How to Build a Two-Sided Marketplace: Cold Start, Take Rate and Trust Mechanics

Forty-Two Features, Four Transactions in Three Months

A supplier in central Taiwan had spent ten years matching customers with plumbers and plasterers through a single LINE group, roughly 60 jobs a month. The owner decided to turn it into a platform: NT$850,000, five months, 42 features including live chat, calendars, nested categories, two-sided reviews, coupons and push notifications. Three months after launch, 310 tradespeople had registered, but only 41 had complete profiles with photos, price ranges and service areas. The demand side generated 26 inquiries a month and four closed jobs. The money was not spent on the wrong things; it was spent in the wrong order. The hard part of a marketplace was never the feature list. It is deciding which side to make dense first.

Who Should Build This, and Who Should Wait

Good fit

• You already match manually, through a LINE group or a Facebook group, at a steady 30-plus jobs a month. You are already the intermediary; you just have no system.
• Transaction size sits between NT$3,000 and NT$200,000. Below that a take rate cannot fund the platform; above it both sides have a strong incentive to route around you.
• Supply can be filtered on hard attributes: licences, service area, availability windows, vehicle model, floor area, capacity.
• You already hold one side of the market, at least 100 suppliers you can call and list.
• The transaction repeats. The same buyer comes back at least twice a year.

Wait if

• Both sides start from zero and you have no budget for three to six months of subsidy or field recruiting.
• Unit price is under NT$500 with no repeat purchase; the take rate will not even cover payment fees.
• The transaction is inherently one-and-done once the parties meet, and you cannot name a post-transaction value you provide.
• The supply is your own staff or your own inventory. That is an ecommerce or booking system, not a marketplace.
• You plan to monetise through listing fees but have no demand-side traffic yet.

Four Paths, With Costs and Trade-offs

The pragmatic route is staged: run the group-plus-form version for 60 to 90 days to prove matching volume, then jump to SaaS or a custom build. What that saves is usually more than one development fee.

Four Phases From Requirements to Launch

• Phase 1: single-side audit and matching economics (5–8 working days). Nail down three numbers: how many staff minutes each match costs today, the maximum matching fee a buyer will pay, and the maximum take rate a supplier will concede. Deliverables: unit-economics model, a listable supplier roster, an MVP scope ranked with MoSCoW. Tools: Google Sheets, Notion.
• Phase 2: trust mechanics and state machine (8–12 days). Draw inquiry to quote to match to fulfilment to review to payout as a state machine, defining who can act in each state and where a dispute freezes the flow. Design identity-verification tiers and double-blind review rules alongside it. Deliverables: high-fidelity Figma, state diagram, dispute SOP. Tools: Figma, FigJam.
• Phase 3: matching engine and payments (30–45 days). Build filtered search and ranking, in-platform messaging, notifications, split payouts, and an admin arbitration console. Deliverables: a platform that can run a real transaction, reconciliation reports. Tools: Laravel, Meilisearch or Algolia, ECPay, Stripe Connect, Cloudflare R2, LINE Messaging API.
• Phase 4: seed supply and closed beta (15–20 days). Lock one city and one category, push 30 real transactions through by hand, and fix the flow as you go. Deliverables: support SOP, GA4 funnel events, Sentry and UptimeRobot monitoring.

Real Costs: Three Tiers and the Hidden Line Items

• Entry, NT$350,000–550,000. One category, with listings, vetting, filtered search, inquiries and messaging, payment handled offline. The platform does not touch funds.
• Mid, NT$700,000–1,200,000. Adds split payouts, escrow-style holds, double-blind reviews, a search engine, admin arbitration and reconciliation reports.
• Advanced, NT$1,500,000–2,800,000. Multi-category, ranking and recommendation logic, apps for both sides, automated reconciliation and invoicing, a risk-rules engine, and a supplier operations back office.

The line items most often missed

• Payments. Per the ECPay rate card, standard sellers pay 2.75% on domestic credit cards plus NT$1 per order in processing, with 5% business tax on settlement; contracted sellers can negotiate 1.85%–2.75% but pay a one-off NT$5,000 setup fee and NT$13,000 a year and up for the payment service. For cross-border splits, check Stripe Connect pricing.
• Compliance. If the platform collects buyer funds and then pays the seller, that is agency collection. Under Article 6, Paragraph 3 of the Money Laundering Control Act, you must complete third-party payment AML and service capacity registration with the Ministry of Digital Affairs. Advisory and documentation runs NT$50,000–150,000.
• Search. Algolia Grow includes the first 10,000 search requests a month, then charges US$0.50 per 1,000. Self-hosting Meilisearch on a 4GB VPS costs about NT$600–1,200 a month.
• Verification. SMS OTP at NT$0.7–1.2 per message; manual licence and identity checks at 3–5 minutes each, which is about 20 hours for 300 suppliers.
• Image storage and transcoding: NT$500–3,000 a month.
• Support and dispute arbitration: half to one full-time person for the first six months. This is the line most often ignored entirely.
• Annual maintenance at 15–20% of build cost.

What Clients Expect vs What Actually Happens

• Expectation: open registration on both sides and the market grows itself. Reality: open both and both stay thin. A buyer who lands on three results the first time does not come back for a second.
• Expectation: 15% is a reasonable take rate. Reality: labour-matching in Taiwan lands mostly at 8–20%, and the first year often needs a flat matching fee or no take rate at all to buy supply density. Charge too early and suppliers walk your customers out the door.
• Expectation: the matching algorithm is the hard part. Reality: manual matching plus rule-based ranking is enough for the first 500 transactions. The hard part is who decides when something goes wrong: refunds, hold periods, burden of proof.
• Expectation: five stars and you are done. Reality: without double-blind reviews, released only after both sides submit, you get retaliation and collective inflation. A site-wide average of 4.9 carries no information.
• Expectation: suppliers will maintain their own profiles. Reality: for the first three months someone has to phone them one by one for photos and prices, or six out of ten profiles stay empty.

