Independent PoC proposal

Eight weeks to establishwhether arrival drives payment

The MVP shows that transactions occur. Whether they are incremental, caused by the offer rather than coinciding with it, cannot be judged without a control group. Eight weeks in the Yeongilman district, randomised at the trip level, to measure that increment.

District: Yeongildae Beach, Pohang
Target merchants: 20 shops
Radius: 5-min walk · ~400m
Duration: 8 weeks

Starting point: a limitation we recorded ourselves

“This is a transaction confirmed in a demo. It does not establish that the customer was new, or that the purchase was caused by the offer. Validating the business requires a controlled experiment.”
— DANGDO MVP, merchant performance screen

We left this note in the product because presenting unverified results as verified is a recurring failure in this category. The PoC is designed to resolve that limitation rather than restate it.

The site: Yeongilman

The Yeongildae Beach district in Pohang is the first site. Trips there concentrate on weekends and evenings, and cafés, bakeries and restaurants cluster within a five-minute walk. Validating first in a regional district rather than the capital area is central to this proposal.

Experiment design

Unit of randomisation: The unit is a single trip. Randomising per trip rather than per user allows one person's other trips to serve as controls.
Treatment: Offers surfaced fifteen minutes out, orderable through Kakao Pay.
Control: A pure holdout. No offers are surfaced; payments in the same district over the same period are observed.
Allocation: 50 / 50
Blinding: Customers cannot tell whether they are in the control group. Merchants receive weekly aggregates only, never individual assignments.

Metrics

Primary

Share of trips with a Kakao Pay merchant payment inside 400m within 60 minutes of arrival

We do not count only payments routed through Dangdo. Counting all district payments is what separates demand created from demand merely redirected.

Secondary

Average payment value
Share of first-time payments at that merchant
Incremental payment per ₩1 of benefit
Weekly settlement per merchant
Return visit within four weeks

Guardrails

Offer notification opt-out rate
Cancellation and no-pickup rate
Merchants leaving mid-experiment
Budget burn rate

Statistical power

The eight-week window is derived from a power calculation. The basis for it is set out below.

Assumptions

Baseline conversion: 6%
Detectable effect: +3.0pp (+50% rel.)
Significance: 0.05, two-sided
Power: 80%

Required sample

1,209 trips

Required sample · per arm

2,418 trips

Total

~302 trips

Per week

~43 trips

Per day

Limitation. Detecting +2.0pp under the same design would require 2,554 trips per arm, 5,108 in total, or 91 a day. One district over eight weeks does not supply that. Measuring it would require a second district or a twelve-week run.

Eight weeks

Weeks 1–2
Onboarding and baseline

Sign 20 merchants and load their menus and margins. No offers yet; measure what the district does normally.
Week 3
Randomisation on

Begin 50/50 trip-level assignment. The first week is for validating the data pipeline, not for reading the effect.
Weeks 4–5
Steady state

Run with the pricing policy frozen. A deliberate no-peeking window prevents early-stopping bias.
Weeks 6–7
Policy iteration

Adjust exactly one pre-registered variable, the benefit cap, to find where incremental payment per won begins to fall.
Week 8
Holdout restoration and analysis

Return part of the treatment group to control and check whether the lift holds, separating durable effect from novelty.

What would count as success

Fixed in numbers before the experiment begins, so the bar cannot move once the result is in.

Primary: +3.0pp or better, with a 95% CI lower bound above zero
Unit economics: At least ₩2.5 of incremental payment per ₩1 of benefit
Merchants: 80% retained through week 8, 60% intending to renew
Riders: Opt-out at or below 3%, no-pickup at or below 5%

If the bar is not cleared, we stop. The finding would be that trip data alone does not meaningfully lift district payments, which is still a result worth having.

Data and privacy

Destination and arrival time are used for offer discovery on that trip only. Per-trip identifiers are not retained once it ends.
Analysis uses trip-level aggregates and assignment flags only. No route history or fine-grained location data is collected.
Offers are opt-in and can be switched off at any time. The control group receives no notifications.
Payment data is limited to the amount, timestamp and merchant needed to classify a conversion.

What we would need from Kakao

The minimum scope needed to run the PoC. We are not requesting more than this.

Kakao T: Destination and ETA for trips bound for the named district, plus a trip identifier to attach assignment to
Kakao Pay: Authorisation and cancellation events for participating merchants, settlement integration, and payment timestamps for conversion
Operations: Approval to communicate with district merchants during the run, and agreement to keep the holdout intact

After the PoC

The expansion plan if the eight weeks clear the success criteria.

Months 3–6

All of Pohang

Extend from Yeongildae to Jukdo Market and Pohang Station, and test how pricing behaves across district types.

Months 6–12

Regional hub cities

Expand to Ulsan, Daegu and Gwangju. The smaller the local advertising market, the more an arrival signal is worth.

Beyond 12 months

A pre-arrival commerce layer

Extend the same structure beyond taxis to rail, coach and air arrivals: any trip with a committed arrival time.