The 2026 Marketplace ops Playbook

Why this changed

A marketplace is an operations company with a website on top, and every ops leader who has run one knows it. You do not sell a product. You sell the promise that a stranger on one side will show up, do the work, and get paid, and that a stranger on the other side will be glad they came. Every one of those promises is kept by a queue somewhere.

For fifteen years the answer to a growing queue was more people, usually offshore, usually reading a macro. A new seller applies, and someone checks a license against a state website. A listing goes live with a blurry photo, and someone writes to the host. A buyer says the item never arrived, and someone reads both sides of the thread and picks a winner. The work is repetitive, but it is not simple, so it never automated well with rules.

Agents change the cost of the reading part. An agent can open the license lookup, read the photo, read both sides of the thread, check the payout history, and write a recommendation with its reasons, in under a minute and at a few cents a case. What it cannot do is carry the risk of being wrong about a person’s safety or a large customer’s money. This playbook is about drawing that line well.

The short version

Five operators helped us write this, and they run very different marketplaces: elder care aides, used restaurant equipment, event crews, camera rentals, and furnished rentals. They still made the same five moves, in roughly the same order, and they agree on where people stay in the loop.

1. 01Give the agent the whole onboarding file, not one check. It gathers documents, runs the checks, and hands a person a complete case or a clean approval.
2. 02Score every new listing before it goes live, and send the fix request to the seller in the channel they answer.
3. 03Let agents clear the low-risk review queue end to end. Route anything about physical safety, fraud rings, or a top account to a named person.
4. 04Have an agent build the dispute file from both sides and the evidence, and let it settle small claims under a dollar cap you set.
5. 05Reconcile payouts every day, hold only what the rules say to hold, and tell the supplier why in plain words.

The companies to study

Two public marketplaces set the bar that every contributor measured against. Both run supply onboarding, trust review, and payouts at a scale where small error rates become thousands of cases a day, and both publish enough about their host and Dasher programs that you can study the shape of the work from the outside.

• AirbnbMarketplace for stays and experiencesAn Airbnb for every kind of trip
• DoorDashLocal delivery marketplace for restaurants, groceries, and retailSame-day delivery from local stores

You do not need their headcount to copy their habits. Their policies are written down, their supply gets one clear reason when it is held or removed, and they treat a payout delay as a customer issue, not a finance issue. Those three habits matter more than any tool in this playbook.

Onboard supply in a day

Slow onboarding is the most expensive leak in a marketplace, because the supplier you lose is the one who had other options. At Hearthline, Marisol’s team found that an aide who waited more than three days for approval was 40 percent less likely to take a first shift. The checks were not slow. The waiting between the checks was.

The old flow had a person run each step and then wait: the ID check, the background check, the license lookup, two references, and a training quiz. Each step sat in a different tool, so a file moved only when someone remembered to look. Now an agent runs the file. It starts every check the moment the aide uploads a document, texts the aide when something is missing, calls the references through a scripted phone flow, and reads the results as they arrive.

1. 01The applicant submits an ID and a selfie. The agent starts the ID check and the background check in parallel.
2. 02The agent looks up the license on the state registry and saves a screenshot of the result to the file.
3. 03The agent texts each reference a short form, then follows up by phone if the form is not back in 24 hours.
4. 04When every check is clear and the references agree, the agent approves the applicant and books an orientation slot.
5. 05When any check returns a hit, a mismatch, or a vague reference, the agent stops and hands a person the full file with a one-paragraph summary.

Fix listings before buyers see them

Listing quality decides conversion more than any ranking change your product team will ship. A listing with dark photos, a missing dimension, or a price far above the comparables does not just fail to sell. It teaches buyers that the search results are not worth reading, and that cost spreads to every good seller on the page.

Every contributor now scores a listing before it goes live. The agent reads the photos, the title, the description, and the price, then compares them to a short written standard for that category. At Stallward, a used fryer listing must show the data plate, the gas type, the BTU rating, and a photo of the inside of the vat. If the data plate photo is missing, the agent writes to the seller and names the one photo it needs.

