation-to-contract rates for Gangnam Unni, Babitalk, Soomgo, Lawtalk, Zipdoc, or Zigbang.
Grades follow How to Read the Evidence and appear at the end of the first table cell. Choose inputs in the Korean Conversion Atlas.
Consultations, Appointments, and Finance
Appointment to Attendance
Most Korean data are hospital records dominated by returning patients. Borrow the principle of separating appointment types, not the rates for new consultations.
- Outpatient bookings arranged in emergency care56.4%11,000 appointments
- Tertiary-hospital diagnostic bookings30.1%Subset of 1.25M appointments
- Colonoscopy, already sending two SMS reminders16.0%2,401 people
- Online bookings, university hospital10.8%3,921
- New patients, university hospital9.3%8,515
- Ten national university hospitals7.1%13.61M people
- University hospital overall6.3%300,000
- Tertiary hospital overall6.1%1.25M appointments
- Phone bookings, university hospital4.9%43,000
| Source | Figures | Caveats |
|---|---|---|
| Seoul tertiary hospital, B | Overall 6.12%; consultation 4.5%; treatment/surgery 16.6%; diagnostics 30.1% | Three months in 2018; 94.1% follow-up bookings. Same hospital reported 5.0% in 2013–2014 |
| Same hospital, B | In-person booking 10.6%; follow-up 5.9%; online/phone 4.7% | Some adjusted models reverse the online/phone versus in-person ordering |
| One university hospital, B | General 6.5%; phone 4.9%; internet 10.8%; new patients 9.3%; returning 6.2% | Unadjusted 2016–2017 comparison. Abstract says 6.8%; table implies 6.3%, calculated |
| Ten national university hospitals, B via press | 7.1%; hospital range 1.2–14.1% | January 2022–June 2023, lawmaker's data. Numerator reported as same-day cancellations, potentially a different definition. Advance notices already in use |
| Post-emergency outpatient bookings, B | 56.4%; under one week 51.0%; three to four weeks 84.0% | Booked by clinicians, not patients; supports separating third-party bookings only |
| Kim's Eye Hospital, D | Attendance 84.3% with SMS versus 72.5% without; 1.16×, calculated | 2004, nonrandomized; only patients supplying numbers received messages |
| Asan Medical Center endoscopy, B | Bundled changes including two to three SMS reminders reduced no-shows 16% → 14.1% | 2008 before/after, not SMS versus none |
| National forest campsites, D | No-shows 10.4%, 84,881 cases | 2015, secondary citation; rare nonmedical figure with denominator |
| Kakao Hairshop, D | Prepaid no-shows 0.5%, 0.09%, 0.39% | Company announcements. 0.5% covers a two-month 2016 preview; other periods undisclosed. Does not identify prepayment's causal effect |
We found no Korean randomized SMS attendance experiment comparable to Cochrane's RR of 1.14. Kim's Eye Hospital's similar 1.16 ratio may be coincidental. Phone connection figures include 24.7% in a Daegu screening experiment, about 12% in a Goyang pilot, or 8.7% of 2,065 eligible people, calculated, and Gimhae health center's usual 32% versus experimental 46.9–72.2%. These are public-agency outbound calls, unlike direct inquiry leads.
Insurance and Loans
Public data compare post-purchase retention for direct and solicited sales.
- Agency, face-to-face88.9%Year 5: 47.2
- Bancassurance88.6%Year 5: 38.5
- Exclusive agent, face-to-face88.4%Year 5: 47.2
- Telemarketing75.9%Year 5: 42.4
- Online70.0%Year 5: 60.8
Excluding Other, online is lowest in year one but highest in year five, month 61. These durations refer to different acquisition cohorts and must not be multiplied sequentially. Regulators also explain that cancellations rise after commission-clawback periods end. Month-25 bancassurance retention moved from 67.7% to 79.5% in one year, while online moved from 74.0% to 71.0%; a single year cannot establish channel rankings.
