How to Set Up Your Moderation for the 2026 Holiday Season
TL;DR
- New users and new scams show up during the holidays and require different configurations than in the fall.
- Act on repeat offenders first. Set a rule that applies to accounts that keep getting flagged, rather than just to individual messages.
- Check any rule that treats new accounts as risky, because many holiday users are new. Be sure to check both where links go and whether they include your brand name.
- Finish the changes by Friday, November 20th, then re-check each setting on Monday, January 11th, 2027, and revert it or keep it on purpose.
Most moderation teams plan for the holidays as a traffic problem, so they staff up for higher volume and assume the abuse will look roughly the same as it did in October.
Yes, the volume will surely arrive, but the bigger change will be who shows up and what they try once they're there. For example, gift recipients open their first account on Christmas morning, and scammers start copying your shipping notices because buyers expect a lot of them in December.
I’m the moderation product lead here at Stream, and I've been looking at moderation data across the platforms we work with this fall. A handful of patterns keep showing up. Below is a checklist of what I'd revisit before the holidays.
The Pre-Holiday Moderation Checklist
Note: Thanksgiving falls on Thursday, November 26th, and Black Friday traffic usually starts building up the week before, so every date below falls before that. None of this depends on a particular tool; most moderation setups can handle each item in some form.
| Area | What to revisit | Why it's important* | Do it by |
|---|---|---|---|
| Repeat offenders (do this first) | Make sure something acts on accounts that keep getting flagged, rather than only on individual messages. A temporary restriction is a good place to start. | Detection fires per message and bans fire per account, so nothing connects them unless a rule counts across messages. | Tuesday, November 17 |
| Flood and rate limits | Base thresholds on your busiest periods rather than your daily average. | Most companies' busiest day ran about 1.3 times a normal day, but event-driven apps hit 3 to 13.6 times. | Tuesday, November 17 |
| New-account rules | Soften rules that penalize new accounts for the holiday window, and let behavior make the case instead. | Many holiday users are first-timers, and rules keyed to account age treat them as risky. | Tuesday, November 17 |
| Brand impersonation | Judge links by where they actually lead, not by whether they contain your brand name. | Lookalike domains carrying your brand name pass an allowlist. | Wednesday, November 18 |
| Off-platform contact (marketplaces) | Decide whether it's allowed, and make sure your handling matches that decision. | On three marketplaces, off-platform signals averaged 11% of harm labels. | Wednesday, November 18 |
| Borderline content | Use lighter, reversible actions so it's contained while reviewers catch up, and save removal for clear-cut harm. | On average, two-thirds of automated actions were flags or shadow blocks, not removals. | Friday, November 20 |
| Live shopping video | Revisit how often you check video during streams, weighing cost against coverage. | Text is scored inline, but video is only covered by the frames you sample. | Friday, November 20 |
| Holiday settings cleanup | Review everything you loosened and either revert it or keep it on purpose. | Temporary settings tend to become permanent. | Monday, January 11, 2027 |
*Based on Stream moderation data from nine companies across marketplaces, gaming, prediction markets, fantasy sports, fintech and fundraising: about 3.2 billion AI moderation checks between October 10 and November 9, 2026.
The sections below go through each of these eight items, including how to best configure your moderation policy.
Count Flagged Messages Per Account
Detection fires on each message while bans fire on each account, so if nothing connects the two, an account can rack up hundreds of flags without ever tripping anything.
On the platforms I've reviewed, a small group of accounts usually produces a large share of what gets flagged, so a rule that acts on the account takes a lot of volume out of the queue before a reviewer ever opens it.
If you only make one change before the volume spike, I'd make it this one, because it takes a single rule and pays off more as volume goes up.
What to do: Make sure something in your setup looks at the account as a whole and acts once it crosses a set number of flagged messages in a short window, rather than judging each message on its own. A temporary restriction is a sensible place to start, since it stops a burst of bad behavior without permanently banning someone for one bad night.
Set Flood and Rate Limits Against Your Peak Minute
Most teams set their flood and rate limits once, based on an average day, and then leave them alone. That works until holiday traffic hits. A limit based on the average hour blocks real messages as soon as a sale starts. If you raise the limit to stop the complaints, it becomes too loose for the rest of the day, and a spam run can get through without tripping it.
The swing is usually bigger than teams expect. Across nine companies we moderate, the busiest day between October 10 and November 9 ran about 1.3 times a normal day for most of them. The ones built around live events were different: a fantasy sports app peaked at 3.1 times its median day, and a prediction market hit 13.6 times. A Black Friday sale behaves like a live event.
Detection itself usually isn't what breaks at peak. In our own benchmark, text moderation kept latency around 50 ms at more than 1,500 requests per second, and most modern detection holds up similarly, so the settings that tend to break are the thresholds you wrote yourself.
