Every business in this atlas introduces two parties who would both be richer if the introducer went away. That is not a flaw in the model, it is the model. The only question worth asking is how long the introduction stays worth paying for.
There is a formula for it, and it is worth stating plainly because most operators reason about leakage in adjectives:
Leakage ≈ (value of a single relationship) ÷ (frequency × switching friction). One-shot, high-ticket, low-friction is the maximum-leakage corner. Repeat, low-ticket, embedded is the minimum. Everything else is interpolation.
The numerator is what a defection pays. The denominator is what it costs. A $60,000 recruiting fee captured once, by two people who have already met and will not need each other again for eighteen months, is a defection that pays for itself immediately. A $400 annotation batch, one of eight hundred this quarter, routed through a vendor that also handles the contractor's tax residency in Manila, is not.
The one honest datapoint in the sector
Nobody publishes leakage rates. Not because they are embarrassing — because nobody can compute them. The single most authoritative statement available anywhere is a risk factor in Upwork's FY2025 Form 10-K:
"Customers circumvent our platforms and other workforce solutions, which adversely impacts our business… Despite our efforts to prevent them from doing so, customers circumvent our platforms and other workforce solutions and engage with or take payment through other means to avoid fees, and it is difficult or impossible to measure the losses associated with circumvention. … In addition, circumvention is likely to increase during a macroeconomic downturn, as customers may be more cost-sensitive."
That is from Upwork's 10-K for FY2025. Three things follow from it, and each one is load-bearing.
First, leakage is material enough to stay a named risk factor two decades into the category's life, at the largest and best-instrumented freelance marketplace in the world. Second, it is unmeasurable by the one party with full payment, messaging and contract telemetry. If Upwork cannot count it, your dashboard is not counting it either. Third — and this is the part operators underweight — it is procyclical to buyer cost pressure and countercyclical to your pricing power. Circumvention rises exactly when your take rate is under strain. The defence is weakest at the moment it matters.
No quantitative leakage rate exists anywhere in this research. The one study with a measured disintermediation rate — Gu & Zhu, "Trust and Disintermediation: Evidence from an Online Freelance Marketplace," Management Science — sits behind INFORMS, which blocks automated retrieval. Thumbtack, the most-cited home-services leakage case, publishes nothing; the public record holds only $400M of revenue in FY2024 — Thumbtack's own net lead and subscription revenue, not job value transacted — ~300,000 active pros, and a valuation that peaked at $3.2B in July 2021 before a $75M debt raise in 2024. This page therefore reasons structurally from one confession. Treat every ranking below as an argument, not a measurement.
The nearest thing to a documented death-by-leakage is Homejoy, whose Wikipedia entry records "leakage of its best workers to direct employment arrangements" with clients as a contributing cause of its July 2015 shutdown — alongside classification litigation and a $19 first-clean promotion against an $85 market price. Even there, leakage is listed, not sized.
What actually holds, ranked
Six mechanisms are commonly claimed. They are not equally real.
| Mechanism | Strength | Fails when |
|---|---|---|
| Workflow embedding — you are the buyer's system of record | Strongest | Never fully; going around you creates internal work |
| Liability transfer — EOR status, chaperoned calls, bonds, indemnity | Strong and rises with buyer sophistication | The buyer recreates the compliance function in-house |
| Payments and guarantee — escrow, dispute resolution, money back | Moderate | The counterparties come to trust each other |
| Non-portable reputation — platform-locked ratings and rank | Weak alone | The supplier's income no longer depends on rank |
| Frequency | Decisive but largely uncontrollable | The category is genuinely episodic |
| Exclusivity | Almost never achievable | Immediately |
Only the top two survive a sophisticated buyer. Concentrix's largest-client relationships average 16 years because the client's process lives inside Concentrix; Robert Half's own FY2025 10-K describes its staffing contracts as "generally terminable on short notice and without penalty." Same labour, same buyers, opposite hold.
Exclusivity deserves burial. GLG's S-1 concedes that network members are "almost always non-exclusively contracted" and reports declining responsiveness to project invitations — and GLG is the best-designed version of this business anyone has built. In Expert data for frontier labs, Surge AI and Scale AI have historically drawn on overlapping annotator pools through multiple sub-brands, and Mercor contractors work across labs. Nobody owns supply. See Expert networks for how a ~70–80% take survives anyway.
Contractual non-circumvention clauses belong in the "speed bump" tier with ratings and escrow. They are worth having — they convert a silent defection into an affirmative, discoverable act — but they price a lawsuit, not a lock. If a fee increase can be defeated by a WhatsApp message, price accordingly.
The formula, applied
Recruiting is the maximum-leakage corner and there is no way to design out of it. Transaction values run $30–65K per placement, so a single defection captures the entire lifetime rake. The relationship is identity-revealing by construction — you cannot anonymise a recruiter from the hiring manager they run a search with. And buying is episodic, so no habit forms. Paraform's public terms bar using the platform in a way that violates another user's contract but do not explicitly bar an employer hiring an introduced candidate off-platform; the real terms sit in a non-public recruiting agreement. Its actual defence is economic rather than legal: a ~70% recruiter split and no 60-day payment hold mean a defector gains at most ~30 points on one deal and loses inbound access to a thousand companies. That calculus holds only while req supply is abundant. When demand thins, it inverts — which is Upwork's countercyclical point restated in a different vertical. Note also that candidate-ownership windows are an openly negotiated lever, with employers pushing agencies from 12 months to 6; every month conceded is leakage bought by the buyer.
Frontier data is the defensible end, but not for the reason people assume. Frequency is high, the buyer's spend is continuous, and the vendor absorbs work the lab cannot legally or practically do itself: payroll and contractor liability across 45+ countries, and the ability to stand up thousands of vetted credentialled strangers in weeks. That is liability transfer plus workflow embedding — the two mechanisms that hold. What it is not is supply ownership. The leakage here runs upward, not sideways: OpenAI's Project Mercury hired 100+ ex-bankers directly at $150/hour, with its own 20-minute AI interview and modelling tests — a lab building a Mercor internally, screening funnel and all (Entrepreneur, citing Bloomberg). xAI cut 500 generalist annotators in September 2025 while pledging a 10x-larger in-house specialist team (TechCrunch). The in-housing trigger is legible: the function becomes large enough to matter in the buyer's P&L, strategic enough to differentiate, and hireable. All three are already true for the commodity tier.
Paid creators, clipping and UGC ad ops leaks trivially and the take rate already reflects it. Whop keeps 10% of Content Rewards payouts; competitors run 9–12%. That is a payments fee, not a marketplace margin, and it is priced that way because a brand and a clipper who have worked together once can transact by direct message forever after. Nothing in the stack — no compliance, no indemnity, no system of record — makes the second campaign harder to run off-platform than the first. The money in that vertical sits in the managed layer, where the operator buys views and bills a retainer, precisely because the marketplace layer cannot hold a spread. See Whop.
What this means for a take rate
The rule that falls out is not about leakage at all. It is about what you are selling. A take rate above ~20% is only durable if you sell something other than discovery — compliance, indemnity, capital, or a workflow the buyer runs their business inside. Discovery is the cheapest function to replicate and the first thing search, SEO and now language models attack. Bill Gurley's rake essay makes the point through oDesk cutting from 30% to 10% and passing its competitors, and through Bezos's line: your margin is my opportunity.
So model leakage explicitly, in the numerator and denominator, before you set a price. And when you cannot get the number — you cannot, nobody can — at least be honest about which of the six mechanisms you actually have. Most operators in this atlas have two speed bumps and a story. See What a rake can actually be and Marketplace, staffing firm, BPO or agency.