Search and Matching in Marketplaces
Turn this article into takeaways for your work.
Each assistant summarizes the article only for you and suggests best practices for your work.
A marketplace can have plenty of buyers and plenty of sellers and still fail, because the right buyer never meets the right seller. Supply sits unseen on page nine. Demand types a query and gets nothing back. Both sides are present, and no transaction happens.
That gap is the job of search and matching. It's the layer that turns a pile of listings and a crowd of buyers into actual deals. This article covers the three main ways marketplaces do it, the signals that go into ranking, the research behind two-sided matching, what platforms are required to say about their rankings, and the metrics that tell you whether it's working.
We'll stay conceptual. We won't tell you which ranking formula to use, because that depends on your category. We'll give you the vocabulary, the evidence, and the questions to ask.
Three ways to connect buyers to supply
Most marketplaces use one of three approaches, or a mix.
Search and ranking. The buyer states what they want, and the platform returns an ordered list. The buyer does the choosing. Etsy, Airbnb, and Upwork search work this way. The platform's influence is in the order, filters, and what gets surfaced.
Algorithmic matching. The platform picks the counterpart, or a very short list, for the buyer. Ride-hailing is the standard example: the rider doesn't browse drivers, the system assigns one. This works when the goods are interchangeable enough that the buyer cares about speed and price more than about choosing.
Managed matching. People do the matching. An operator reads the request, picks suppliers, and sometimes quotes the job. It costs more per transaction and is common where quality varies a lot or the job is complex. We cover that model in managed marketplaces.
| Approach | Who chooses | Fits when | Main cost |
|---|---|---|---|
| Search and ranking | The buyer, from a ranked list | Goods are differentiated, buyers want control | Ranking quality, buyer effort |
| Algorithmic matching | The platform | Goods are near-interchangeable, speed matters | Needs enough nearby supply, see density |
| Managed matching | An operator | Complex or high-stakes jobs, uneven quality | Human time per transaction |
These aren't exclusive. A marketplace might rank listings for most buyers and route its largest or hardest requests to a human. The useful question is which approach matches how much choice the buyer wants and how different the supply is from one seller to the next.
What goes into a ranking
Ranking is a prediction problem with a business objective attached. The platform estimates which listing a given buyer is most likely to want, and orders results accordingly. The signals usually fall into a few families.
- Relevance. How well the listing matches the query, category, dates, or stated needs.
- Quality. Reviews, completion history, response behavior, and similar proof that the seller delivers.
- Availability. Whether the seller can actually do the job for the dates, quantity, or time requested.
- Distance. For local or delivered goods, how close the supply is to the buyer.
- Price and fit. Whether the price falls in the range the buyer seems to want.
- Commercial factors. Fees, promotions, or paid placement, which we'll return to below.
Airbnb has published more detail on this than most companies. Its 2018 paper, Applying Deep Learning to Airbnb Search, describes the task as responding to a guest with an ordered list of a handful of listings out of thousands available. It says the first search ranking was a manually crafted scoring function, and that replacing it with a gradient boosted decision tree model gave one of the largest step improvements in bookings in the company's history, before gains plateaued. The features feeding its neural network were mostly simple listing properties such as price, amenities, and historical booking count.
Two details in that paper are worth keeping. First, the model is trained on logs where the booked listing is the positive example and the others the negatives, so the ranking learns from what buyers actually chose, not from what a human thought was relevant. Second, the authors say this model is one piece of an ecosystem that also includes models predicting whether a host will accept a request. In a two-sided marketplace, a result the supplier would decline isn't really a good result.
That second point is where marketplace ranking differs from ordinary web search. A search engine optimizes for the searcher. A marketplace has to care about both sides.
Balancing both sides
Airbnb's 2023 paper, Optimizing Airbnb Search Journey with Multi-task Learning, states the problem directly. Guests often spend weeks exploring and comparing before making a request, a reservation request may then be rejected or cancelled by the host, and the model has to balance guest and host preferences. Its answer was to use intermediate guest actions, positive and negative, as milestones that move the guest toward a successful booking.
A different paper from the same group, Learning to Rank Diversely at Airbnb, challenges another common assumption. Traditional ranking estimates each listing's booking probability independently. The authors argue that similarities between listings affect user decisions, and that reducing redundancy among similar results improves diversity, which they validated in production A/B tests.
