The law opened a direct sales channel.
One of six producers walked through it. Six Korean traditional-liquor makers · D2C storefronts and online sales · public-data analysis

01What the financials show, and where they stop

Start with the P&L. FY2025 against FY2024, standalone basis throughout.

Revenue growth clusters between +0.7% and +3.2% across all six — the market was uniformly quiet. Yet the change in operating margin runs from −14.13pp to +7.72pp, a spread of 21.85pp. They sold at the same pace and their economics came apart.

The ones that spent more on advertising lost margin; the ones that spent less gained it. Company A raised ad spend 75.8% and gave up more than 14pp of margin, while Company D cut 55.9% and crossed from loss into profit. That said, the increase in ad spend accounts for only 25.6% of Company A's profit decline. The remaining three quarters is not visible from here.

One caveat. Company D's return to profit is on a standalone basis; consolidated, FY2025 is still an operating loss of ₩0.83bn (prior year −₩2.28bn). The table is standardized on standalone because three of the six do not prepare consolidated statements at all.

02Why the analysis cannot continue on financials alone

Statements tell you what was spent and what was left. Not one line tells you whether the customer that money brought in ever came back. There is a line item for advertising, but no note anywhere records who saw the ad, or how many of them bought a second bottle.

So I changed direction. Repeat purchase rate cannot be measured from public sources. But whether a company is equipped to measure it can be. Whether a storefront exists, whether subscriptions exist, whether membership tiers exist, which tools are installed, whether anyone is being hired to read the output — all of it is public.

I downloaded and parsed the page source of all six sites, and went through every job posting they had.

03First problem: there is no storefront

Start with what makes this category unusual. Alcohol is, as a rule, barred from online sale in Korea. The basis is not a statute but a National Tax Service notice — Article 17 of the Liquor License Act delegates the authority, and Article 6 of the resulting notice on mail-order sale of alcohol prohibits it.

Traditional liquor is carved out. More precisely, what the carve-out opens is not a drink but a person: Article 3 of the same notice permits mail-order sale by the producer who made the traditional liquor, where that producer has approval from its district tax office. Article 4 then lists the permitted channels, which include both an own storefront and open marketplaces.

These six therefore hold a direct-to-consumer right the rest of the drinks industry does not have. So I opened each of their official sites.

Online orders for five of them flow through Naver Smart Store, Kakao Gift and convenience-store pickup apps. Which means the buyer ID, the purchase interval and the cart abandonment all stay inside the platform. What reaches the producer is a settlement statement.

Company B is the one that stings. It runs paid search into its marketplace listing, and yet its own homepage carries no link to that store at all. Traffic bought with its own money pools nowhere it owns.

04Inside the one company that does have a store

Company E is doing this properly. Seven subscription products (weekly through monthly, 10% off across the range, auto-billing at 8am on the dispatch date), cancellation and skip-one-delivery from the account page, and five membership tiers from Bronze up to Diamond. The earn rate climbs from 0.5% to 8% based on the last three months of spend, with tier-specific coupons on top.

Even payment failure is documented — the terms state that where the card limit is exceeded or the payment otherwise fails, the system retries two or three times before cancelling the subscription. There is retry logic behind that sentence.

All of which is unmatched in this cohort. Then I read the membership terms to the end and stopped on one line.

Subscribers are excluded from the tier benefits. The customers who buy most often, and most predictably, sit outside the retention programme. And the tier itself, as the terms spell out, is calculated on spend alone, independent of purchase count. Someone who spends ₩500,000 in a single order and someone who spends ₩500,000 across ten land in the same tier.

Put those two together and it converges: this company's retention design has no frequency axis at all. It owns the single best repeat-purchase asset available — a subscription — and its tier system does not count it.

Meanwhile the guide page a new member reads first still contains lines like these.

Six placeholders from the e-commerce platform's default template, left exactly as shipped. The actual courier and the actual payment methods differ from what the page says. The terms of service and the privacy policy, by contrast, have been filled in with real names — meaning the documents carrying legal exposure got attention, and the documents customers read have not been opened since launch.

