Marketplace Moderation Statistics: What 550M+ Checks Show
Harmful content on an online marketplace shows up in very different places at very different rates.
This post uses moderation data from Stream's Moderation API to show where it appears most often across live video, speech, chat, photos, and text.
Below are the key findings, a comparison across all five content types, and how marketplace teams can use these numbers.
Key Findings
- Seller event photos have the highest flag rate of any content type at 1.4% on average (1 in 73). Livestream video frames have the lowest, at 0.004% (1 in 25,000).
- 97% of flagged seller event photos contain sexual or nudity content.
- Live chat messages are flagged at 0.8% on average (1 in 124). That is the second-highest rate and about 40 times the rate for livestream speech (0.02%).
- Threats and hate speech each account for 38% of flagged livestream video frames. Sexual or nudity content accounts for 18%, self-harm for 4%, and weapons or violence for 1%.
- Vulgarity (5,054 flags) and racism (4,739 flags) are the most common categories in flagged livestream speech. Sexual harassment (3,335) and insults (3,260) follow.
- Links were 56% of all harm labels on seller event text (417 of 744), and 231 of those were forbidden links. Other labels include vulgarity (92), hatred (59), personal info (28), and off-platform deals (15).
- Livestream flags went to human review 100% of the time, and none were removed automatically. In live chat, 35% of automated actions were removals, and 32% of seller event flags were removed automatically.
What is Marketplace Moderation?
Marketplace moderation is the process of checking user-generated content on an online marketplace and acting on content that breaks platform rules. This includes seller photos, listing text, chat messages, and live video. Checks can be automated, done by human reviewers, or a mix of both.
On a marketplace, moderation also covers scams, prohibited items, off-platform payment requests, and spam links.
Methodology
This analysis uses moderation data from five marketplace customers using Stream's Moderation API. It covers September 1st - October 1st, 2026, and includes more than 550 million checks across five content types: 1) livestream video frames, 2) livestream speech, 3) live chat messages, 4) seller event photos, and 5) seller event text.
Definitions
The terms below are used throughout this post:
- Checked: An item that was run through moderation. For livestream video, an item is a single frame. For livestream speech, it is a speech segment. For chat, it is one message. For seller event content, it is one photo or one text submission.
- Flagged: An item that matched at least one moderation category, rule, or blocklist entry, such as sexual content, hate speech, or a forbidden link.
- Flag rate: Items flagged divided by items checked.
- Sent to review: A flagged item placed in a queue for a human moderator to decide on.
- Auto-removed: A flagged item removed without a human decision.
- Blocklist hit: An item that contained a word or phrase on a platform's blocklist.
- Rule hit: An item that triggered an automated rule, such as the spam rule or the anti-harassment rule.
- User-reported abuse: Content that a user reported through the platform, separate from automated flags.
Side-By-Side Comparison: Content Type
Seller event photos have the highest flag rate at 1.4%, followed by live chat at 0.8%. Livestream video frames have the lowest at 0.004%.
| Content Type | Items Checked | Items Flagged | Flag Rate | Auto-Removed | Sent to Review |
|---|---|---|---|---|---|
| Seller event photos | 78,543 | 1,075 | 1.4% | 32% | 68% |
| Live chat messages | 18.6M | 149,471 | 0.8% | 35% | 65% |
| Seller event text | 347,331 | 571 | 0.16% | 32% | 68% |
| Livestream speech | 57.9M | 10,567 | 0.02% | - | - |
| Livestream video frames | 482M | 19,135 | 0.004% | - | - |
Statistics by Content Type
The sections below cover each content type in order of flag rate, from highest to lowest.
Seller Event Photos
Seller events are live selling sessions hosted by marketplace sellers. Sellers upload photos to promote them. Here's how the moderation metrics break down:
| Metric | Value |
|---|---|
| Photos checked | 78,543 |
| Photos flagged | 1,075 |
| Flag rate | 1.4% |
| Flagged as sexual or nudity content | 97% |
| Flagged for other reasons | 3% |
| Auto-removed (photos and text combined) | 32% |
Almost every flagged photo was sexual or nudity content.
