Disintermediation: Why Users Go Around the Platform

By Brian Tr · Updated Oct 7, 2026

A marketplace spends money to bring a buyer and a seller together. Then the two of them swap phone numbers, agree a price over chat, and pay each other directly. The platform did the hard part and collected nothing for it. That's disintermediation, also called platform leakage or circumvention.

It's one of the oldest worries in marketplace strategy, and it's more contested than most people assume. The research shows it clearly in some settings and finds almost none in others. This article covers what disintermediation is, why it happens, which marketplace types are most exposed, the levers platforms use against it, what each lever costs, and how to measure it without fooling yourself.

We'll stay with evidence we can cite. Where we're reasoning from first principles instead, we'll say so.

What disintermediation is

Harvard Business School's technical note by Benjamin Edelman and Philip Hu, Disintermediation in Two-Sided Marketplaces, describes the problem like this: users rely on the marketplace to find each other, but then perform related future transactions, or even the current transaction, without the platform's involvement and without paying any fees the platform charges.

Two distinctions are worth making early.

Current versus future transactions. Some users pay outside the platform for the very job they found on it. Others complete the first job on the platform, then rebook directly. The second kind is harder to prevent and harder to see.

Leakage versus a fair exit. If a platform's fee exceeds the value it adds to a repeat relationship, a user leaving isn't betrayal. It's a price signal. We'll come back to that, because it changes how you respond.

Fees come up in every definition, but they're not the only motive. Buyers and sellers also go around platforms to avoid friction, to keep messaging in tools they already use, or to speed up a deal. The common thread is that the platform's contribution has become smaller than its cost in the user's eyes.

Why it happens

Three forces repeat across the cases we can document.

1. Trust built on the platform makes the platform less necessary

The best-known empirical result here is Grace Gu and Feng Zhu's paper, Trust and Disintermediation: Evidence from an Online Freelance Marketplace, published in Management Science in 2021. The researchers ran a randomized control trial on a major freelance marketplace. According to the abstract, enhanced trust increased the likelihood that high-quality freelancers were hired. But when trust was sufficiently high, it also increased disintermediation, which offset the revenue gains from hiring better freelancers. They also found that clients and freelancers differ in how likely they are to disintermediate.

Read that carefully. Trust is what the platform sells. The same mechanism that makes a buyer comfortable hiring a stranger also makes the buyer comfortable paying that stranger directly next time. The paper describes how the trust signal helps and hurts at once.

According to a later working paper's literature review, Gu and Zhu detected disintermediation by scoring chat messages against a dictionary of phrases such as "my phone number," "avoid fees," and "Venmo." We'll use that idea in the measurement section.

2. Fees are paid on every transaction, so repeat relationships are expensive

A fee that feels small on a first job feels different on the twentieth job with the same person. By then the buyer and seller know each other, and the platform's matching is no longer doing anything. This is reasoning, not a finding, but it follows from how take rates work. For the mechanics of the fee itself, see marketplace take rate.

3. The platform's job is finished early

Matching happens once. After that, the platform adds value only if it continues to do something: payments, protection, scheduling, reviews, tools. Where it adds nothing after the match, the incentive to leave grows with each repeat transaction.

The counterexample: when leakage barely shows up

It would be tidy to say every marketplace leaks. The evidence says otherwise.

A 2023 working paper by Ekaterina Astashkina, Robert Bray, Ruslan Momot, and Marat Salikhov, A Disquieting Lack of Evidence for Disintermediation in a Home-Cleaning Platform, looked at a home-cleaning app. The dataset recorded the geographic distances between 5,391 cleaners and the 95,633 residences they cleaned, with frequent readings during working and off hours. The authors tested whether cleaners returned to those homes for off-the-books jobs. They found no evidence of it. The paper states that no more than one in 83 relationships formed on the app would end in disintermediation. This is a working paper, so treat it as evidence under review, not a settled result.

The authors said the setting looked ripe for leakage: a sizeable commission, an informal economy, and in-person meetings that make it easy to swap contacts. They then listed possible reasons, after the fact, for why it didn't happen:

• Cleaners were relatively plentiful in that city, so someone wanting an under-the-table arrangement could find one without using an app.
• The platform's security may have mattered to customers letting strangers into their homes.
• The average cleaning cost about 31.58 USD, small relative to the incomes of the customers.
• Repeat relationships rarely formed. About 90% of cleaner and residence pairs had just one cleaning.

The same paper notes that other studies did find signs of leakage. As it summarizes them, one study of a European cleaning platform estimated that the platform would enjoy about 24 percent more cleanings if not for disintermediation, though the Astashkina team points out that this was reached partly by eliminating alternative explanations. We're reporting these figures as that paper describes them, not as our own verification of the underlying studies.

The takeaway is not that leakage is a myth. It's that exposure depends heavily on the category and the market, so you need to measure your own before spending on defenses.

Which marketplaces are most exposed

Edelman and Hu's note says it assesses which marketplace models are most vulnerable, but we can't reproduce its analysis here. The framework below is our own reasoning, informed by the cases above. Treat it as a checklist for hypotheses, not a ranking.

Services marketplaces with skilled, differentiated sellers sit toward the high-exposure end of this table. Commodity goods and anonymous, on-demand matching sit toward the low end. The Gu and Zhu setting (freelance work, where a client may want the same freelancer again) fits the first pattern. The cleaning study, with plentiful supply and mostly one-time pairs, fits parts of the second. Repeat purchasing patterns differ a lot between B2B and B2C marketplaces, which is one reason exposure varies by model.

Counter-levers and their trade-offs

Platforms rarely rely on one defense. The usual toolkit falls into five groups. We'll describe each and what it costs.