05Twenty-one tracking scripts, zero customer tools

I counted the advertising and analytics scripts in the page source across all six. Twenty-one in total: GA4, the Meta pixel, the Kakao pixel, Naver conversion tracking, Criteo, and assorted retargeting tags.

In the same source I looked for CRM, marketing automation, review or chat tools. Zero across all six.

Traffic and conversion are measured diligently, channel by channel. Nothing installed anywhere brings those customers back.

And the twenty-one that are installed are in uneven condition.

• Company A's tag manager container is empty. I pulled the container file and parsed it: zero tags, zero triggers. Installed, and never filled.
• Three companies still load a legacy analytics product that stopped collecting in July 2023. Every page fetches a script that has recorded nothing for three years.
• Company F's site carries a retargeting script built for e-commerce. That site has no e-commerce.
• Company E's storefront loads the same GA4 property twice on one page, and initializes the Meta pixel three times. Double-counted pageviews and conversions are settled at the code level.

That last one does real damage. Sum the platform reports as they come and total ROAS exceeds actual revenue. Reconciling at the order-ID level to strip the duplicates changes the numbers on its own.

06And no one is being hired to read any of it

Finally, hiring. I collected every posting from the six across two recruitment platforms and their own careers pages — 14 open, roughly 388 closed postings still viewable.

Then I downloaded the posting text and counted words.

Not one data role across all six. The platforms' own auto-assigned job tags say the same — one company carries 39 tag types, another 29, another 36, and among them zero relate to data. None of the six maintains an engineering blog.

Only Company D has a careers site of its own, with two postings on it in total.

07How far public sources reach

Visible — the P&L and cost structure; whether a storefront exists and what commerce functions it has; subscription products and payment-failure policy; how membership tiers are calculated; which tracking tools are installed and in what state; whether they double-load; job postings and role tags.

Not visible — actual repeat purchase rate, subscriber counts and retention curves, contribution profit by channel, the incremental effect of advertising. Those need the storefront order ledger and platform settlement data.

So this piece is a hypothesis about where to look, not a measurement. One thing, though, is already answered: whether these companies are equipped to measure repeat purchase. Five of six do not hold the data to do it, and the one that does is counting in a way that leaves frequency out.

08So the answer is to open a storefront?

No. A storefront is a means, not the goal, and both the build and the running of one cost real money. Something has to be measured first.

"Known" here does not mean you have a name or a phone number. It means you can tie two separate orders to the same person. Without that, neither repurchase rate nor lifetime value can be computed at all.

The order gets reversed often. Install a CRM while identification is low and the tool runs perfectly with nobody inside it. Five of the six producers here are close to that state.

09Three choices when raising the rate

• Never count platform settlement records as customers. A settlement line is a transaction, not a person; one buyer purchasing three times appears as three. Treat those as customer counts and repurchase rate comes out structurally low — and the false conclusion "we have no loyal customers" follows.
• Rank identification points by proximity to the purchase, not by cost. The closer an action sits to the moment of buying, the more reliably it joins to the order. A code included in the box always connects better than a survey emailed a month later. This is a design we propose, not a value measured at these six companies.
• Decide on the storefront after measuring, not before. "Identification is high but revenue isn't" and "revenue is fine but nobody is identified" are different problems. For the first one, a storefront is not the answer.

Public sources reveal only whether a point of identification exists. The rate itself requires opening the order ledger. The sequence holds either way — measure before you buy.

10What this piece is really saying

The law opened a door and one producer in six walked through it. But the answer here isn't "open a storefront" — measure your identification rate first. What share of revenue can you tie to a known person across two separate orders? Above 50%, you can quote repurchase rates as numbers. Below 20%, buying a CRM first is the sequence running backwards. Right now five producers' data lives on someone else's servers, and the one that has its own doesn't count frequency.