Latest / Marketplace Businesses with Fexingo: Two-Sided Networks, Liquidity, and Take Rates / How Marketplaces Use Pricing Experiments to Discover the Right Take Rate
Transcript
- Lucas: So you're running a two-sided marketplace. You've got supply and demand, you've got some liquidity going. And you're staring at the single most important number in your business model: the take rate. Luna: The fee you keep from every transaction. Which, if you get it wrong, can kill your marketplace overnight. Lucas: Exactly. Too high and suppliers leave, buyers balk, liquidity dries up. Too low and you leave money on the table — or worse, you can't cover the costs of trust and safety, payment processing, customer support. The question is: how do you find the right number? Luna: And that's where pricing experiments come in. But it's not as simple as A/B testing a button color. Lucas: Not even close. Because on a marketplace, every transaction involves two sides. Change the fee on one side and the other side's behavior shifts too. You're not just testing price elasticity — you're testing network elasticity. Let's start with a concrete example. In 2020, Airbnb ran a quiet experiment on its host-facing fee structure. Historically, they charged hosts roughly 3 percent and guests somewhere between 6 and 12 percent, depending on the booking total. But they wanted to understand: what happens if we move more of the fee onto the host? Luna: Right, because the host is the one setting the price. If you increase their cost, they might raise prices, which affects demand. Lucas: Bingo. So Airbnb tested shifting the host fee from 3 percent to something closer to 8 or 10 percent, while lowering the guest service fee proportionally. The hypothesis was that hosts would absorb part of the increase because they value the booking volume, and guests would see lower upfront prices and book more. But the actual result was more nuanced. In some markets, hosts did raise prices — and demand dropped. In others, hosts kept prices flat and the marketplace actually grew because the total price to guests was more competitive. Luna: So the same experiment gave opposite results depending on the micro-market. That's the nightmare of marketplace experimentation — you need segment-level statistical power. Lucas: And most companies don't have it. Which is why a lot of platforms end up using a simpler heuristic: look at comparable marketplaces and pick a number in the middle. But that's a dangerous shortcut, because your network's specific liquidity profile matters a lot. A marketplace with concentrated supply — like a few big sellers — can sustain a higher take rate because those sellers have fewer alternative channels. A fragmented marketplace with lots of small suppliers competing on price — think Etsy sellers — is much more sensitive to fee increases. Luna: Etsy is a great example. They've raised their transaction fee twice in the last few years: from 3.5 percent to 5 percent in 2018, and then from 5 percent to 6.5 percent in 2022. Each time sellers protested loudly. But Etsy's argument was that the extra revenue funded marketing, seller tools, and buyer trust features — and overall gross merchandise value kept growing. Lucas: Right, and that's the thing. A take rate increase isn't just a price hike — it's a value proposition shift. If you reinvest the incremental revenue into things that improve the marketplace experience, the fee increase can be positive for everyone. But the sequencing matters. Etsy didn't just raise fees and hope for the best. They first invested heavily in search relevance, promoted listings, and buyer guarantees. By the time they raised the fee, sellers could point to tangible benefits. Luna: So the experiment isn't just 'what fee do we charge?' It's 'what fee, combined with what investment, maximizes long-term GMV?' Lucas: Exactly. And the best way to answer that is with a structured experiment. There's a playbook emerging from the data teams at companies like Uber, DoorDash, and Airbnb. The first step is to define your primary metric — usually something like gross transaction value per user over a 90-day window, or maybe supplier retention rate. Then you pick a small, isolated segment — say, one metro area or one category — and you vary the take rate by a small amount, like 50 basis points. Luna: Fifty basis points is tiny. At Uber's scale, that's millions of dollars. But it's small enough that you can catch the effect before it damages the network. Lucas: And you run it for long enough to see the full dynamic. A two-week experiment might show no change because buyers don't notice a half-percent price difference. But suppliers — drivers, hosts, sellers — they notice. Their margins are thin. So you need to watch for supply-side responses: do they reduce availability? Do they shift to competing platforms? Do they raise prices to compensate? Those effects can take weeks or months to fully materialize. Luna: I want to dig into a specific case I came across. A mid-sized home-services platform — think Thumbtack or TaskRabbit, but local to the UK — ran an experiment where they lowered their take rate from 15 percent to 12 percent for a segment of handymen. They hypothesized that lower fees would attract more suppliers, increase competition, and drive prices down for consumers — ultimately growing the total number of transactions. Lucas: And did it work? Luna: Partially. Supplier sign-ups increased by 22 percent in that segment. But the number of completed transactions only grew by 3 percent. What happened was, the new suppliers weren't getting booked — because the existing suppliers already had relationships with repeat customers. So the lower take rate mostly just reduced revenue on existing transactions without generating much new volume. Lucas: That's a classic 'experiment cannibalization' trap. You changed the fee for new suppliers, but the demand side didn't see a different price — because the platform didn't lower the price to consumers. The lower take rate didn't translate into lower buyer prices, so there was no demand response. If they had also lowered the buyer-facing price, the result might have been very different. Luna: And that gets to the heart of marketplace pricing experiments. You have to decide which side gets the fee change, and whether you pass any of it through to the other side. It's a joint optimization problem. Lucas: One of the cleanest examples of this is Uber's early pricing experiments. They famously tested 'surge pricing' as a way to balance supply and demand in real time. But they also ran experiments on the base take rate. In 2015, Uber tested reducing the driver commission from 25 percent to 20 percent in certain cities. The result? Driver hours increased, wait times dropped, and overall trips grew enough that Uber's absolute revenue actually went up despite the lower per-trip fee. Luna: So a lower take rate can be revenue-accretive if it unlocks enough volume. That's the dream scenario for any marketplace. Lucas: But you can't just copy Uber's number. Their network is high-frequency, low-value, with immediate matching. A different marketplace — say, a platform for