"Pay Me on Another App": Why Keyword Lists Miss Off-Platform Payment Requests

TL;DR

• Off-platform payment requests rarely contain a word you can safely block.
• Keyword lists miss the disguised versions and block buyers asking honest questions.
• The violation is the intent to move a payment off the platform, so detection has to read intent and the conversation around the message.
• Bounce the first attempt, review repeats, and act on the account when the pattern repeats.
• Stream's AI agent Nora reviews each flag against the conversation, the user's history, and your policy, then acts or recommends an action for a moderator to approve.

"Can do $40 if you send it on Venmo instead, saves us both the fees."

There's no profanity in that message, no slur, and no link. Yet, it's one of the most expensive messages a marketplace can receive.

Once payment leaves the platform, the buyer loses purchase protection, the trust and safety team loses visibility, and the app loses the fee its business runs on. If the seller turns out to be a scammer, the buyer still comes back to your support team, and now there's no transaction record to help them.

Most teams try to stop this with a keyword list. Here's why that keeps failing, and what works instead.

What It Looks Like in Practice

Off-platform requests rarely say "pay me outside the app." In the conversations we review, they usually take one of four shapes.

A discount for leaving. "Cheaper if we skip the site fees." It reads like a friendly offer, which is why it works.

Disguised spellings. "V3nm0", "c@sh app", or a phone number written out in words. In October, one UK services marketplace we work with saw 10 buyer accounts post 12 fake "council tender" projects. They drew 39 seller inquiries and pushed each one to WhatsApp and Gmail, with slashes between each character of the contact details so no filter would read them as email addresses.

Moving the chat first. "Text me at...", "find me on WhatsApp", "check my bio." The payment ask comes later, somewhere you can't see it.

Split across messages. The handle goes in one message and the ask goes in the next. Each message on its own is clean.

Why Keyword Lists Miss It

A blocklist fails in both directions here.

It under-blocks. Disguised spellings, number words, and split messages all get past exact matches, and new variations show up as soon as people learn a word is blocked. In our own data, AI harm detection labeled 231 forbidden links, 28 cases of personal info, and 15 off-platform deals in one month of seller event text from five marketplaces. Each one is a way to move a buyer off the platform.

For example, Tradeblock, a sneaker trading marketplace, started with more than 100 keyword rules for phrases like "pay me on Zelle" and "DM me on WhatsApp." It then moved to AI moderation to catch evasive versions, like spaced-out letters and lookalike characters. As its CEO put it, the AI stopped them from "playing whack-a-mole with new spam variants."

It over-blocks. "Do you accept PayPal through checkout?" is a fine question. Block every payment app's name, and you penalize the buyers you most want to keep, then spend your moderators' time on appeals.

The violation isn't a word. It's an intent: getting the payment to happen somewhere else. Keyword rules still have a place for known patterns, so don't rule them out completely.

How to Catch It From Context

To catch intent, moderation has to read the message the way a person would.

Classify the intent. An AI label for platform bypass catches "let's do this outside the app" in any phrasing, including ones nobody has written down yet. Peerspace uses a custom "Platform Bypass" label to flag hosts and guests proposing off-app bookings, and routes more than 1 million text moderation requests a month to its trust and safety team before a booking leaves the platform.

Read the entire conversation. "Sure, it's 555..." means nothing on its own and everything right after "what's your number?"

Check images, too. Sellers who know their text is watched put phone numbers and payment handles in photos. OCR reads the text in the image, and PII masking hides a phone number or email before the other person sees it. Gumtree uses PII masking and OCR together across more than 8.8 million text moderations a month, catching scammers who try to send payment and contact details around text filters.

Weigh the account and the content. A brand-new account, a first message to this buyer, a discount paired with a contact handle, and urgency together are a much stronger signal than any one of them alone.

Keep a few precise rules alongside the AI. One marketplace we moderate went from catching 80% of the violations in its test set to 95% after we tightened its prompts for "cash for pickup" and contact sharing and added a regex for phone numbers broken up with symbols.

Tune by surface. Listing comments, DMs, and live chat need different thresholds. A handle in a public profile may be fine. The same handle in a first DM to a buyer usually isn't.

How Nora Works to Stop a Scam on a Marketplace App

Stream's AI agent Nora reviews flagged content the way your best moderator would. She checks each flag against the conversation, the user's history, earlier actions, and your policy, then takes the action herself or recommends one. Detection catches the message, but someone still has to decide what happens next, and on a busy marketplace that queue grows faster than a team can clear it. That decision is Nora's job.

Here's how she handles the split-message scam from earlier. A buyer asks, "Could you do $60?" The seller replies in four short messages: "if you want it cheaper," "find me on the gram," a handle, and "$10 off there."

The platform bypass label flags the conversation. Then Nora checks four things:

• Context. Does the flag hold up in the conversation? Read together after a price question, the four messages move the deal off the app for a discount.
• History. Has this seller done it before? A first attempt and a pattern across several buyers call for different responses.
• Prior actions. What's already been done? If the seller's last message was bounced with a reminder, this is a repeat, not a mistake.
• Policy. She applies your rules for off-platform payments and then acts or recommends an action.

You choose how much she does alone for each policy. In Auto-act, Nora takes the action herself, such as deleting the messages and banning the seller. In Approve first, she prepares the decision with her reasoning, and a moderator approves or overrides it. Many teams start with Approve first and switch to Auto-act once they trust the results.

Every decision goes into an audit log with who or what made it, so you can see why a message was blocked and handle an appeal. Nora resolves about 80% of flagged items, which leaves your team the hard cases like the coordinated rings and the new tricks that don't have a rule yet.

What to Configure This Week

• Add an AI rule for off-platform payment and contact sharing on DMs and listing comments.
• Bounce the first attempt. The message goes back to the sender with a short note that payments stay in the app, so both sides keep purchase protection.
• Send repeats to review, or shadow-block them so the message looks sent to the sender but never reaches the buyer.
• Act on the account when the pattern spans several conversations or accounts.
• Turn on OCR and PII masking so contact details in images and text are caught, too.
• Track two numbers: your false-positive rate and how many bounced messages get rewritten.

Where This Leaves You

The most expensive message in your marketplace probably isn't the abusive one. It's the polite one, offering a discount to pay elsewhere. A keyword list won't find it, because there's no bad word in it. Context will.

See how Stream's AI Moderation catches off-platform payment requests