Since photo flags are the highest of any content type in this data, seller uploads carry more risk per item than the live video itself.
Live Chat
Live chat is the messaging that runs during a live selling session, where participants send messages to each other in real time. The table below shows the metrics:
| Metric | Value |
|---|---|
| Messages checked | 18.6M |
| Messages flagged | 149,471 |
| Flag rate | 0.8% |
| Automated actions | 166,359 |
| Automated actions that flagged a message | 108,072 |
| Automated actions that removed a message | 58,287 (35%) |
Live chat moderation hits came from four sources, shown below, totaling 34,807. The spam rule had the most hits at 19,637 (56.4%), followed by blocklists with 9,027 (25.9%), the anti-harassment rule with 3,918 (11.3%), and user-reported abuse with 2,225 (6.4%).
Blocklist hits were mostly profanity and sexual terms. These are hits, not messages, so one message can trigger more than one source.
Live Chat Moderation: Detection Sources
Seller Event Text
Seller event text is the written content sellers submit for their live selling sessions, such as event titles and descriptions. These are the moderation numbers:
| Metric | Value |
|---|---|
| Text submissions checked | 347,331 |
| Text submissions flagged | 571 |
| Flagged by AI harm detection | 470 |
| Flagged by blocklist | 101 |
| Flag rate | 0.16% |
| Auto-removed (photos and text combined) | 32% (506 of 1,589 seller event flags) |
Of the 571 seller event text submissions flagged, 470 were flagged by AI harm detection and 101 by the blocklist. The 470 AI-flagged submissions carried 744 harm labels in total.
Links accounted for 417 of them (56%), and 231 of those were forbidden, including off-platform domains and competitor links. Vulgarity followed with 92 labels (12.4%), then hatred with 59 (7.9%), sexually explicit content with 47 (6.3%), personal info with 28 (3.8%), ads with 18 (2.4%), and off-platform deals with 15 (2.0%). The remaining 68 labels (9.1%) fall in other categories.
Links were about 4.5 times as common as the next category.
Seller Event: Flags Per Category
Seller text is flagged far less often than seller photos, but the categories differ from the other content types. Links, personal info, ads, and off-platform deals are all ways to move a buyer away from the platform or to promote something outside it.
For example, Furnished Finder, a rental marketplace, saw many landlord and renter conversations move to email, phone, and text before it added Stream Moderation, which left its team unable to see whether those conversations were productive or legitimate.
Livestream Speech
Livestream speech is the spoken audio during a live selling session. Here's how the numbers break down:
| Metric | Value |
|---|---|
| Speech checks | 57.9M |
| Speech checks flagged | 10,567 |
| Flag rate | 0.02% |
Of the 10,567 livestream speech checks flagged, 5,054 (47.8%) were flagged for vulgarity and 4,739 (44.8%) for racism. Sexual harassment accounted for 3,335 flags (31.6%), insults for 3,260 (30.8%), sexually explicit content for 1,678 (15.9%), hatred for 1,166 (11.0%), and threats for 752 (7.1%). Vulgarity and racism were each present in nearly half of flagged speech checks, and threats were the least common category.
Livestream Speech: Flags Per Category
Livestream Video Frames
Livestream video is the live video feed of a seller's selling session. This table shows the metrics:
| Metric | Value |
|---|---|
| Frames checked | 482M |
| Frames flagged | 19,135 |
| Flag rate | 0.004% |
Of the 19,135 livestream video frames flagged, threats and hate speech each accounted for 38%, so together they made up 76% of all flagged frames. Sexual or nudity content accounted for 18%, self-harm for 4%, and weapons or violence for about 1%.
