Latest / Marketplace Businesses with Fexingo: Two-Sided Networks, Liquidity, and Take Rates / How Marketplaces Use Data Network Effects to Raise Take Rates
Transcript
- Lucas: Let me ask you something. Have you ever noticed that Uber's take rate — the percentage it keeps from each ride — has crept up over the years, and yet drivers haven't fled en masse to Lyft? Luna: I mean, they complain about it constantly. But yeah, they mostly stay. What's the trick? Lucas: It's not a trick. It's something called data network effects. The idea is that every transaction makes the platform smarter — better at matching, better at predicting demand, better at routing. And that intelligence becomes a moat. Luna: So the platform can charge more because it's actually providing a better service over time, not just the same service with a bigger fee? Lucas: Exactly. And this is the angle I want to drill into today: data network effects as the hidden justification for take rate increases. Most marketplace operators talk about take rate as a simple percentage, but the smart ones frame it as the price of access to an improving system. Luna: It reminds me of how Amazon Web Services used to cut prices every year because they got more efficient. But in a marketplace, the platform often raises take rates as it gets better. Lucas: Right. Because in a marketplace, the value to each side grows with data. Let's use Uber as our first case. In 2015, Uber's take rate in the US was around 20 percent. Today it's closer to 27 percent in many markets. But Uber's estimated time of arrival predictions are more accurate, surge pricing is smarter, and driver routing reduces deadhead miles. Luna: So drivers see more trips per hour, even if they're giving Uber a bigger cut. Their net hourly earnings might be flat or even slightly up compared to a platform with a lower take rate but worse matching. Lucas: That's the calculus. And it only works if the data advantage is real. If Uber's matching were identical to Lyft's, drivers would switch the moment the take rate diverged. But because Uber has more trips, more data points, better ETAs, the switching cost is higher than just the fee difference. Luna: Airbnb does something similar with its service fee. Hosts pay around 3 percent, guests pay up to 14 percent. That's a hefty total take rate. But Airbnb's search ranking, its review system, its fraud detection — all improve with every booking. Lucas: And they've been able to maintain that structure even as competitors like Vinted and VRBO offer lower fees. The reason is that a host on Airbnb gets access to a massive demand pool that is better sorted, better insured, and better reviewed. The platform's data about guest behavior reduces the risk of bad bookings. Luna: There's a paper by a couple of economists — I think it was from 2022 — that showed that Airbnb's review system actually substitutes for host experience. New hosts with good review scores from guests get bookings faster than experienced hosts with mediocre scores. Lucas: That's a pure data network effect. Every review trains the algorithm to predict host reliability. And that prediction makes the marketplace more liquid — guests book more confidently, hosts accept more requests. The platform can charge for that confidence. Luna: But here's the danger: if you raise take rates too fast, you break the trust loop. Etsy found that out in 2019 when it raised transaction fees from 3.5 percent to 5 percent. Sellers revolted. Some left. Lucas: Etsy's situation is interesting because its data network effects are less about matching and more about discovery. Etsy's search algorithm personalizes results based on past purchases across the entire platform. The more people buy, the better the algorithm becomes at showing relevant products. That's a data moat. Luna: Right, but sellers didn't see that value clearly. They saw a 42 percent increase in fees overnight. Etsy had to spend months explaining that the fee increase funded better marketing and search improvements. Lucas: And that's the communication gap. Marketplace operators often assume users understand the data feedback loop. They don't. So the take rate increase feels like rent extraction, not value sharing. Uber learned to message take rate increases as 'service fee adjustments' tied to specific improvements. Luna: Do you think there's an upper bound? Like, is there a point where even great data network effects can't justify a higher take rate? Lucas: Absolutely. The theoretical limit is the point at which the platform's contribution to the transaction is less than the fee. If Uber's take rate hit 40 percent, drivers would find private ride-hailing apps or form cooperatives. The data advantage has to keep growing faster than the fee. Luna: So it's a treadmill. You have to keep investing in data infrastructure to sustain your take rate. Lucas: Exactly. And that's why marketplaces with weak data network effects — like generic classifieds — have very low take rates. They can't differentiate. The moment they raise fees, users go to the next site. Luna: Let's talk about a marketplace that built data network effects from scratch: OpenTable. Every reservation, every cancellation, every no-show