And Improving Deep Learning for Airbnb Search organizes its work as architecture, bias, and cold start. Position bias (people click higher results partly because they're higher) and cold start (new listings have no history) are problems every marketplace ranking system meets.
These papers describe one company's system. They're evidence of what's possible and what the problems look like, not a template. A marketplace with 500 listings doesn't need a neural ranker. It does need to decide what it's optimizing.
Two-sided matching: the research roots
The idea of matching in a market has a deep academic history that predates the internet.
The 2012 Nobel Prize in economics went to Alvin Roth and Lloyd Shapley, and the Royal Swedish Academy's press release frames their subject as a central economic problem: how to match different agents as well as possible. Its examples are students with schools and organ donors with patients in need of a transplant.
Several ideas from the release matter for marketplaces.
Stability. A matching is stable if no two agents would prefer each other over their current counterparts. The release credits Shapley and colleagues with methods, "in particular, the so-called Gale-Shapley algorithm," that always produce a stable matching. Think of it as a test: if a buyer and a seller are both better off ditching their assigned partners for each other, the match will fall apart, and the system hasn't really solved the problem.
Manipulation. The release says these methods also limit agents' motives for manipulating the matching process. That matters in any market where participants can game what they report.
Which side benefits. It adds that Shapley showed how the specific design of a method may systematically benefit one side of the market over the other. Design choices aren't neutral. Who proposes, who accepts, and who ranks all shift outcomes.
Practice. Roth recognized that this theory could explain real markets. He and colleagues showed empirically that stability is the key to understanding the success of particular market institutions, and he helped redesign institutions for matching new doctors with hospitals, students with schools, and organ donors with patients. The release notes that those reforms were based on the Gale-Shapley algorithm, modified for specific circumstances and ethical restrictions such as the preclusion of side payments.
How much of this transfers to a commercial marketplace? Less than you might hope, honestly. Those centralized matches typically have both sides submit ranked preferences and a clearinghouse resolves them in one go. Most online marketplaces run continuously, preferences aren't fully stated, and prices move. What does transfer is the framing. Matching is a design problem, stability is a useful standard, and the rules of the mechanism shape who wins.
Search frictions: why good matches still get missed
A separate line of research explains why supply and demand coexist unmatched. The 2010 Nobel Prize, as the Academy's press release describes, went to Peter Diamond, Dale Mortensen, and Christopher Pissarides "for their analysis of markets with search frictions."
The release's core observation is plain. On many markets, buyers and sellers don't make contact immediately. The search process requires time and resources, so it creates frictions. On such markets some buyers' demands go unmet while some sellers can't sell as much as they'd like, at the same moment. It notes that the theory applies beyond labor markets, and it cites the housing market as one example.
For a marketplace operator, "friction" is a useful word. Each of these is a friction:
- A buyer can't describe what they want in terms your categories understand.
- Results are relevant but the buyer can't tell which seller is good.
- A seller's calendar or stock is stale, so the result can't be fulfilled.
- A buyer contacts three sellers and two never reply.
Better search and matching reduces friction. It doesn't remove it, and some friction is useful: a buyer who does a little comparison often makes a better choice. The goal is to remove the avoidable kind. Some of that is a trust problem, which is why supply quality controls belong in the same conversation as ranking.
What platforms must say about ranking
Ranking is also a legal subject in some regions. In the European Union, Regulation (EU) 2019/1150 (often called the Platform-to-Business Regulation) sets transparency requirements for online intermediation services and search engines.
The regulation defines "ranking" as the relative prominence given to goods or services offered through online intermediation services, or the relevance given to search results by online search engines. Article 5 then requires:
- Intermediation services to set out in their terms and conditions the main parameters determining ranking and the reasons for the relative importance of those parameters compared with others.
- Search engines to give an easily and publicly available description, in plain and intelligible language, of the main ranking parameters, kept up to date.
- Paid influence to be disclosed. Where ranking can be influenced by payment, the provider must explain that possibility and the effect of the remuneration on ranking.
The regulation also says providers don't have to disclose their algorithms or information that could allow consumers to be deceived by manipulation of search results. So the obligation is about the main parameters and their relative importance, not about publishing the model.
