Our brand safety tool flags too many false positives on video. How do we fix it?
Start by finding out which rule is doing the blocking, because video false positives have a small number of causes and each one has its own fix. Most come from four places: keyword and metadata rules that react to words in a title or transcript, frame-by-frame scoring that judges a still image without the scene around it, a single safety floor applied to every brand, and default confidence thresholds set low against content that is rarely unsafe. Measure precision per category on a labeled sample of what the tool blocked, then fix the categories that are wrong most often.
Where do brand safety false positives on video come from?
Keyword and metadata rules. Many tools still decide from text: the title, tags, description, or words in the transcript. A cooking show that says "shoot the pan into the oven", a basketball recap full of "shots" and "attacks", and a news report about a wildfire all match blocklists written for violence. The rule fires on the word and never looks at the picture.
Topic without treatment. A news segment that reports on a war and a clip that glorifies violence share a topic. Brand suitability frameworks grade them differently, because the risk sits in how the content handles the subject. A tool that only knows the topic has to treat both the same way, and it usually picks the cautious answer.
Single frames judged alone. A classifier that samples a frame every few seconds and blocks the whole video when any frame scores high gives every frame a chance to misfire. One frame of a prop gun in a comedy sketch, a red sauce that reads as blood, or a news anchor standing in front of a protest image can sink an hour of inventory.
One floor for every brand. Brand safety is a floor no advertiser wants to run against. Brand suitability is a tolerance that differs per brand: a war documentary is fine for a streaming service's promo and wrong for a children's cereal. A tool that applies the strictest brand's settings to everyone blocks inventory that most of its brands would accept. Our guide to brand safety vs brand suitability covers the difference in depth.
Low thresholds against a low base rate. Moderation APIs return labels at fairly low confidence unless you set a cutoff. Amazon Rekognition, for example, returns moderation labels at 50% confidence or higher when no MinConfidence is given (AWS). Truly unsuitable content is also rare in most premium video. When the thing you are looking for is rare, even an accurate classifier produces more wrong flags than right ones, because its small error rate applies to the large pool of safe content. The arithmetic is worked through in why content moderation flags so many false positives.
How do we measure our false positive rate?
The tool's dashboard reports what it blocked. It cannot tell you how much of that was wrong, so build the answer from a sample.
1. Pull a random sample of blocked impressions or segments, a few hundred per category if the volume allows, with the video and timestamp for each. 2. Label each one against your own written suitability policy, per brand or per brand tier. Two reviewers per item, with disagreements settled by a third, keeps the labels honest. 3. Compute precision per category: of everything blocked for "violence", the share that was truly unsuitable. A category with low precision and high volume is where the lost inventory is. 4. Sample what passed as well, at a smaller rate, so a fix that cuts false positives can be checked for misses it introduced. 5. Keep a fixed test set and rerun it after every change. IAB Tech Lab publishes media content human-labeled against the GARM standard for this kind of testing (Brand-Suitability-Test-Benchmarks), which is a useful start before you have your own.
Ask your vendor for per-category scores and the segment timestamps behind each block. A vendor that returns only pass or fail leaves you unable to do steps 3 and 4.
Which fix matches which cause?
| What you see in the sample | Likely cause | Fix |
| Sports, cooking and news blocked for words like "shoot", "attack", "fire" | Keyword or transcript rules | Score the scene with picture, speech and on-screen text together, and drop word-only rules or demote them to a review signal |
| News and documentaries blocked alongside genuinely harmful clips | Topic without treatment | Classify treatment as well as topic (reported, depicted, glorified) and set tiers on that |
| Long videos blocked for a single moment | Frame-level scoring, video-level verdict | Score segments, report the timestamp range, and block the break near that segment instead of the whole asset |
| One brand's concerns blocking inventory for every brand | A single shared floor | Keep one floor for content nobody accepts and set suitability tiers per brand or per campaign |
| Many low-confidence flags across categories | Default thresholds | Set a threshold per category from your labeled sample and route a middle band to human review |
| The same publisher's content blocked again and again | No memory of past decisions | Keep reviewed allowlists for trusted publishers and shows, and feed review outcomes back into the thresholds |
Is GARM still the standard for brand suitability?
The Global Alliance for Responsible Media is gone as an organization. The World Federation of Advertisers announced on 9 August 2024 that GARM would discontinue its activities (WFA). Its Brand Safety Floor and Suitability Framework still shapes how the industry names sensitive categories and risk tiers, and the IAB Tech Lab benchmark set above is labeled against it. For describing what content is about, the current IAB Tech Lab standard is Content Taxonomy 3.1 (IAB Tech Lab). When a vendor says its categories are "GARM-aligned", ask which version of the framework it maps to and who maintains that mapping now.
Where should the classification run?
Run it where the video is. A verification tag in the ad slot sees the page or the app, and for CTV it often sees little more than an app ID. The publisher, SSAI service or FAST channel operator holds the video files, so it is the party that can score each segment and pass a reliable signal into the bid request or the post-campaign report. Our guides on CTV bid requests with no content category and finding out what content CTV ads ran next to cover how that signal travels.
How does Mixpeek handle this?
Mixpeek reads your files where they already live, in S3, GCS or any bucket, and turns what's in them into data your software can search, classify and moderate. For brand suitability that means each video is split into scenes, and each scene is indexed with what is in the frames, what is said, and the text on screen. Taxonomy classification assigns your categories and tiers against that index, at the scene level and with timestamps. Because the analysis is stored, a new brand's tolerance or a stricter campaign setting is a new query over the same index, without paying to analyze the video again.
It fits publishers, SSAI and FAST operators, and platforms that hold the video and need per-segment signals. It is the wrong tool for a buyer that only sees bid requests and has no access to the video files. Plans are a flat monthly fee with usage billed on top; current rates are on pricing. The brand safety verification use case shows the workflow, and the content moderation tools list compares the APIs teams use for the classification step.
Frequently Asked Questions
Why does our brand safety tool block news content?
Most tools judge topic and keywords, and news covers the topics brand safety lists were written for: conflict, crime, disasters, death. A report about a shooting and a video that depicts one share those words and that topic. Unblocking news safely means grading treatment (reported, depicted, glorified) as well as topic, scoring the segment rather than the headline, and giving each brand a tier for news instead of a blanket block.
What is a false positive in brand safety?
A false positive is content the tool blocks or downgrades that your own suitability policy would have accepted. It is measured against your policy, so the same block can be correct for one brand and a false positive for another. The useful number is precision per category: of everything blocked in that category, the share that was truly unsuitable.
Should we lower our brand safety thresholds or switch vendors?
Measure first. Pull a labeled sample of blocked content and compute precision per category. If one or two categories cause most of the wrong blocks and the vendor exposes per-category scores, a threshold change per category may be enough. If the sample shows keyword hits, whole videos blocked for single frames, or news blocked as a topic, the cause is how the tool classifies, and a threshold change only trades false positives for misses.
How do we reduce false positives without letting unsafe content through?
Keep a fixed, labeled test set that includes truly unsafe content, and rerun it after every change, tracking both precision and the share of unsafe items caught. Send a middle confidence band to human review instead of auto-blocking it, keep a strict floor for content no brand accepts, and loosen suitability tiers only for the brands that want the inventory.
Can brand safety be checked per scene instead of per video?
Yes, when the classifier runs on the video itself. Scene-level classification gives each segment its own categories and a timestamp range, so a single unsuitable moment affects the ad breaks near it and leaves the rest of the asset available. CTV impressions can then be joined to the scene that aired before the break using the VAST content ID and playhead macros.