Friday Night Wiki
TopicsTagsAbout
Buy me a coffee
Marketplace Dynamics Tutorial

Matching: How Supply and Demand Find Each Other

The Matching Problem

Once you have supply and demand, you have to connect them. Matching is the platform's answer to the question "given a buyer's intent, which subset of supply do we show, and in what order?"

This is not the same as listing everything. A marketplace with 100,000 listings is only useful if each buyer sees the 5 to 50 listings most relevant to them. The other 99,950 are noise. The platform's job is to decide which 50.

Matching shapes every other part of the marketplace:

  • Liquidity depends on matches succeeding
  • Pricing is meaningless without a match
  • Trust is the prerequisite for matches converting
  • Network effects accumulate value in matching algorithms over time

Three Matching Paradigms

User types a query; the platform returns ranked results. eBay, Amazon, Airbnb search.

Works when:

  • The buyer has clear intent
  • Supply is long-tail (many distinct items)
  • The buyer is willing to filter and compare

2. Browse / Discovery

User explores a feed or category; the platform suggests options. Etsy's "discover" feed, App Store's editorial picks, Instagram Shop.

Works when:

  • The buyer is inspirational, not goal-directed
  • Supply has aesthetic variety the buyer enjoys exploring
  • The platform has enough data to suggest meaningfully

3. Matching (push)

The platform pushes a match to both sides. Uber assigns a driver. Tinder shows you one profile. Upwork recommends jobs to freelancers.

Works when:

  • Speed matters more than perfect fit
  • The platform has enough signal to make a defensible choice
  • Over-optimising search would be worse than a "good enough" match

Most marketplaces use all three. Uber has search (browse driver types), discovery (ride recommendations), and push matching (actual ride assignment). Airbnb has search primary and discovery secondary.

Ranking

Within any of these paradigms, ranking is where the real decisions happen. What goes on the top of the page, what goes on page 5. Ranking changes who on the supply side makes money.

A reasonable ranking function takes into account:

  • Relevance: does this listing match the query/intent?
  • Quality: is this a good seller? High ratings, fast responses, low cancellations?
  • Price: is this competitively priced? For some users, cheap wins; for others, quality does
  • Freshness: is this listing live and recently updated?
  • Diversity: does this result set give the buyer a useful range of options?
  • Expected outcome: will this match transact? Or will it waste both sides' time?

The last one is key. The platform's incentive is not "show the best match"; it is "show the match most likely to transact". These overlap but aren't identical.

Expected value ranking

Most modern marketplaces use some form of expected-value ranking:

score(listing) = P(click) × P(book | click) × value_to_platform(booking)

You predict click probability, then conditional book probability, then assess the transaction value (take rate times expected price). Rank descending.

This makes matching a machine-learning problem as the marketplace matures. The data network effect kicks in: more transactions make the model better, which improves matching, which produces more transactions.

Fit Quality vs Speed

Every marketplace trades off between matching speed and matching quality. Real-time marketplaces like Uber lean toward speed; you'd rather have a 4.7-star driver in 3 minutes than a 4.9-star driver in 15.

Asynchronous marketplaces lean toward quality. A freelance hiring decision that takes two days can afford to examine twenty candidates carefully.

Where on the spectrum you sit determines:

  • How fast the platform pushes matches
  • Whether users see a list or a single choice
  • How much context the UI shows (full profile vs rating)
  • What happens when no good match exists (fail fast, or wait?)

A common early mistake: pushing a real-time pattern onto an asynchronous marketplace. "Swipe to match your plumber" is less useful than "browse three plumbers and compare".

The Cold Start of Matching

Matching algorithms need data. When you launch, you don't have it.

Two ways around this:

  1. Hand-tuned heuristics with obvious signals: distance, price, ratings from any source, listing age. Enough to be useful. Not enough to be a moat
  2. Explicit preferences from the user. "What are you looking for?" lets the buyer tell you what click-through data would normally teach. Tinder's "interested in men / women / everyone" is a matching-cost-shortcut

Algorithms improve with data. Early-stage marketplace algorithms are heuristic. That is fine. Premature ML is a common waste.

The Quiet Decisions Matching Makes

Ranking is political. Every tiebreaker is a choice about whose economic outcome to prioritise. A few quiet decisions that make or break marketplaces:

  • Do new listings get a boost? Yes, usually. Without it, new supply never gets a chance. With it, existing supply complains about unfair competition
  • Do top listings get locked in? Hard to climb past them? Easy? The answer determines whether the platform's supply side becomes winner-takes-most or stays competitive
  • Do listings with bad outcomes get suppressed? Yes, but over what time window? Bury a seller over one bad review or 50?
  • Does the platform favour its own paid placements? Most do, but the line between "native ad" and "search result" is a trust decision
  • What happens on an empty result set? Fallback to broader criteria? Suggest alternatives? Show nothing?

These decisions are not visible in the product, but they determine the economics of every user's experience.

Matching Under Constraint

Matching gets interesting when it's constrained: not every supply fits every demand. Lodging has dates, location, number of guests, pet policy. Labour has skills, availability, budget. Dating has geography, age, gender preferences.

Constraints make matching harder but also more defensible. The platform that best handles constrained matching owns the vertical. It is also why verticals matter: a general-purpose marketplace for everything has to handle every constraint badly. A narrow marketplace can handle its constraints well.

Anti-Patterns in Matching

  • Ranking by volume only. Sellers with more reviews always win; new sellers never break in. The supply pool calcifies
  • Ranking by price only. Race to the bottom. Quality drops. Platform reputation follows
  • Opaque ranking. Sellers don't know why they rank where they rank, so they game what they can observe. You get SEO-style games. Airbnb's ranking transparency shifts over the years are a case study
  • Matching by geography only. Usually close, but "closest" and "best" are not the same. Uber's early surge pricing made this worse: it pushed the closest driver, not the best one, and rewarded drivers for being in the right spot at the right time rather than being good drivers
  • No matching. A giant search box over a full dataset. Sounds democratic; produces a user experience dominated by whoever games search. eBay suffered from this for years

Good matching requires good supply. If half of your listings are fake, outdated, or low-quality, no ranking function will save the experience. Matching's job is to surface the best in the pool; if the pool is bad, so is the match.

This means matching is partly a supply-side product decision. Do you let anyone list? Do you verify sellers? Do you remove stale listings? Do you deactivate sellers with repeated bad experiences? Every "no" above hurts matching; every "yes" costs supply.

Common Pitfalls

"Better matching will fix our conversion." Matching is upstream of conversion, but if conversion is bad for structural reasons (bad pricing, bad trust, poor supply), matching improvements only lift you a few percent. Look at the whole funnel before investing

"We need an ML matching algorithm." You need one eventually. You probably don't need one this quarter. Simple heuristics and tight loops of "try, measure, adjust" beat premature ML

"We'll let the user figure out the match." A pure search box with no ranking. Some marketplaces survive this (Craigslist) but pay for it in user experience. For most businesses, the platform's opinion is the product

"We're a neutral platform; we don't influence matches." Every ranking is an opinion. Every default is a nudge. Claiming neutrality in public is fine; believing it internally is dangerous

Next Steps

Continue to 05-pricing-and-take-rate.md to learn how marketplaces make money off the transactions they enable.