What to do: Pull your busiest minute from last year's holiday weekend, or from this fall's biggest event if you don't have last year's data, and set your flood and rate thresholds from that number. Before you ship the new values, check how they would have handled that hour.
Recheck Any Rule That Treats New Accounts as Risky
Account age is one of the most common trust signals, and for most of the year it works well. Most abuse comes from accounts with no history, so adding friction for new accounts catches a lot of it. December changes things, though, as gift recipients set up their first account and first-time buyers come for a single deal, and those are exactly the users your marketing budget paid to bring in.
New faces already show up in the flags on a normal day, so once the holiday influx lands on top of that, a rule keyed to account age ends up blocking real customers on their first day.
What to do: List every rule that uses account age or prior history as a condition, and for the holiday window, soften those rules so they send content for review instead of acting on it automatically. A new account sending the same link to 40 people will still look like spam on just behavior signals.
Watch for Scams That Copy Your Own Notices
Scams that convert the best in December look like they came from you. Buyers expect shipping updates and account notices this time of year, so a message that looks like one gets clicked.
These messages usually come from lookalike domains that carry your brand name, which is exactly what makes them convincing.
An allowlist won't help much here, because a rule that trusts any link containing your brand name will happily pass a domain like yourbrand-orders-help.com, which was built for exactly that purpose.
What to do: Judge links by where they actually lead, following any shorteners or redirects to the final destination, rather than by whether they mention your brand. It also helps to keep a list of lookalike spellings of your brand name so the obvious ones never reach your users.
Decide What to Do With Off-Platform Contact
Fee avoidance rises with transaction volume, since buyers and sellers who found each other once will try to close the next deal somewhere cheaper. Across the three marketplaces in our data, off-platform signals (platform bypass, personal info, and link labels) averaged 11% of harm labels, ranging from 8.5 to 16 percent.
The common failure is a queue full of off-platform flags that nobody acts on. Either the team decided off-platform contact was fine and never updated the action, or nobody made a decision and the queue filled up anyway, and both use up reviewer time during the season when you have the least of it to spare.
For example, Furnished Finder, a rental marketplace, saw conversations move to email, phone, and text before it added moderation.
What to do: Make the call before the season starts. If off-platform contact is allowed, stop flagging it so it doesn't fill the queue. If it isn't, stop the first attempt before it reaches the other person, ideally with a short note explaining that payment stays in the app so the buyer keeps purchase protection, and send repeat attempts to a reviewer.
Plan for the Queue After the Event
The heaviest moderation hours tend to come after the event, once people get back online to argue about it, and holiday staffing is thinnest at exactly that point. Think of the night after Thanksgiving, or the days around Christmas, when half the team is on PTO.
Shadow actions are the safety valve for those hours. A shadow block hides a message from everyone except the sender, who sees no error, so the harm stops spreading without anyone having to wait for a reviewer. Across eight of the companies in our data, about two-thirds of automated actions were flags or shadow blocks rather than removals, so most teams already lean on lighter actions at normal volume.
What to do: For borderline content, use an action that limits its reach right away and can be undone later, and save permanent removal for clear-cut harm. When reviewers are back, they can work through whatever was held and release anything that was caught by mistake.
Set Up Live Shopping Streams Separately From Chat
If you run live shopping, the video needs its own settings. Text gets scored inline before a message posts, but video can't work that way, so frames are sampled from the stream at an interval you choose, and that interval sets both your cost and your coverage. Sampling more often catches more at a higher cost, while sampling less often means a product shown for a moment can slip between frames.
In September, across five marketplace customers, we checked 482 million livestream video frames and flagged 19,135 of them, or 0.004%. Flags that rare are easy to miss if your sample interval is too long.
Speed usually isn't what limits you. In our benchmark, image and live-video requests return in about 40 to 50 ms at the API, and scoring a single frame takes about 300 ms, which is fast enough to act during the stream as long as the frame you needed was sampled.
What to do: Set the interval by stream type, with shorter intervals for large branded sales and longer ones for small seller streams, and check what that costs at holiday volume before you commit.
What to Watch in January
Although the holidays might be over, your work isn’t done. Returns and payment disputes follow the gifts, and chargeback-driven abuse follows the disputes, usually in the form of messages pressuring the other side to cancel a claim or settle outside the platform.
Then there's the cleanup, because every setting you loosened in December should either go back to normal or stay that way on purpose. I'd put a date on the calendar now, Monday, January 11th, and walk through each row of the checklist above against what actually happened.
Get Your Setup Checked Before Black Friday
If you'd like a second pair of eyes, we're running sample detection audits ahead of the holidays, where you send a sample of your own content and we show you what your current setup catches and what it misses. To set one up, email me at kenzie.wilson@getstream.io. You can also see how Stream AI Moderation works.