Trust and Safety for Marketplaces

A marketplace is two products bolted together — a place to sell and a place to buy — and fraud exploits the seam between them. The same account system, payments rail, and reputation mechanics that make a marketplace work also give a bad actor several ways to profit: register as a fake seller, defraud as a buyer, or, most damagingly, play both sides at once to launder money, farm reviews, or cash out incentives. Trust and safety on a two-sided platform is therefore not one problem but a set of overlapping ones, and the connective tissue that catches most of them is a durable sense of who the device behind each account really is.

This article maps the main abuse vectors on a marketplace and shows how device intelligence, tied together with identity clustering, addresses them without turning the platform into an obstacle course for honest users.

The abuse vectors, and why account signals miss them

Almost every serious marketplace abuse depends on operating multiple accounts that appear unrelated. Email addresses are free, phone numbers are cheap, and payment instruments can be rotated. So any defense keyed on account attributes alone is playing whack-a-mole. The recurring vectors:

• Fake and reincarnated sellers. A seller banned for scamming buyers simply re-registers. Without device continuity, the new account is a stranger.
• Buyer-side fraud. Stolen cards, chargeback abuse, and item-not-received scams run through disposable buyer accounts.
• Collusion and self-dealing. One operator runs both a seller and a set of buyer accounts to fake sales volume, launder funds, or trigger payouts.
• Review manipulation. Rings of accounts post reviews to inflate a seller or bury a competitor.
• Incentive abuse. New-user credits and referral bonuses harvested across many accounts.

The unifying weakness the attacker has is that all of it runs on a limited set of physical devices. That is the lever.

Device identity as the connective tissue

A device fingerprint with a confidence score gives you a stable handle on the machine behind an account, independent of the email or card attached to it. On a marketplace this single capability collapses several distinct problems into one:

• The reincarnated seller shows up on the same device as the banned account — a link no amount of new email addresses hides.
• The colluding buyer and seller accounts share a device or a tight device cluster, exposing self-dealing.
• The referral abuser’s twenty accounts trace to two phones.

Building this into an identity graph — nodes for accounts, devices, payment instruments, and shipping addresses, edges where they co-occur — turns isolated signals into visible rings. A cluster of accounts sharing a device and a payout address is not a coincidence; it is a fraud ring rendered legible.

Abuse vector	Primary device signal	Graph link that confirms it
Reincarnated seller	Fingerprint matches banned account	Same device, new account
Buyer-seller collusion	Shared device / cluster	Device edge between the two sides
Review ring	Reviewer devices cluster	Reviewers link to seller device
Incentive abuse	Many accounts, few devices	Fan-out from device to accounts

Applying friction where the risk is, not everywhere

The trap in trust and safety is protecting the platform by degrading it for everyone. The discipline is to spend friction only where a suspect score justifies it, and to make the response proportional to the action and the stage of the relationship.

• Onboarding. New sellers with a device linked to prior bans or a datacenter/proxy egress get held for verification; clean new sellers list immediately. Protect signups with device-aware signup checks.
• Transacting. Score payments with device signals so a stolen-card checkout is challenged while a returning buyer sails through.
• Payouts. The highest-value moment. Hold or review payouts where the seller device clusters with the buyers who generated the volume — the classic collusion tell.
• Reviews. Suppress or flag reviews whose authoring devices link to the seller or to a coordinated cluster, rather than moderating review text by hand.

The result is that the median honest user — the buyer who checks out once, the seller who lists a few items — never encounters a challenge, while the concentrated abuse that a small number of devices generates absorbs the scrutiny.

Operating it over time

Trust and safety is a program, not a launch. A few practices keep it working:

• Preserve the device link across bans. When you ban an actor, retain the device identity so the next account inherits the reputation. This is where a reputation network extends the reach beyond your own platform.
• Watch new-versus-returning device ratios. A sudden surge of brand-new devices on a listing category or a promo often precedes an abuse wave.
• Instrument false positives. Track false-positive rate explicitly; a marketplace lives on liquidity and cannot afford to choke legitimate sellers.
• Explain decisions. Reason codes let your operations team and your appeals process justify a hold, which matters when a real seller is caught in a cluster by accident.

Frequently asked questions

Why is marketplace fraud harder than single-sided fraud?

A marketplace has to protect both buyers and sellers, and the same person can play both roles to defraud the platform. Collusion and self-dealing across the two sides create fraud patterns that single-sided defenses miss.

What is the single most useful signal for marketplace abuse?

A durable device identity, because most marketplace abuse depends on one actor operating many accounts, and the device fingerprint links those accounts even when emails, payment methods, and IPs differ.

Can you stop review manipulation without a review police team?

Largely, yes. Linking reviewer accounts to the seller device or to a coordinated device cluster catches the bulk of fake-review rings automatically, leaving only edge cases for manual review.

Marketplace trust and safety comes down to seeing through the account layer to the small number of devices actually behind the abuse, linking them into a graph, and spending friction only where that graph lights up. Get the device identity right and fake sellers, collusion, and review rings stop being invisible. See the account takeover guide for the login side and the playground to try the signals.