Latest / The CTO Podcast with Fexingo: Technical Leadership, Architecture, and Engineering Org / How Airbnb Rebuilt Its Search for 100 Million Listings
Transcript
- Lucas: So Airbnb now lists over a hundred million properties worldwide, and every one of them has to be ranked, filtered, and served to a traveler in under half a second. Luna: Half a second? That feels impossibly fast when you consider all the variables — location, price, amenities, reviews, availability. Lucas: It is a beast of a problem. And it's not just about speed — it's about relevance. How do you define the 'best' match for someone who's looking for a beach house in Tulum while someone else wants a studio in downtown Tokyo? They're searching on the same platform but have entirely different intent. Luna: And that's where their machine learning ranking comes in. But I've read that they completely rebuilt this system recently. What was wrong with the old one? Lucas: The old system was essentially a single model that tried to predict one thing: the probability a user would click on a listing. It worked, but it had blind spots. It didn't understand travel context — like how a week-long trip should differ from a weekend getaway. And it treated every listing equally, which meant brand-new properties had almost no chance of surfacing. Luna: That's the cold-start problem, right? A host lists their apartment, and it's just invisible because there's no click data yet. Lucas: Exactly. And that's a real problem for Airbnb because inventory is their moat. If new listings don't get any traction, hosts leave, and the whole marketplace thins out. Luna: So what did they do? I'm guessing they didn't just tweak the old model. Lucas: No, they went modular. Instead of one monolithic ranking model, they split it into multiple specialized models, each handling a different part of the ranking problem. Think of it like a team of specialists rather than one generalist. One model focuses on the user's immediate intent, another on long-term preferences, and another on the quality of the listing itself. Luna: That's a big architectural shift. But how do they ensure these models don't pull in opposite directions? I mean, what if the quality model loves a listing but the intent model thinks it's not relevant? Lucas: That's the key trade-off. Airbnb's team spent a lot of time defining a set of objectives that balance user satisfaction with host success. They use a framework where each model outputs a score, and then a final blending stage decides the overall ranking. It's not a black box — they can inspect why a listing ranks where it does, which is crucial for debugging and for helping hosts understand how to improve. Luna: That transparency is smart. But I've also heard about them using something called a two-tower neural network. Can you break that down for me? Lucas: Sure. A two-tower model is a neural network with two separate branches — one that encodes the user and their search context, and another that encodes the listing. The outputs are then combined to produce a similarity score. This is great for personalization because you can precompute the listing embeddings and then in real time just compute the user embedding and do a quick dot product. It's fast and scalable. Luna: So the two towers are like, one is the user's taste profile and the other is a listing's feature profile, and they match them up. That's elegant. But how do they handle the cold start now? The listing embedding has to be learned somehow. Lucas: Right. They use a hybrid approach. For new listings, they rely on feature-based embeddings derived from the listing's attributes — location, price, amenities, photos. So even with zero interaction data, the model can make a reasonable guess. Then as the listing gets more bookings and views, they gradually shift to learned embeddings that incorporate user behavior. Luna: That's a smart bootstrap. And it must help hosts feel like they have a fair shot at getting discovered. Lucas: It does. But there's another piece: experimentation. Airbnb runs thousands of A/B tests every year, and they've had to build a robust platform to make that safe and fast. The old approach of 'let's try this in production and see' is just not viable when you're ranking for a hundred million listings. Luna: And that's where a lot of engineering orgs struggle — the discipline to test properly. It's a cultural thing as much as a technical one. Lucas: Absolutely. They treat experimentation as a core part of the engineering culture. Every change to ranking goes through a rigorous evaluation, not just on click-through rate but on long-term metrics like booking value and guest satisfaction. It's a lesson for any team building recommendation systems: the offline metrics can lie to you. Luna: That's a great point. If you only optimize for clicks, you might end up with clickbait listings that don't actually convert to bookings. The real goal is the booking, not the click. Lucas: Exactly. And that's why they look at something they call 'guest satisfaction' — which includes things like the accuracy of the listing description and the overall experience. It's a more holistic measure. Luna: You know, we talk a lot on this show about architecture choices, but the cultural piece is often what makes or breaks these systems. Do you think Airbnb's approach is replicable for smaller companies? Lucas: I think the principles are, yes. You don't need a hundred million listings to benefit from a modular ranking system. Even a small marketplace can start with a simple heuristic and then gradually introduce machine learning. The key is to design your system so it can evolve. That's what Airbnb did — they didn't boil the ocean overnight. Luna: And they didn't do it alone. They built a strong data science and ML engineering team. That's a lesson in hiring too, right? Lucas: Totally. They invested heavily in that talent. And speaking of investment, that's something that applies to us as a show too. If these conversations are useful for what you're building or running, and you want to keep them ad-free, a small contribution to buy me a coffee dot com slash fexingo goes a long way. It's what keeps the podcast independent and lets us keep diving into these topics. Luna: Yeah, every little bit helps. And it's a nice way to say thanks if you've gotten value out of episodes like this one. No pressure, just appreciated. Lucas: Right. Now, back to Airbnb. One of the most interesting pieces to me is how they handle the real-time aspects. When a user changes their search dates or adds a filter, the ranking has to update instantly. That's not trivial when you're dealing with millions of potential matches. Luna: So how do they achieve that sub-half-second response? Is it all precomputed embeddings and caching? Lucas: Partly. They use a combination of caching, fast approximate nearest neighbor search, and efficient serving infrastructure. But the real magic is in the architecture — they separate the retrieval stage from the ranking stage. Retrieval quickly picks out a few thousand likely candidates from the full inventory, and then the ranking model scores those in detail. That's a classic funnel approach. Luna: So the heavy lifting is done on a small set, not the whole hundred million. That makes sense. Lucas: Exactly. And they've built a feature store so that all the real-time signals — like current location, device type, time of day — are available to the models without having to recompute them. It's a lot of moving parts, but when it works, it feels seamless. Luna: I also read that they experimented with a 'guest favorite' badge. How does that fit into the ranking? Lucas: That badge is a great example of how they incorporate qualitative signals into the model. 'Guest favorites' are listings that have consistently high ratings and reviews. They baked that into the ranking as a feature, so those listings get a boost. It's a way to surface quality without relying solely on the cold hard numbers. Luna: It's almost like a trust signal. And I bet it helps new users who are overwhelmed by choice. Lucas: For sure. The whole system is designed to reduce friction. The better the ranking, the less time a user spends searching, and the more likely they are to book. That's a win for Airbnb, for hosts, and for guests. Luna: And it's a continuous improvement loop. They're always tweaking and testing. What's the next big challenge do you think? Lucas: I think the next frontier is incorporating more personalization across the entire trip — not just the search, but the recommendations for things to do, restaurants, experiences. And also making it work across multiple devices. But those are stories for another day. Luna: True. For now, I'm just impressed by how they've managed to scale without losing the human touch. That's a fine line. Lucas: It is. And it's a good reminder that even the most technical problems have a human dimension. The ranking algorithm isn't just about math — it's about connecting people with places that make them happy. Luna: Well said. And on that note, if you're a CTO or an engineering leader, what's the one takeaway you'd want listeners to remember? Lucas: I'd say this: don't let your ranking system become a black box. The more you can explain why a result appears, the more you can debug, improve, and build trust with your users and your team. Airbnb's modular approach gives them that visibility, and it's a principle any tech org can adopt.