What a rake can actually be

Every middleman business in this atlas is an argument about one number: what fraction of the money passing through you, you get to keep. The number has a floor set by what it costs you to operate and a ceiling set by what the two sides will tolerate before they route around you. Almost nobody argues about the floor. The whole business is the ceiling.

The canonical statement of the ceiling is Bill Gurley's A Rake Too Far (2013). His examples are still the reference set: iTunes, the App Store and Facebook at ~30%; Groupon at a 38% rake on top of a ~50% merchant discount, an effective rake of 70%; eBay at ~10% in 2011 ($6.6B of revenue on $68.6B of GMS); Amazon's marketplace at 6–15% by category; Booking.com's 10% agency model out-competing rivals charging 30%+; and oDesk cutting from 30% to 10% and passing its competitors on the way past. His three conditions for a durable rake: high volume at a modest rake beats the reverse; the value delivered must exceed the transaction cost so suppliers stay; and the pricing friction must not make the platform worse than the non-platform alternative. The warning underneath is Bezos's — your margin is my opportunity.

Albert Wenger's rule of thumb, quoted by Version One, is blunter: take rates above 20% are not sustainable long-term, because they suppress transactions and invite entry. Version One's own valuation heuristic assumed 10–15% at scale.

And yet expert networks have held 70–80% for four decades. Both statements are true. The reconciliation is the subject of this page.

#The benchmarks

All computed from FY2025 primary filings unless noted.

The distribution is not random. The 8.5% end is freight brokerage, where supply is homogeneous, buyers are price-transparent and switching costs are near zero. The 70% end is a chaperoned phone call.

#The 20% line, and what buys you the right to cross it

Look at the two poles.

C.H. Robinson brokers freight at ~8.5% and books $17.0B of gross revenue to do it. It has a licence, a bond and cargo liability — real regulatory furniture — and it still cannot hold double digits, because a shipper can call three brokers and a carrier can answer four. Discovery is most of the job, and discovery is a price.

GLG does the opposite. Its buyer is a hedge fund or a consultancy paying with someone else's money against extreme urgency, and the actual product is not the expert. An expert who sets a rate of $200/hour produces a client bill of $800–$900 (Inex One forum) — a 70–80% take that has survived four decades, three compliance scandals and every attempt to route around it. What the fund is buying is the chaperone: it cannot legally cold-call a competitor's ex-employee on its own. See Expert networks and GLG.

The other durable rakes in the record run on the same logic:

• Permanent placement — one-shot, high-stakes, and the buyer's alternative is an expensive internal process. Robert Half's perm-placement segment runs a 99.8% gross margin against 39.0% in contract staffing, in the same company with the same salespeople (RHI 10-K).
• Payments-adjacent rakes. Gurley's "payment flow" factor: when the fee is netted out of the supplier's payout rather than invoiced, it reads as a distribution cost, not an expense — and that raises the sustainable rake.
• Regulated or bonded intermediation: staffing's employer-of-record liability, a freight broker's bond. The buyer cannot disintermediate without recreating the function.

And the places the rake gets competed away: commoditised supply with price-transparent buyers; any position where the contribution is discovery alone; and any fee level high enough that an aggrieved supplier will fund the alternative. Amazon's third-party seller fees rising from 35% to 52% over six years is on a16z's list precisely as a migration trigger.

#Take rate is not gross margin

The two get confused constantly and they answer different questions.

Upwork takes 18.7% and runs a 77.2% gross margin. Robert Half marks contract labour up by roughly 65% and runs a 39.0% gross margin. The first is a marketplace: the fee is nearly all margin because the worker is never Upwork's cost. The second is a labour reseller: the worker's wage is cost of revenue. A high take rate on a gross-booked staffing spread and a high take rate on a net-booked marketplace fee are not the same asset, and GMV is not revenue is where that difference does the most damage.

#The tell: a rising rake on flat volume

This is the single most useful diagnostic on this page, because it is computable from two published numbers and it is not something a company can spin. Rake up plus volume up is pricing power. Rake up plus volume flat is a harvest, and the harvest ends.

#Where the atlas sits

The verticals in this atlas span nearly the whole observed range.

Read down the column on the right and the rule reads itself. The spreads at the top are paid for liability, physical capital or an output the buyer cannot assemble. The spread at the bottom is paid for a payment rail, which is why Whop keeps 10% of Content Rewards payouts and ~5.5% blended across its platform, and why nobody in that vertical is defending a number.

The interesting middle is Expert data for frontier labs. A 27–33% gross margin is below the 20%-rule threshold in spirit even though the headline take is above it — because that spread is a staffing spread on a gross-booked flow, not a fee on a net-booked one. Mercor is not GLG with better software; on this axis it is closer to Robert Half's contract-staffing segment. The lab is buying speed and volume of sourcing, and speed of sourcing is a service, not a licence.

#Where the record is thin

Every take rate in the private half of this atlas is either self-reported or backed out of two other self-reported numbers. The public filings above are the only entries anyone has audited — which is why they are worth more than the more exciting numbers around them.