Latest / Marketplace Businesses with Fexingo: Two-Sided Networks, Liquidity, and Take Rates / How the Best Marketplaces Use Search Relevance to Optimize Discovery
Transcript
- Lucas: So Luna, you've been shopping on Etsy for years. When you search for something there — say, 'vintage brass lamp' — how often do you actually buy the first thing you see? Luna: Almost never. I usually scroll through at least a dozen listings. Half the time the top results feel like they're just keyword-stuffed. Lucas: Right, and that friction is exactly what Etsy's search team spent 2023 trying to fix. They rolled out a new relevance engine that essentially stopped treating search like a keyword-matching game and started treating it like a conversion-prediction problem. Luna: Meaning they rank results based on how likely you are to buy, not just how many times the word 'brass' appears in the title? Lucas: Exactly. Etsy's head of search said in their 2023 investor day that the old system was optimized for 'matching,' but buyers were clicking and bouncing. So they built a model that predicts the probability of a purchase given the query, the listing, and the buyer's historical behavior. And the results were dramatic — gross merchandise volume from search went up 12% in six months. Luna: Twelve percent is huge for a platform that already does over $13 billion in GMV. But doesn't that optimization risk burying new or niche sellers? Lucas: That's the classic trade-off in marketplace search — relevance versus serendipity, or what some economists call 'exploit versus explore.' If you only show people what they're statistically likely to buy, you starve the long tail. But if you show too many obscure items, you frustrate buyers and they leave. Luna: And Etsy's long tail is its whole identity. They have over nine million active sellers, most of them micro-businesses. Lucas: So what they did was blend two signals. The primary ranking is still conversion probability, but they reserve about 20% of search slots for what they call 'discovery inventory' — lower-volume listings that have high relevance but low historical conversion. It's a bandit algorithm approach. Luna: That makes sense. You get the liquidity boost from showing buyers what they'll actually buy, but you also give new sellers a chance to earn their first reviews. Lucas: And this isn't just an Etsy problem. Every two-sided marketplace lives or dies on search relevance. Think about Airbnb — if you search for 'apartment in Paris' and the top result is a shared room in the suburbs, you're closing the app. That lost session is a liquidity failure. Luna: So how do smaller marketplaces handle this without a team of machine learning engineers? They can't all build custom prediction models. Lucas: Most start with a simple rule-based system. You can get surprisingly far by ranking on a weighted combination of three things: text relevance, a seller quality score, and a freshness signal. Text relevance can be done with tf idf or basic embedding similarity. The seller quality score could be their average rating times their fulfillment rate. Freshness just means giving newer listings a temporary boost. Luna: So a newer seller with good ratings still gets a shot even if their listing isn't perfectly optimized? Lucas: Exactly. And then you iterate. The best marketplaces treat search as a continuous experiment. You run A/B tests on ranking changes, measure not just click-through rate but conversion rate and repeat search rate. If buyers are coming back and searching again within the same session, your relevance engine is failing. Luna: I've heard that some marketplaces actually optimize for time to purchase. If someone searches and buys within five minutes, that's a perfect match. Lucas: Right, but that can backfire. If you optimize purely for quick conversion, you incentivize sellers to use manipulative tactics — keyword stuffing, misleading photos, artificially low prices. That erodes trust over time. So you need a counterbalancing signal like post-purchase satisfaction or return rate. Luna: Which is exactly what Amazon does. They've been sued for allegedly prioritizing their own private-label products, but their search algorithm is known to heavily weight things like defect rate and return probability. Lucas: Yeah, Amazon's search is probably the most scrutinized in the world. But even they struggle with the explore-exploit trade-off. Back in 2019, they introduced a 'randomized' slot in search results specifically to surface new products. They found that without it, the long tail was shrinking by about 8% per year. Luna: So the algorithm has to be self-correcting. If you only optimize for current conversion, you kill future conversion by starving your supply pipeline. Lucas: That's the core insight. Search relevance isn't just a UX feature — it's a liquidity management tool. The way you rank results determines which sellers survive and which buyers return. Misaligned search can kill a marketplace faster than a bad fee structure or a clunky checkout flow. Luna: Let's talk about a marketplace