The Short Answer
Brand safety and brand suitability are different decisions, made at different granularities. Brand SAFETY is a floor: content no advertiser should ever appear against (illegal content, terrorism, CSAM), binary, universal, non-negotiable. Brand SUITABILITY is a per-brand tolerance curve: legitimate content that is fine for one advertiser and wrong for another, a war documentary is acceptable inventory for a streaming service and unacceptable for a children's cereal, so it needs tiered risk classification (high/medium/low) across sensitive categories, not a yes/no. The GARM Brand Safety Floor + Suitability Framework remains the industry's de-facto taxonomy for both: the organization behind it wound down in 2024, but the framework's category and tier definitions are still what platforms, verification vendors, and IAB Tech Lab's human-labeled benchmarks classify against.
What does the framework actually classify?
The GARM-style taxonomy defines roughly a dozen sensitive-content categories (adult content, arms and ammunition, crime, death and injury, online piracy, hate speech, military conflict, obscenity, illegal drugs, spam, terrorism, and sensitive social issues) and, for each, a floor (never monetize) plus three graded risk tiers. The tier is a judgment about TREATMENT, not just topic: news reporting on a shooting, a documentary about one, and glorification of one are the same category at three different risk levels. That treatment-sensitivity is what makes suitability classification genuinely hard for machines: topic detection alone over-blocks journalism (starving news publishers of revenue, the classic false-positive failure) and under-blocks glorification.
How does AI classify video for suitability?
Text pages are one signal; video is four. Production systems decompose each video into parallel channels and classify their combination:
| Signal | What it carries | Example failure if skipped |
|---|---|---|
| Visual frames/scenes | Violence, weapons, nudity, alcohol, distressing imagery | A muted fight scene passes an audio-only check |
| Speech (ASR transcript) | Profanity, hate speech, drug references, tone of discussion | A calm-looking talking-head video discussing self-harm |
| On-screen text (OCR) | Slurs in memes, captions, lower-thirds | Text-in-image hate content invisible to visual classifiers |
| Metadata + context | Title, description, channel history, adjacency | A compilation channel whose individual clips look benign |
Each channel produces category scores; a fusion step maps the combination onto floor/high/medium/low per category; per-brand profiles then threshold those tiers (the cereal brand blocks medium-risk conflict; the streaming service allows it). For video specifically, ADJACENCY matters as much as content: what plays in the seconds before and after an ad slot, which is why scene-level classification with timestamps beats whole-video labels. The decomposition pipeline is the same architecture as multimodal content moderation generally; the difference is graded output instead of binary policy calls.
Why can't I just use a moderation API for suitability?
Three mismatches. Moderation APIs return POLICY categories (violates/doesn't), not per-brand risk tiers: you can't express "medium-risk debated social issues are fine for us" through a binary endpoint. They classify at upload time against one static ruleset, while suitability standards and brand tolerances change, reclassifying a large catalog usually means re-paying full analysis (systems that classify at query time over an already-indexed catalog sidestep the reprocessing bill). And most operate per-file, while suitability for video advertising needs scene-level timestamps for adjacency decisions. If your buyer is a trust-and-safety team, a moderation API or NSFW detector fits; if your buyer is an ad platform, publisher, or agency, you're in suitability territory: the vendor conversation is different (see the best ad-tech AI platforms).
How does this work on Mixpeek?
Mixpeek treats suitability as a classification problem over an indexed catalog rather than a per-upload verdict. Videos are decomposed at ingest (scenes, transcripts, on-screen text, faces, objects) into a searchable index; taxonomy classification then assigns categories and tiers against YOUR taxonomy (GARM-style or bespoke), and because classification runs over stored features, changing a threshold or adding a category re-scores the catalog without re-running extraction. The same index answers adjacency queries ("what is in the 10 seconds around every mid-roll slot?"), powers video compliance monitoring, and stays searchable for everything else the archive is for: suitability becomes one taxonomy among several rather than a separate vendor silo. Rates for the underlying indexing are on pricing; the classify-at-query-time pattern is documented in the NSFW list's economics FAQ.
Screening footage before it becomes training data is a related job: how to keep footage of minors out of a video training dataset.