Even if you don't operate in the EU, this is a good template for a marketplace's own ranking explainer. Sellers who understand roughly what drives visibility make better decisions about photos, pricing, and response times. And if paid placement exists, saying so protects trust. We look at how sponsored placement works commercially in marketplace advertising revenue.
Metrics for search and matching
Search quality is easy to feel and hard to measure. These are the metrics teams commonly track. None come with a universal benchmark, so set targets from your own data.
| Metric | Definition | What a problem looks like |
|---|---|---|
| Zero-result rate | Searches with no results divided by total searches | Buyers asking for things you don't have, or filters that are too strict |
| Click-through rate | Share of searches with at least one click | Results look irrelevant or unappealing |
| Search-to-contact | Share of searches that lead to a message, request, or inquiry | Relevant results, but low trust or high effort to reach out |
| Search-to-transaction | Share of searches that end in a booking, order, or hire | Weak match, or fulfillment fails after contact |
| Match rate | Share of buyer requests that end with a supplier accepting | Supply is thin, or suppliers decline too often |
| Time to match | Time from request to supplier acceptance | Too little supply nearby, or slow responses |
The first two have a documented definition. Algolia's analytics documentation defines no results rate as the percentage of searches with no results, and defines click-through rate as the percentage of tracked searches with at least one click event. The other rows are our own working definitions. "Search-to-contact" and "match rate" have no single standard meaning, so write yours down before you compare numbers across teams or time periods.
A few cautions.
- Segment the zero-result rate. A blended number hides the queries that matter. Look at the top zero-result queries: they're a free list of unmet demand and a prompt for supply acquisition.
- Watch both sides. A high match rate with suppliers burning out on low-quality leads is not a win. Pair buyer metrics with supplier acceptance and response data.
- Don't optimize the click alone. Airbnb's 2018 paper trains on bookings, not just clicks, because a click isn't the goal. The closer your objective is to the real outcome, the less you reward curiosity.
- Separate search quality from supply depth. A zero-result query can mean bad search or missing supply. They need different fixes.
Where you place these metrics along the buyer's path matters too. We lay out the stages in the marketplace conversion funnel, and the way price signals shape what buyers choose in marketplace price discovery. The booking model also changes what match means: with instant book versus request to book, the supplier's acceptance step is either removed or made explicit.
A practical order of operations
If you're building or improving matching, this is a sensible sequence. It's reasoning, not a rule from the sources above.
- Start with filters and sorting, not models. Let buyers narrow by the attributes that decide purchases. A small, well-built filter set beats a clever ranker over messy data.
- Fix the data before the algorithm. Ranking can only use what sellers provide. Required fields, consistent categories, and fresh availability do more than model changes at small scale.
- Pick one objective. Decide whether you're optimizing for completed transactions, buyer satisfaction, or revenue, and be aware they can conflict.
- Add seller-side constraints. Include acceptance likelihood, availability, and response time, so you're not showing buyers results that go nowhere.
- Log everything. Impressions, positions, clicks, contacts, outcomes. Airbnb's approach of learning from logged sessions is only possible if you kept the logs.
- Write down the main parameters. Even a short ranking explainer for sellers forces you to know what your ranking actually does.
- Route the hard cases to people. Where a request is complex or high-value, human matching may beat any ranker. See managed marketplaces.
Key Facts: Search and Matching in Marketplaces
- The 2012 Nobel Prize in economics recognized Alvin Roth and Lloyd Shapley, and the Academy's press release says the Gale-Shapley algorithm always ensures a stable matching, meaning no two agents prefer each other over their current counterparts.
- The 2010 Nobel Prize went to Diamond, Mortensen, and Pissarides for analyzing markets with search frictions, where buyers and sellers don't make contact immediately and some demand and supply stay unmatched at the same time.
- Regulation (EU) 2019/1150, Article 5 requires intermediation services to disclose the main parameters determining ranking and the reasons for their relative importance, and to explain the effect of any payment on ranking.
- Airbnb's 2018 search paper says replacing a manual scoring function with a gradient boosted decision tree model gave one of the largest step improvements in bookings in the company's history, before gains plateaued.
- Airbnb's Journey Ranker paper describes guests spending weeks comparing options, hosts potentially rejecting or cancelling requests, and the need to balance guest and host preferences in ranking.
- Algolia's documentation defines the no results rate as the percentage of searches with no results.
Related Reading
Co-Founder & COO