trains their algorithm. Restaurants get better fill rates, diners get better availability. OpenTable's take rate is around 2.50 per diner, but it's effectively a tax on the value of predictive data. Lucas: And they've raised it over time, quietly. In 2015, the per-diner fee was about 1 dollar. Now it's closer to 2.50. Restaurants grumble, but they stay because OpenTable drives enough incremental business that the net is positive. Luna: What about marketplaces where data network effects are weaker, like Thumbtack? You're matching a plumber to a homeowner. That's a one-time transaction. There's less repeat data to learn from. Lucas: Thumbtack's model is different. They charge pros per lead, not per completed job. Their data network effect comes from predicting which leads are likely to convert. The more data they have on homeowner behavior, the better they can score leads. But it's a harder sell because the pro pays upfront. Luna: So the take rate is more visible and more painful. That might explain why Thumbtack has struggled with profitability compared to Uber or Airbnb. Lucas: Precisely. The transparency of the fee matters. When the fee is embedded in a transaction — like Airbnb's guest fee — it's less salient. When it's an upfront charge — like Thumbtack's lead fee — it's more scrutinized. Data network effects help, but they don't override pricing psychology. Luna: I want to bring this back to something practical for marketplace operators listening. How do you know if your data network effects are strong enough to support a take rate increase? Lucas: You look at two metrics: switching cost and value-added. Switching cost is the effort a user would have to make to replicate their current experience elsewhere. If they'd lose their review history, their personalized recommendations, their trusted network — that's a moat. Value-added is the improvement in outcomes: are sellers selling faster? Are buyers finding what they want more easily? Luna: And if both are high, you have room. If not, you're just extracting. Lucas: Right. The worst thing a marketplace can do is raise take rates without contemporaneously improving the data product. That destroys trust and liquidity. We saw that with eBay in the early 2010s when it raised fees without noticeable improvements. Sellers started multi-listing on Amazon. Luna: eBay actually had to reverse some fee increases. They learned the hard way that data network effects degrade if you don't reinvest in them. Lucas: One final thought: the best marketplaces treat take rate as a dynamic number. They experiment with different levels for different segments. Uber does it — new cities get lower take rates to build liquidity. Mature cities get higher take rates because the data is richer. Luna: So the take rate is not a universal constant. It's a function of data maturity. Lucas: Exactly. And that's the framework I hope listeners walk away with. Don't ask 'what's the industry take rate?' Ask 'how much value does our data generate per transaction?' Luna: If today was worth a coffee to you, that's the link — buy me a coffee dot com slash fexingo. It keeps the show ad-free and independent. Lucas: Yeah, listener support is what lets us spend time on deep dives like this. No pressure, just if it's useful. Luna: Alright, back to it. So let's talk about a marketplace that got this exactly right: Poshmark. Lucas: Poshmark's take rate is 20 percent for items under 15 dollars, and it's flat for higher-priced items. That's high relative to other resale apps. But they've built a social layer — parties, follows, shares — that generates data on style preferences. Luna: That social data feeds the algorithm, making recommendations better. So sellers get more visibility, buyers find things they love. The 20 percent starts to feel like a marketing fee. Lucas: And Poshmark has been able to maintain that take rate while growing. Their gross merchandise value hit 2 billion in 2023. The data network effect is real. But they also face competition from Depop and Vinted, which have lower rates. The question is whether Poshmark's data advantage is enough to keep sellers loyal. Luna: I think the answer will depend on whether Poshmark continues to invest in its data infrastructure. If the recommendation engine plateaus, sellers will start multi-homing. Lucas: Multi-homing is the enemy of data network effects. If every seller lists on three platforms, none of them gets a complete picture. The platform with the most exclusive, high-quality data wins. That's why Uber and Airbnb are so focused on loyalty programs and exclusive inventory. Luna: So the ultimate moat is not the network size — it's the data density per user. Lucas: Exactly. And that's a subtle but crucial distinction. A million users with ten transactions each is more data-powerful than ten million users with one transaction each. Marketplaces that drive repeat usage build stronger data network effects. Luna: We should probably wrap, but I want to leave listeners with a concrete question: if you run a marketplace, can you prove that your take rate is tied to data value? If not, you're vulnerable. Lucas: Good note to end on. Thanks, Luna. Luna: Thanks, Lucas. See you next time.