This is the complete reference for the video rule types in Cloudinary Moderation. Video rules analyze a video over its whole duration - sampling frames from the footage, and, in some cases, the audio - and give the video a single overall result.
All rules share the common settings described in "Create a rule" (name, description, Review threshold, optional rejection reason). Video rules add one recurring setting worth understanding up front:
Frames per second to sample - how many frames per second of footage are analyzed. Higher sampling catches brief appearances more reliably but takes longer to process. The defaults are a good balance for most content.
Video Content Detection
What it checks: what the video shows, against a requirement you describe in your own words - for example, "The video must not show alcoholic drinks" or "The video must be live-action, not animated." Moderation checks one frame per second and uses AI to confirm what it finds.
Settings:
What the video must or must not show - describe your requirement in a sentence, then click Prepare rule. Moderation turns your requirement into the checks it runs on the video's frames and shows a summary under What this rule checks. Click Prepare rule again whenever you change the requirement - the rule can't be saved until it's prepared.
Add examples to broad terms - on by default. Adds examples to broad categories, such as weapons, so the AI recognizes them.
Advanced settings - optional. Shows the values Prepare rule filled in - for example, whether the video passes when the content is Found or Not found, and whether One frame is enough or Most frames must match - so you can fine-tune them. Running Prepare rule again replaces your edits to these values.
Good to know: rules created with the earlier version of Video Content Detection, which matched a list of Labels to detect, show Unknown as their rule type on the Rules page, and their settings can no longer be edited. They can't be used in new runs or automations. To keep checking for the same content, create a new Video Content Detection rule and describe it as a requirement.
When to use it: spotting competitor products, alcohol, or other unwanted items in footage - or confirming your product actually appears in creator-submitted videos.
Video On-Screen Text Check
What it checks: the text that appears on screen (captions, overlays, signs, packaging), evaluated against a yes/no question you write.
Settings:
Question - a yes/no question about the on-screen text, for example: "Does the video contain profanity or offensive language?"
Pass this rule when the answer is Yes - choose Pass on Yes or Pass on No, depending on whether your question describes acceptable or unacceptable text.
Frames per second to sample
When to use it: policing burned-in captions and overlays - profanity, prohibited claims, outdated promotions, competitor mentions.
Video Transcript Check
What it checks: what is said in the video. The audio is transcribed, and the transcript is checked against a yes/no question you write.
Settings:
Question - a yes/no question about the spoken content, for example: "Does the transcript contain profanity or offensive language?"
Pass this rule when the answer is Yes - Pass on Yes or Pass on No.
No speech is detected in the asset's audio - what should happen when the video has no spoken content: Approve (the default), Needs Review, or Reject.
No audio transcript file is detected - what should happen when no transcript could be produced for the video: Needs Review (the default), Approve, or Reject.
When to use it: spoken claims and compliance - required disclaimers, forbidden phrases, offensive language in voiceovers or interviews.
Good to know: this rule listens to the audio; it doesn't look at the visuals. Pair it with Video Content Detection or Video On-Screen Text Check for full coverage.
Video Text Metadata Match
What it checks: whether the text shown on screen matches the asset's own metadata - for example, that the product name burned into the video matches the product name in your metadata.
Settings:
Question - a yes/no question comparing on-screen text to metadata. Insert metadata fields using the field chips (or type
#followed by the field name), for example: "Does the on-screen text match #product_name?"Metadata fields - select which metadata fields Moderation may compare against the video's text.
Pass this rule when the answer is Yes - Pass on Yes or Pass on No.
Asset's metadata field is empty - what should happen when the selected metadata field has no value: Needs Review (the default), Approve, or Reject.
No text is detected on the asset - what should happen when no on-screen text is found in the video: Needs Review (the default), Approve, or Reject.
Frames per second to sample
When to use it: catching mismatches between creative and catalog data at scale - wrong product names, prices, or campaign codes in localized video variants.
Video Transcript Metadata Match
What it checks: whether what is said in the video matches the asset's own metadata - for example, that the narration mentions the product name stored in your metadata.
Settings:
Question - a yes/no question comparing the spoken content to metadata. Insert metadata fields by typing
#followed by the field name, for example: "Does the narration mention #product_name?"Pass this rule when the answer is Yes - Pass on Yes or Pass on No.
Asset's metadata field is empty - what should happen when the selected metadata field has no value: Needs Review (the default), Approve, or Reject.
No speech is detected in the asset's audio - what should happen when the video has no spoken content: Approve (the default), Needs Review, or Reject.
No audio transcript file is detected - what should happen when no transcript could be produced for the video: Needs Review (the default), Approve, or Reject.
Good to know: this rule compares the audio transcript to metadata. To compare the on-screen text to metadata instead, use Video Text Metadata Match.
When to use it: verifying spoken content against catalog data - product names, campaign claims, or disclaimers that must be said in each video variant.
Reading video results
Whatever the rule, a video gets one overall result: Approved, Rejected, or Needs review. The rule's explanation tells you what was found, and when it refers to a specific moment it includes a clickable timestamp that jumps the player straight there. A brief appearance is enough to trigger a rule - the result reflects the worst moment in the video, not the average.
Things to keep in mind
Processing time: video is analyzed frame by frame, so runs with many or long videos take noticeably longer than image runs.
Limits: rules that look at the frames check up to 600 frames per video - about 10 minutes of footage at 1 frame per second, or 5 minutes at 2. Longer videos, and video files over 500 MB, end with an error for that rule instead of a result.