| Source | Figures | Caveats |
|---|---|---|
| Insurance comparison/recommendation service, B | Comparison → contract: auto 9.2%; travel 62.5% | FSC data. Auto: January 2024–February 2025; travel: February 2025 average. Before March 2025 price unification, version 2.0. FSC suggests people comparing then buying directly from insurers are omitted |
| Same source, B | About 1.85 million uses and 280,000 contracts: 15.1%, calculated | Products and periods pooled |
| Korea Insurance Research Institute 2022, B | Long-term nonlife month-13 retention: direct 85.0%, phone 74.4%, agent 85.7% | 2021. The 2025 all-product table reverses online/phone ordering; products and commissions differ |
| Seomin Finance Itda, B via press | Referral → funded loan 4.0% in first-half 2026, 5.4% in 2025; primary banks 2.5%, secondary lenders 4.3% | Denominator is product referrals, not applications. Counts such as 1,498,168 disclosed in lawmaker data. The high share of borrowers with low credit scores limits generalization to other lending contexts |
| Finda, D | Application → agreement about 60%; 820/1,400 = 58.6%, calculated | Company announcement; applicants had already viewed rate/limit. One product's first month; nine-company aggregate after about two months was around 50% |
Consultations
| Company | Figure | Caveat |
|---|---|---|
| Bullsone, D | Consultation → purchase up to 15% | 100 purchases from 1,300 monthly consultations imply 7.7%, calculated; tool-vendor announcement |
| Bullsone, D | Three-month repeat purchase 18% → 55.4% | May include only consulted customers |
| Homefit, D | Inquiries 6–7 times higher after chat | Customer CEO comparing with phone-only service; inquiry count only |
| Ohouse construction, D | Contract conversion more than doubled | Baseline undisclosed |
| Naver Place tools, D | Use associated with booking traffic 3.1×, TalkTalk 2.2×, Smart Call 1.2× | Traffic metrics and self-selection among businesses adopting tools |
No consultation-to-contract rates were found for Gangnam Unni, Babitalk, Soomgo, Lawtalk, Zipdoc, or Zigbang. Medical-agency claims are reviewed in Consultation and Photo-Upload Forms.
Academic and Public Experiments
The studies below use Korean data and disclose random assignment, except that Ghose randomizes only coupon recipients. Few measure purchases, and no-coupon controls are uncommon.
| Study | Design | Result | Caveats |
|---|---|---|---|
| Hwang et al., 2023, A | Ringtone-app coupon pushes, 2017; 6,984 assigned | 24-hour purchase: immediate coupon 4.17%, expiring held coupon 7.13%; negative wording 6.24%, positive 5.03%; four combinations 3.65%, calculated, to 7.83% | No no-coupon control. Excludes 168 held-coupon users who redeemed before the push; authors call this conservative. Some industry-affiliated authors. Low-price digital goods with 30% discount |
| Lee et al., 2026, A | Naver Webtoon comment recommendations; about 595,000 per arm; four weeks | Control CTR 1.819%, click → read 41.095%; LLM content prior CTR −5.68%; gender prior +1.48%, p=0.144 | Company preprint; reading, not purchasing |
| Lee et al., 2020, A | Santa TOEIC-app score-prediction model; 17,019 versus 50,451 users | Paid purchase 2.73% versus 2.37%; ARPU $3.23 versus $2.83 | Company preprint. Random assignment and arm sizes disclosed, no significance test; back-calculated counts give p≈0.009. Tests a prediction-model replacement, not a purchase prompt |
| Hong and Kam, 2014, A | Screening invitation letters to 453 Daegu men aged 50–59 | Gastric screening 22.2% versus 17.9%; colorectal 16.5% versus 13.5%; nonsignificant | About 225 per arm, so nonsignificance does not establish no effect. Control also received national-insurer notices. Phone arm analyzes only reached people and is not causal |
| Jung et al., 2023, A (summary) | Review-request pushes in travel/apparel marketplace; over 300,000 people | Magnitude unverified; immediate requests are not always best | Full text blocked; press release only |
| Chu and Joo, 2024, A (summary) | Mobile coupons at health/beauty chain; over 500,000 people | Magnitude unverified; amount-emphasis, personally connected messages, and weekday sending perform better | Abstract only |
| Ryu and Kwon, 2025, B | Advance SMS to 480 high-risk Gimhae health-center clients | Phone pickup 72.2% versus 46.9%. Visits per assigned person: tailored conversation 7.5%, standard 3.75%, no SMS 2.5%, calculated; paper reports 11.7%, 4.7%, 5.3% per reached person | Body says randomized, ethics section says observational. p-values mismatch counts; pre-call pickup differs by 15.6 points between arms that should not yet differ, raising allocation-quality concerns |
| Ghose et al., 2019, B | 14,741 free transport-app coupons over 21 days | Redemption 34.8%; one-coupon model predicts commute route 37.8%, noncommute 19.5% | Commuting not randomized. Abstract's roughly threefold effect comes from a multiple-coupon model; one-coupon ratio 1.94. Accepted manuscript |
Papers by Naver, Kakao, and Karrot authors also report online A/B gains without baselines: Naver long-tail search CTR +5.63%, Naver Cafe recommendation clicks +5.4%, Karrot ad CTR +7.46%, and Kakao Brunch CTR 1.212×. Under this book's rules they are D because arm-level samples and baselines are undisclosed; Karrot also lacks country separation. Public-nudge coverage of National Tax Service arrears wording gives no effect size.