freelance graphic designers — has much lower transaction frequency and longer decision times. The take rate elasticity is completely different. For a high-frequency marketplace, a small fee change compounds over thousands of transactions per user per year. For a low-frequency marketplace, the fee might only be salient once or twice a year, so you can charge a higher percentage without scaring users away. Luna: That's why the best experiments start with a hypothesis about buyer and supplier behavior, not just a spreadsheet optimization. You need to understand the psychology of your users. Lucas: And that's actually a perfect segue into something I want to mention. If you're building or running a marketplace, or just obsessed with how these networks work, these conversations are exactly what we try to explore here. And they take time to produce — research, data analysis, interviews. A couple of dollars a month is genuinely what keeps these going. If you've gotten something out of them, buy me a coffee dot com slash fexingo really does make a difference. No pressure at all, but it's how we keep this ad-free and focused on the mechanics. Luna: Yeah, it's a small thing that adds up. And it keeps us independent. So thank you to anyone who's ever contributed. Lucas: Alright, back to pricing experiments. So we've established that you need to test on the right side, with the right metric, for long enough. But there's another layer: the structure of the take rate itself. Flat percentage versus tiered versus fixed fee. Luna: Right, because not all take rates are created equal. A flat 10 percent fee is simple, but it might overcharge on small transactions and undercharge on large ones. Lucas: Exactly. Some marketplaces use a tiered structure. For example, a freelance platform might charge 20 percent on the first $500 of a project and 5 percent on anything above that. That's designed to capture more from small projects where the platform's value — discovery, payment processing, dispute resolution — is proportionally higher, and less from large projects where the supplier might be tempted to go direct. Luna: Fiverr does something similar. They charge 20 percent on all transactions, but they cap the fee at $50 for orders over $250. So for a $1,000 project, the take rate drops to 5 percent. That's an implicit tier. Lucas: Which is smart because it encourages high-value transactions without losing the supplier to off-platform work. Another structure is the fixed fee plus percentage — like a $1 booking fee plus 5 percent. This is common in travel and ticketing. The fixed fee covers the fixed cost of processing the transaction, and the percentage aligns with the platform's value as a share of the total. Luna: But each structure changes the incentives. A fixed fee hurts low-value transactions more, which might discourage small purchases. A tiered fee can be complex to communicate. There's always a trade-off. Lucas: And the right choice depends on your transaction size distribution. If your marketplace has a long tail of small transactions — like an app store with lots of 99-cent purchases — a flat percentage with a low minimum might be best. If you're a B2B services marketplace with $10,000 contracts, a tiered or fixed plus percentage structure makes more sense. Luna: Let's talk about how to actually run the experiment. The gold standard is a randomized controlled trial at the user level. But that's hard on a marketplace because of network effects. If you change the fee for some suppliers but not others, the ones with the higher fee might see less demand and leave, which affects the experience of all buyers — including those in the control group. Lucas: So you get spillover effects. That's why many marketplace experiments are done at the market level — city, category, or even time-based. You roll out the change to the entire city for a period, and compare against a similar city that didn't get the change. That controls for cross-user contamination, but it reduces your sample size and makes it harder to detect small effects. Luna: There's also the risk of confounding factors. What if one city has a local holiday or a competitor promotion? You need to run the experiment long enough to average out those shocks. Lucas: One technique that's gaining traction is switchback experiments, where you alternate the fee structure every few hours or days within the same market. This is common at Uber and Lyft. They can test a 20 percent commission vs 25 percent by switching every four hours. The advantage is that supply and demand conditions are roughly similar within a day, and you get many switches, so you build statistical power quickly. But the downside is that suppliers might notice the change and game it — only working during the higher-commission hours, for example. Luna: And that's exactly what some drivers did. They learned the pattern and started driving only during low-commission windows, which biased the results. Lucas: So the experiment design itself changes behavior. That's the Hawthorne effect in marketplace form. The best experiments are often ones where users don't know they're being tested. For instance, you can roll out the fee change to a random subset of new users only — because they have no baseline expectation. Or you can test changes to the fee disclosure language — like framing the fee as a 'service charge' vs. 'commission' — to see if that affects conversion. Luna: I've seen a fascinating experiment from an early-stage marketplace that sold online courses. They tested three different fee structures: 30 percent flat, 20 percent plus a $2 fixed fee, and a tiered 25 percent on the first $100 then 15 percent after. The tiered structure actually increased both supplier sign-ups and buyer conversion, because suppliers felt it was fair, and buyers saw lower prices on high-value courses. Lucas: That's the holy grail — a structure that both sides perceive as fair. And fairness perception matters just as much as the actual economics. If suppliers think the fee is exploitative, they'll churn even if the net benefit is positive. So part of the experiment is measuring sentiment, not just transactions. Luna: Right. You can survey suppliers after the experiment. Ask them: 'Do you feel the fee is reasonable for the value you receive?' That qualitative data can save you from a purely quantitative mistake. Lucas: To wrap this up, I think the key takeaway is that pricing experiments on marketplaces are not just about finding the revenue-maximizing number. They're about understanding the dynamics of your specific network — how sensitive your supply and demand are to price changes, how your fee structure affects behavior, and how to test in a way that doesn't destroy the liquidity you've built. There's no universal answer, but there is a universal process: hypothesize, segment, test, measure, iterate. Luna: And start small. A 50-basis-point test in one city can tell you more than a boardroom debate about what the industry standard is. Lucas: Exactly. The marketplace itself will tell you the answer, if you listen to the data. And that's the real art — designing the experiment that lets the network speak.