Make staying more valuable than leaving

If the platform keeps adding value after the match, leaving becomes a worse deal. That value can be software the pair uses to run the job, scheduling, invoicing, reviews, insurance, or access to a wider pool of customers. Our article on value-added services covers the options.

The trade-off: services cost money to build and run, and they only deter leakage if users actually want them. A tool nobody needs doesn't hold anyone.

Protect the money

Holding payment until the work is done, and offering dispute handling, gives users a reason to pay on-platform. Airbnb states this directly in its help center: paying for a booking outside Airbnb means losing the AirCover protection that comes with booking through the platform. Upwork's support article on direct payment says that when a contract moves outside the platform, Upwork can no longer offer any support, including payment protection and dispute resolution.

Both platforms are using the same logic: the fee buys a safety net, and the safety net disappears off-platform. For how funds are held and released, see marketplace payments.

Buyer guarantees work the same way, and most of them apply only to on-platform transactions; see guarantees and buyer protection. The trade-off: protection is only persuasive to people who worry about being cheated. Repeat partners who trust each other may not care, and that's the segment most likely to leave.

Write the rule into the terms, and attach a price to leaving

Policy is the bluntest tool. Airbnb's help center says it's against its policy for hosts to ask a guest to take a booking off Airbnb or to communicate outside the platform before booking. Upwork's support page defines circumvention as a client and freelancer meeting on Upwork, then conducting their relationship outside it by sending payment another way without paying the conversion fee. The same page treats the conversion fee as one of the ways a relationship can legally move off-platform. We're deliberately leaving out the amounts, which change; check the vendor's current page.

That second idea is worth pausing on. A conversion fee admits that some relationships will outgrow the platform, and prices the exit instead of forbidding it. It converts an enforcement problem into a buyout.

The trade-off: rules need detection and enforcement to mean anything, usually run by the trust and safety function, and heavy enforcement sours relationships with the very sellers you want to keep. Penalties also hit honest users who simply wanted to use another channel.

Mask or delay contact details

Many platforms hold back phone numbers, emails, and last names until a booking is confirmed or paid for, and filter messages that contain them. Airbnb's rule against pre-booking communication elsewhere is one example of the policy side.

The trade-off: masking adds friction to legitimate conversations, and determined users find workarounds. Gu and Zhu's dictionary approach, as described in the later paper, shows that such attempts can be detected in chat, but detection doesn't stop someone who rewrites "my phone number" as something else. In some categories, like in-person services, the parties meet anyway and masking buys little.

Redesign fees so leaving costs less than staying

Fee structure shapes the incentive. Options include lower fees on repeat transactions, fees that fall as the relationship ages, charging the buyer instead of the seller, or moving from commission to subscription. The Astashkina paper's literature review cites theory work on how information design and pricing strategies can deter disintermediation, so this is an active research area. We haven't verified those papers ourselves, and a given redesign needs testing on your own data.

The trade-off: lower fees on repeat jobs reduce revenue on the transactions that were staying anyway. See marketplace revenue models for how different structures change who pays and when.

Summary of levers

How to detect and measure leakage

Leakage is secretive, so it's hard to observe. Astashkina and colleagues say it has proven difficult, if not impossible, to estimate. That doesn't mean you can't try. It means you should triangulate.

Message scanning. Search chat for contact-sharing phrases and payment app names, as Gu and Zhu did. This measures attempts, not completed leakage, and it undercounts people who move the conversation without typing anything flaggable.

Repeat-rate patterns. Compare how often buyer and seller pairs transact again. A sharp drop-off after the first job in a category where repeats should be common is a clue. The Karacaoglu study, as summarized in the Astashkina paper, used a version of this: it looked at whether customers left at higher rates after their cleaner's final cleaning than after a typical one, with the figures of about 66% and 74% reported in that summary. The Astashkina team notes the method reaches its conclusion partly by process of elimination.

Supply-side exit analysis. Track what happens to a seller's repeat customers when the seller leaves the platform. If customers leave with them, some of that relationship was probably going to continue elsewhere.

Behavioral signals around trust. Gu and Zhu found disintermediation rose when trust signals got stronger. Watch whether pairs with strong reviews show more contact-sharing, since that's where the effect concentrates.

Direct observation, where you can get it. The cleaning study used location data. Most platforms can't, and shouldn't assume a method like that is available or appropriate.

Whatever you use, write down the definition before you compare numbers. "Leakage rate" can mean attempts, completed off-platform deals, or lost revenue, and those aren't the same figure. And remember what the counterexample shows: a clean zero is a real possible answer. Don't build a defensive program around a problem you haven't confirmed.

Key Facts: Marketplace Disintermediation

• Edelman and Hu's HBS note defines the risk as users finding each other on a marketplace and then transacting without the platform's involvement and without paying its fees.
• Gu and Zhu's randomized trial on a freelance marketplace found that when trust is sufficiently high, it increases disintermediation, offsetting the revenue gains from hiring better freelancers.
• A 2023 working paper on a home-cleaning platform found no evidence of disintermediation and argued that no more than one in 83 relationships formed on the app would end in it.
• Airbnb's help center says paying outside Airbnb means losing AirCover protection.
• Upwork's support page defines circumvention as moving a relationship off Upwork without paying the conversion fee, and says payment protection and dispute resolution aren't available off-platform.
• Exposure depends on the category: repeat relationships, differentiated sellers, and little platform value after the match raise it (our reasoning, consistent with the cases above).

Related Reading

• Marketplace Take Rate
• Value-Added Services in Marketplaces
• Marketplace Payments
• Marketplace Revenue Models
• Managed Marketplace
• Guarantees and Buyer Protection in Marketplaces
• Reviews and Rating Systems in Marketplaces
• Trust and Safety in Marketplaces