that got this wrong. Do you remember the early days of Fab? Lucas: Oh, Fab is a textbook case. They were a design-focused marketplace that grew explosively in 2011 and 2012. But their search was terrible — it was basically just keyword matching with no personalization. Buyers would search for 'minimalist desk lamp' and get results that were completely irrelevant. Conversion rates tanked, and they burned through their venture capital trying to fix it with marketing instead of product. Luna: They ended up selling for pennies on the dollar. And the irony is they had great curation — their editorial team picked beautiful products — but the search engine couldn't surface them. Lucas: Exactly. Curation without discovery is like having a library with no card catalog. So the lesson is: invest in search relevance early, even if your marketplace is small. You don't need a giant AI team — you need a feedback loop. Log every search, every click, every purchase. Use that data to rank better tomorrow than you did today. Luna: And if you're bootstrapped, you can use off-the-shelf solutions like Algolia or Elasticsearch with some basic business rules on top. That's what many early-stage marketplaces do. Lucas: Sure. But the rules matter more than the engine. The most common mistake I see is marketplaces that rank purely by seller-paid promotion — essentially turning search into an auction. That works for short-term revenue but destroys the buyer experience over time, because the highest bidder isn't always the best match. Luna: It's the same tension as advertising versus organic results on Google. If you let money dictate relevance, you lose the user's trust. Lucas: And once trust is gone, liquidity follows. So the best marketplaces keep search relevance as a separate function from monetization. Promoted listings exist, but they're clearly labeled and they don't dominate the results. Luna: Speaking of trust and liquidity — and I don't mean to interrupt the thread — but this is exactly the kind of practical, operator-level insight that keeps this show ad-free. We keep it independent because we believe in sharing what actually works, without sponsor pressure. If these conversations are useful for what you're building or running, and you'd like to support that, you can at buy me a coffee dot com slash fexingo. No pressure at all, just a low-key way to help us keep doing it. Lucas: Yeah, I really appreciate you bringing that up. And it's true — the model works because we don't have to sell ad slots. Anyway, back to search relevance. Let's talk about one more example that I think is particularly instructive — The RealReal. Luna: Oh, the luxury consignment marketplace. Their search must be incredibly hard because every item is unique. You can't just match a brand name; you have to match condition, era, authenticity. Lucas: Exactly. And they've invested heavily in what they call 'visual search.' You can upload a photo of a handbag you like, and their algorithm finds visually similar items. That's a whole different level of relevance — it's not text-based at all. Luna: It also solves the problem of sellers not knowing how to describe their items. A vintage Chanel bag might be listed as 'vintage flap bag' when a buyer might search for 'quilted lambskin crossbody.' The visual search bypasses that vocabulary gap. Lucas: Yes, and that's where machine learning really shines. The RealReal trained a model on hundreds of thousands of images tagged by their in-house authenticators. Now the search engine can identify not just the brand but the specific collection and year. Their conversion rate from visual search is reportedly 30% higher than from text search. Luna: Thirty percent is massive. But that's a huge data investment. How does a smaller luxury marketplace compete? Lucas: They probably can't, not directly. But they can leverage pre-trained models. There are open-source image embedding models you can fine-tune on your own catalog. Even a few thousand tagged images can get you a decent visual search. It's not as good as The RealReal's, but it's better than keyword-only. Luna: So the barrier to good search relevance is lower than it was five years ago, but it still requires intentionality. You can't just set up Elasticsearch and walk away. Lucas: No, you can't. And that's the takeaway for anyone running a marketplace. Search relevance is not a one-time build. It's a continuous investment in data collection, model iteration, and experimentation. The marketplaces that survive are the ones that treat search as a core product, not a feature. Luna: And the ones that fail are the ones that think it's just about putting a search bar on the page. Lucas: Exactly. So if you're building a marketplace today, ask yourself: what does your search engine know about the buyer that it isn't using? And what does it know about the inventory that it's ignoring? Those gaps are where liquidity leaks. Luna: Good place to leave it. Next time we'll look at how marketplaces handle pricing signals — another hidden liquidity lever. Lucas: We'll see you then.