Seven Traps and How to Avoid Them

• Cold-starting both sides at once. Fix: pick one side to subsidise and make it dense within a single city and category. In practice 80 to 120 available suppliers in one area is the threshold before you open the other side.
• Pretending leakage is zero. Fix: attach the value to what happens after the match, through escrow holds, invoicing, arbitration and warranty. Masking phone numbers and LINE IDs only delays it. Healthy platforms still leak 20–40%, so build that into the unit economics.
• Treating split payouts as a pure engineering problem. Fix: decide early whether you will register as a third-party payment provider. If you do not want the compliance load, use ECPay or Stripe Connect split payouts and keep funds on a licensed institution balance sheet.
• Filter pages eating your crawl budget. Fix: follow Google guidance on managing crawling of faceted navigation URLs: separate parameters with &, return 404 for combinations with no results, canonicalise back to the main category, and block non-indexable combinations in robots.txt.
• Slow listing pages. Fix: the hero image on a listing page is almost always the LCP element. Serve it from a CDN as AVIF or WebP with fetchpriority="high" and keep LCP under 2.5 seconds. Marketplace bounces happen on the first screen.
• Fake reviews and fake accounts. Fix: phone binding, reviews only after a completed transaction, double-blind release, plus licence upload and manual review on the supply side.
• Launching without an arbitration SOP. Fix: before launch, fix the refund policy, payout timing such as seven days after completion is confirmed, burden of proof, and support response times, then publish them on a public rules page.

Success Metrics and the 90-Day Roadmap

• Day 30: supply density. At least 100 available suppliers in one city and category, profile completeness above 80% covering photos, price range and service area, and a search-with-results rate above 90%. Do not look at revenue in this phase.
• Day 60: the matching funnel. Target a 24-hour response rate above 70% and a quote-to-close rate above 25%. Response rate is the killer metric here, and push reminders plus automated nudges usually lift it 15 to 20 points.
• Day 90: unit economics and retention. Compute real gross margin per transaction, which is take rate minus payment fees minus support hours. Interview 30 suppliers to estimate leakage. Track 30-day buyer return rate against a 15% target. Only when all three look healthy should you raise the take rate or open a second category.

Decision Checklist

• ☐ Are you already matching manually or via LINE at 30-plus jobs a month?
• ☐ Do you hold at least 100 listable suppliers on one side?
• ☐ Is the transaction size above NT$3,000?
• ☐ Does the same buyer return at least twice a year?
• ☐ Can you calculate the labour cost of one match today?
• ☐ Can you live with a take rate under 10% in year one?
• ☐ Do you have three to six months of supply-side subsidy or field-recruiting budget?
• ☐ Does the platform need to collect buyer funds and pay sellers?
• ☐ Have you determined whether third-party payment registration applies?
• ☐ Is someone accountable for arbitration and support, at least half a headcount?
• ☐ Are the refund policy and payout timing already fixed?
• ☐ Can you define precisely what a qualified supplier is?
• ☐ Will you accept manual matching for the first 500 transactions?
• ☐ Do you have a way to measure leakage?

Nine or more checked and the mid-tier budget usually pays back. Fewer than six, run the group-plus-form version for three months first.

FAQ

What take rate should I charge?

Labour-matching platforms in Taiwan mostly land between 8% and 20%. The formula matters more than the benchmark: your take rate must exceed payment costs of roughly 3%, plus support and dispute cost per transaction, plus amortised acquisition cost. Start year one with a flat matching fee of NT$200–500 per closed job to test willingness to pay, then move to a percentage once suppliers have real reasons to stay: steady lead flow, payment protection, reconciliation reports.

How do I stop suppliers from going around the platform?

Accept that leakage exists, then do three things: make paying on-platform more convenient than paying privately with invoicing, instalments and convenience-store codes; make dispute handling and escrow the reason buyers insist on the platform; and offer tiered take-rate discounts to high-frequency suppliers. Technical blocking of phone numbers and keywords only delays leakage and hurts the experience.

Do I have to handle the money myself?

Not necessarily. There are three models. Pure referral never touches funds, launches fastest, but cannot guarantee fulfilment. Gateway collection uses ECPay or Stripe Connect split payouts, leaving funds with a licensed institution while the platform takes its share, which is the practical answer for most small and mid-sized Taiwanese platforms. A self-built wallet handles stored value and balances, carries the highest bar, and may fall under the Act Governing Electronic Payment Institutions. Start with one of the first two.

Should I build an app on day one?

Suppliers such as tradespeople, landlords and dealers need instant notifications and on-site photo upload, but a PWA plus LINE push covers that. Buyers almost all arrive from search, so a solid responsive site does far more for SEO than an app. Revisit native apps once monthly closed transactions pass 300.

How long until I know whether the platform will survive?

Ninety days. Watch three numbers only: search-with-results rate, 24-hour response rate, and 30-day buyer return rate. If all three hit target but revenue does not, that is a pricing problem and it is fixable. If supply density never forms, the premise itself is wrong and you should cut losses.

Next Step

ScriptWalker builds marketplace platforms starting at NT$350,000, including the unit-economics model, a MoSCoW-ranked scope, and an MVP that can run real transactions. If you already run a LINE group that matches buyers and sellers every day, send us the last three months of matching records and we will return a free cold-start feasibility assessment showing which side to densify first and where your take rate should start.

• Email: [email protected]
• Phone: 0916-224-047
• LINE: @ufv9089p