• Write the standard for each category in plain words, with one good example and one bad example. The agent and your reviewers read the same page.
• Ask for one fix at a time. A list of nine problems gets ignored; one clear request gets done.
• Let the agent publish a listing that meets the standard. Hold a listing only for a policy reason, never for style.
• Check a random 2 percent of agent-approved listings each week, and change the standard when reviewers disagree with it.

Price is the one area where contributors keep a person close. An agent can flag a listing priced 60 percent below similar listings, because that is a classic sign of a scam. It should not tell a seller what to charge. At Roostly, pricing advice comes from a person on the host success team, and only when the host asks for it.

Trust and safety review

Trust and safety is where the line between agents and people matters most, so draw it before you build anything. Every contributor uses some version of the same three tiers. The tiers do not depend on how hard a case is to read. They depend on how bad it is to be wrong, and on who pays for the mistake.

1. 01Tier 1, agent closes: spam listings, off-platform contact attempts, duplicate accounts with no money at stake, and profile content that breaks a written rule. The agent acts, logs the reason, and tells the user.
2. 02Tier 2, agent recommends and a person approves: account suspensions, removals of a supplier with good history, and any action that holds money. The agent builds the case; a reviewer clicks approve or changes it.
3. 03Tier 3, person leads from the start: any report of physical harm, harassment, or a threat; any suspected fraud ring; any case that involves a top-100 account by gross bookings. The agent gathers evidence but does not write a recommendation.

The reason the agent writes no recommendation in Tier 3 is anchoring. Dev Ramanathan tested it at Stallward: when reviewers saw an agent’s suggested outcome on fraud-ring cases, they agreed with it 94 percent of the time, including in the cases a second senior reviewer later reversed. When reviewers saw only the evidence, their reversal rate on later audit fell from 9 percent to 3 percent.

Disputes and payouts

Disputes are the most hated queue in a marketplace, because both sides are upset and both sides are partly right. The slow part was never the decision. It was building the file: the order record, the messages, the photos at pickup and return, the delivery scan, the past claims of each party, and the policy that applies.

At Loaner, a typical dispute is a lens returned with a scratch. The agent pulls the check-out photos and the check-in photos, compares them, reads the chat between renter and owner, checks the renter’s past claims, and writes a short file with a proposed outcome. Under $250, with clear photos and no history of claims on either side, the agent settles the case and pays it from the protection fund. Above that cap, or with any doubt about the photos, a person decides.

Payouts follow the same pattern. An agent reconciles every payout against the bookings each morning, finds the breaks, and explains each one. It can release a payout that a rule held by mistake. It cannot hold a payout that no rule covers. A supplier whose money is late hears about it from the agent before they have to ask, with the reason and the new date.

The tools we rely on

Most marketplace ops stacks look alike, because the jobs are alike. You need a way to verify people, a way to score risk, a payments rail that keeps a record for each supplier, a place for the review queue, and a way to see the numbers. We asked each contributor to score the tools they use on a 10-point scale and to name the one they would keep if they could keep only one.

One warning from Dev: do not buy a fraud tool and let its score act alone. Every contributor who did that saw good suppliers removed by a rule nobody could explain. Put the score in the agent’s file, and let the tiers decide who acts on it.

So what do you do?

Start with the queue that hurts the most, not the one that looks easiest to automate. For most marketplaces that is supply onboarding, because every day of delay costs you suppliers you already paid to acquire. For a marketplace with a safety problem, it is the review queue, and you start by writing the tiers before you write a single prompt.

1. 01Write down your tiers this week: what an agent may close, what it may only recommend, and what a person leads. Name the person for each Tier 3 category.
2. 02Pick one queue and let an agent build the full case file for every item for two weeks, while people still decide. Measure how often they agree.
3. 03When agreement passes 95 percent on a case type, let the agent close that type, and audit 5 percent of its closes every week.
4. 04Set a dollar cap for agent-settled disputes. Start at $100, and raise it only after a month with a reopen rate under 2 percent.
5. 05Reconcile payouts daily and send every supplier a message before their money is late.
6. 06Review the tiers every quarter with your trust lead. Move a case type down a tier only with audit numbers, never because a queue is long.

The marketplaces that do this well end up with smaller queues, faster approvals, and reviewers who are less tired. They do not end up with fewer judgment calls. They make the same number of hard calls as before, with better files, and they make them in hours instead of days.