How do I protect my marketplace from fraudulent sellers?

Seller fraud is not a rare edge case. On marketplaces without active screening, roughly 1 in 12 new seller accounts is fraudulent in some form, according to a 2023 LexisNexis Risk Solutions report. The cost is not just chargebacks. It is refunds, dispute resolution labor, brand damage, and the buyers who leave and never come back.

The good news: most seller fraud follows repeating patterns. That predictability is exactly what makes it solvable with a screening model, even on a startup budget.

Four patterns account for the vast majority of marketplace seller fraud. Understanding them is the first step to building a system that catches them early.

The most common is triangulation fraud. A fake seller lists real products at below-market prices, collects payment from buyers, then uses stolen credit cards to order those products from legitimate retailers and ship them to the buyer. The buyer receives the item and never suspects anything. The stolen card's owner later files a chargeback. The marketplace gets caught in the middle, absorbing the reversal.

Dropship-and-disappear fraud follows a similar setup but with no intent to fulfill. The seller collects payments for a short window, typically 7–14 days, then disappears before fulfillment is due. According to Chargebacks911's 2023 marketplace report, this pattern peaks in the 8–11 day window after seller account creation, which means platforms without time-gated payouts are the most exposed.

Account takeover fraud targets existing legitimate sellers. A bad actor gains access to a real account with good reviews, changes the payout details, and runs a burst of fraudulent orders before the original seller notices anything is wrong. Sift's 2023 Digital Trust and Safety Index found account takeover attacks on marketplaces increased 79% year-over-year.

Wholesale identity fabrication is slower and more deliberate. Fraudsters construct synthetic seller identities using real document fragments, real addresses, and phone numbers that pass basic verification. These accounts often stay dormant for weeks to build a credibility window before activating fraud.

A screening model does not look at one signal in isolation. Humans do that, and fraudsters have learned to pass any single check. The model looks at clusters of signals that correlate with fraud, even when each signal individually looks normal.

At registration, the model checks whether the phone number, email domain, IP address, and device fingerprint have appeared on any known fraud networks. It also looks for velocity signals: how many accounts have registered from the same IP block in the past 72 hours, whether the device has been used to create accounts on other platforms that were later flagged, and whether the registration timing matches known fraud campaign patterns.

Once the seller is active, behavioral signals matter more. A fraud model tracks the ratio of high-value orders to low-value orders in the first two weeks, the gap between listing creation and the first sale, and whether the payout bank account was added before or after listing creation. Individually, none of those signals is definitive. Together, a model can assign a risk score that catches 78–85% of fraudulent sellers before they cause a loss, based on internal benchmarks from Stripe Radar's published fraud detection data.

The model runs continuously, not just at signup. That is how it catches account takeovers: a sudden change in listing behavior, payout account, or login location on an established account triggers a review flag, even if the original account passed all registration checks.

Building this in March 2024, an AI-assisted development team can compress the model-building phase significantly. The feature engineering pipeline, the risk scoring logic, and the review dashboard that surfaces flagged accounts are all areas where AI-assisted coding eliminates 40–60% of build time. A model that would have taken 14 weeks to build from scratch in 2022 ships in 6–8 weeks now.

Yes, and by a wide margin on two specific fraud types.

Manual reviewers are good at catching obvious red flags: a listing with stolen product photos, a seller with a brand-new account and 50 listings in 24 hours. They are poor at catching anything that requires comparing a seller's behavior against thousands of other accounts simultaneously. A human reviewer looking at one account cannot know that the bank routing number on that account also appears on 14 other newly registered accounts flagged in the past month. A model knows that in milliseconds.

According to a 2023 McKinsey analysis of marketplace trust and safety operations, human review teams miss approximately 40% of synthetic identity fraud because the individual signals look clean. The fabricated documents pass, the address is real, and the phone number is active. The model catches these accounts by recognizing that the combination of signals, specifically a new device, a recently registered email domain, and a payout account opened the same day as seller registration, matches a known fraud fingerprint even when nothing looks wrong on the surface.

The second area where models outperform humans is speed. A manual review team operating at reasonable capacity can process 200–400 seller applications per day. A model processes every application in under 200 milliseconds. On a marketplace growing fast, that gap between human and model throughput is the difference between a 48-hour approval backlog and instant decisions.

Manual review teams remain useful for one thing the model cannot do well: edge cases. When the model flags a seller with a confidence score in the 55–70% range, a human reviewer adds judgment that the model cannot. The best trust and safety setups use the model for clear approvals and clear rejections, and route the gray-zone accounts to a reviewer. That combination catches more fraud than either approach alone.

The answer depends on whether you are building a screening model from scratch or adding fraud detection to a marketplace that is already live.

For a new marketplace integrating fraud screening during the initial build, the model, risk dashboard, and review queue add $8,000–$12,000 to the overall development cost at an AI-native agency. That assumes the marketplace already has seller account infrastructure and payment processing in place.

For a live marketplace retrofitting fraud detection onto an existing platform, expect $18,000–$28,000. The higher cost reflects the additional work of connecting the model to an existing data architecture without breaking what is already running.

Third-party fraud APIs like Sift, Kount, or Stripe Radar add $0.02–$0.08 per transaction on top of build costs, depending on volume. At 10,000 transactions per month, that is $200–$800/month in API fees. At 100,000 transactions, budget $2,000–$8,000/month. These costs scale with your volume, which means they are proportional to revenue, not a fixed overhead.

The return on that spend is concrete. A marketplace processing $1 million in gross merchandise value per month typically absorbs $10,000–$30,000 in fraud losses without screening. A functioning model cuts that by 70–85%, saving $7,000–$25,000 per month. At $18,000 to build, payback period is under 90 days for any marketplace above $500,000 monthly GMV.

This is the real tension. Tighten your screening too much and you start rejecting legitimate sellers. Loosen it too much and fraud gets through. Most marketplace founders do not realize this is an adjustable dial, not a binary on/off.

The practical approach is tiered onboarding. New sellers start with lower daily transaction limits, say $500 per day, until they clear a behavioral threshold: 10 completed orders with no disputes, or 30 days of clean activity. Once they cross that threshold, limits lift automatically. Sellers who want to onboard faster can submit additional verification documents to skip the waiting period.

This structure means a fraudster who slips through registration is capped at $500 of damage before the model has more behavioral data to act on. A legitimate seller who needs to move volume fast has a clear path to doing so. The fraud team gets a credibility signal from actual transaction behavior rather than relying entirely on the registration snapshot.

On approval speed: a well-tuned model approves 85–90% of legitimate sellers instantly, with no human review needed. The remaining 10–15% go to a review queue. If your review team targets a 4-hour turnaround on queued applications, the median seller wait time across all applications stays under 30 minutes. That is competitive with the fastest marketplaces in any vertical.

One configuration detail worth getting right early: dispute windows. If a seller's payout releases before a buyer's dispute window closes, you have no mechanism to claw back fraud losses. Setting payout release to 7 days post-delivery, with a hold extension triggered by any open dispute, costs nothing to implement and eliminates a whole category of fraud exposure.

Timespade builds fraud screening systems across the Predictive AI vertical, including risk scoring models, behavioral monitoring, and the review dashboards that surface flagged accounts to your trust and safety team. The same team also handles the marketplace infrastructure the model connects to, which means one contract instead of coordinating a fraud vendor and a dev agency separately.

Book a free discovery call