Livestream Video Frames: Share of Flagged Frames
Live video is the content type that gets the most attention, and it produces the fewest flags relative to volume. Seller event photos are flagged about 340 times as often.
Human Review vs. Automated Action
Handling differs by content type. Flagged livestream video and speech always went to human review, while seller event content and live chat had automated removals.
| Content Type | Handling | Share |
|---|---|---|
| Livestream video frames | Sent to human review | 100% |
| Livestream speech | Sent to human review | 100% |
| Seller event photos and text (combined) | Auto-removed | 32% |
| Live chat | Removals as a share of automated actions | 35% |
Livestream video frames and speech had no automatic removals. All 19,135 flagged frames and all 10,567 flagged speech checks went to a human moderator.
Seller event content and live chat used automation:
- About a third of seller event flags (32%) were removed automatically.
- In live chat, 166,359 automated actions were taken, and 58,287 of them (35%) were removals.
- The other 108,072 (65%) flagged a message without removing it.
The content types that were reviewed by people are also the ones with the lowest flag rates:
- Livestream video frames (0.004%) and speech (0.02%) produce relatively few flags, so a review team can look at each one.
- Live chat (0.8%) produces 149,471 flags, and seller event photos (1.4%) have the highest rate, so these are the content types where automation was used.
How Marketplace Teams Can Use These Numbers
Check Seller Photos When They Are Uploaded
Seller event photos had the highest flag rate at 1.4%, and 97% of flagged photos were sexual or nudity content. Sellers upload photos before an event starts, so a platform can check them before any buyer sees them. This gives photo checks more impact than live video checks, since livestream video frames were flagged at 0.004%.
Plan for Live Chat as the Highest-Volume Surface
Live chat produced 149,471 flags at a rate of 0.8%, about 40 times the rate for livestream speech, so it's where automated rules save the most reviewer time. Spam was the most common automated rule hit (19,637), about five times the anti-harassment rule (3,918). A spam rule is a practical first rule to set up for live chat. User reports (2,225) were a small share of detections, so relying on reports alone would miss most flagged messages.
Use Link Rules for Seller Text
Links were the largest flag category in seller event text with 417 flags, and 231 of those were forbidden links. Personal info (28), ads (18), and off-platform deals (15) also appeared. A link rule catches more of this content than any other single rule.
Add Misspellings to Blocklists
Seller event blocklist hits included "18+", "xxx", and the misspelling "whor3s." A list of correctly spelled words will miss variants like this. Include common misspellings and character substitutions, such as numbers in place of letters.
Match the Review Method to the Volume
Flagged livestream video and speech were all sent to human review, and none was removed automatically. Both have low flag rates (0.004% and 0.02%), so a review team can look at every flag. Live chat and seller event content produced more flags, and both used automated removals (35% of automated actions in chat, 32% of seller event flags). Platforms can use flag volume as a starting point when deciding which content types to automate.
Frequently Asked Questions
Can AI moderate a livestream in real time?
Yes. Stream's live video and speech moderation transcribes livestream audio in real time and checks each speech segment against harm categories like hate speech and threats, while analyzing periodic video keyframes for visual harms like nudity and violence.
Teams can send violations to a human review queue or act automatically, for example by issuing a warning, muting audio or video, blurring video, or removing a user.
How fast does moderation need to be for live shopping?
Each surface has a different window. Live commerce chat generally needs moderation decisions in under 100 milliseconds so checks don't disrupt the shopping experience. Stream's text moderation runs in under 50 milliseconds, fast enough to check each chat message before other participants see it.
Seller photos and event text are uploaded before an event starts, so they can be checked before the event is published and before any buyer sees them. Livestream speech is checked segment by segment as it's transcribed, and video keyframes every few seconds, so violations can be flagged or acted on during the broadcast.
How do you keep live shopping moderation from blocking legitimate sales?
Match the action to the content so that not every flag results in removal. Platforms can also check seller photos when they are uploaded, before an event starts, so problems are caught before buyers are in the session.