Marketplaces Still Work — the Liquidity Math Changed

The 2010–2020 playbook was a subsidy machine

For a decade the marketplace playbook was one idea repeated across every category: solve the chicken-and-egg problem with capital, get to liquidity, let network effects do the rest. Uber, DoorDash, Instacart, Airbnb, and the long tail of vertical marketplaces that followed them all ran a version of the same motion. Subsidize one side until the other side shows up. Buy supply, buy demand, hold your breath until the flywheel spins on its own.

The math worked because the assumption underneath it was sound at the time. Matching was a solved problem the moment you had density. Get enough drivers in a city and any rider gets a car in four minutes. Get enough restaurants on the platform and any hungry person finds dinner. The hard part was volume, not quality, so the entire discipline optimized for volume. Take rate, GMV, cohort retention, contribution margin after subsidy all pointed at the same thing. Fill the pool. Liquidity was the moat, and liquidity was bought.

That produced a generation of marketplaces that were structurally identical under the branding. The defensibility story was always “we have the most supply and the most demand, so switching is irrational.” It held right up until it didn’t.

Density stopped equaling a good match

Two things broke at once, and most marketplace writing still hasn’t caught up to either.

The first was that liquidity stopped being defensible. Capital got cheap, then it got expensive, and in both regimes the “buy your way to density” motion turned out to be rentable rather than ownable. A competitor with a fresh round could rebuild your supply side in 18 months. Multi-homing killed the switching-cost story on the demand side. The driver runs Uber and Lyft on the same phone, the restaurant lists on DoorDash and Uber Eats and Grubhub, the freelancer posts on every platform that sends leads. Density was real, and it was also for rent.

The second break is the one that matters more. In most categories, density never actually produced good matches. It produced available matches. A dense pool gets you a car in four minutes, but it does not get you the right contractor for a kitchen remodel, the right candidate for a niche engineering role, the right supplier for a specialty ingredient, the right therapist. The categories where matching was genuinely hard, high-consideration and high-variance and high-stakes on both sides, got the same volume-first playbook applied to them, and the playbook underdelivered. The match was good enough to transact once and bad enough that nobody came back. Churn ate the cohort. The flywheel never spun because a marketplace that produces mediocre matches is just a lead-gen tool with a take rate stapled on.

So the industry drew the wrong lesson. It concluded that most vertical marketplaces don’t work, when the real problem was that most vertical marketplaces had no way to make matching a source of advantage. The tooling to turn match quality into a moat did not exist yet.

AI changed which variable is scarce

Here is the shift, and it is quieter than the AI-eats-everything version you’ve read a hundred times. AI didn’t make marketplaces obsolete. It changed which variable is scarce.

For 15 years, liquidity was the scarce, expensive thing, and match quality was assumed to be a byproduct of it. That relationship inverted. Liquidity is still hard, but match quality, the thing that was previously either trivial (Uber) or nearly impossible to systematize (everything hard), is now something you can build, measure, and compound.

The mechanism is unglamorous. A marketplace sees every transaction, every rejection, every re-engagement, every dispute, every repeat booking. That data was always there and mostly wasted, because turning it into a better match at the moment of decision required either hand-tuned rules that didn’t generalize or a data science team most marketplaces couldn’t staff. Now the same signal feeds a system that gets better at matching with every transaction it observes. The match improves, the improved match drives retention, retention deepens the data, and the data improves the next match. That is a compounding loop, and unlike bought liquidity, it is not for rent. A competitor with a fresh round can buy your supply. They cannot buy the four years of matching signal that makes your placements land and theirs miss.

This is the part the 2019-vintage marketplace thesis can’t see. It still treats matching as a solved commodity and liquidity as the whole game. The economics moved. Match quality is now the durable layer, and liquidity is the cost of entry you pay to start generating the signal that actually defends you.

The honest objection is that AI tooling is available to everyone, so any matching advantage should commoditize the same way liquidity did. It doesn’t, and the reason is the input. The model is a commodity. The proprietary transaction history that trains it against your specific category is not. Two marketplaces running the same off-the-shelf model produce different match quality because they’ve observed different outcomes. The moat was never the algorithm. It’s the exhaust.

Three theses that work in 2026, and three that don’t

The theses that work now share one trait: matching is hard, high-stakes, and repeated enough that better matching visibly changes the outcome.

Specialized labor works: legal, healthcare, senior engineering, skilled trades. The cost of a bad match is high, the buyer has real preferences that generalize poorly, and the platform that learns which placements actually stick pulls away from the one still ranking by keyword and availability. B2B supply for non-commodity inputs works for the same reason. A restaurant sourcing a specific cheese or a manufacturer sourcing a tolerance-critical part is making a match decision, not a price decision, and the platform that remembers what worked last time compounds. High-consideration services, anything where the buyer researches before committing and cares who they get, work because match quality and retention are the same variable there.

The theses that don’t work are the ones where matching was never the constraint. Pure commodity marketplaces, where the only variable is price, gain nothing from smarter matching because there’s nothing to match beyond cheapest-wins. Thin-margin logistics plays where the unit economics only close at subsidized volume are the old playbook with an AI coat of paint, and the paint doesn’t change the margin structure. And any marketplace whose entire value was aggregating supply that is now directly discoverable, the “we’re the search engine for X” positioning, is getting disintermediated by the models themselves, because a good agent doesn’t need an aggregator to find the supply. It needs the thing the aggregator never had, which is judgment about which option is right.

That last category is where the AI-kills-marketplaces argument is actually correct. It’s just narrower than the argument claims. Aggregation without matching is dead. Matching with compounding signal is more defensible than it has been since the category existed.

The marketplaces that raised on liquidity and never solved for match quality are the ones in trouble. The ones sitting on years of transaction outcomes they never knew how to use are holding the most valuable asset in the category, and most of them haven’t noticed yet.