10 SynthID checks a day: which of 40 ad assets to submit

Spend the roughly 10 daily SynthID Detector checks on the final exports that carry the most risk, not on 10 random files out of 40. Press coverage of the October 7, 2026 public launch reports about 10 checks per user per day across images, video and audio, so a 40-asset launch needs a triage order or four days.

Google's announcement does not state the quota in the text I fetched, so confirm the number on synthid.com before you plan around it.

Why the quota exists and what it rules out

Reports say Google limits checks so repeated results cannot be used to build watermark-removal tools. That has two consequences for a team: you cannot treat the detector as a bulk QA step, and you should not script it. Coverage of the terms says they bar automating queries, which I could not confirm from Google's own page, so read the terms on the site yourself.

It also means a check is a scarce signal. Spend it where a hit or a miss would change what you do: a decision to disclose, to re-export, or to answer a client who asks. If the answer would not change either way, skip the check and save it for tomorrow's batch. A reviewer with their own account gets their own quota, so you do not need to pre-check every file for them.

A triage order

The detector only reads watermarks from Google and its named partners (OpenAI, NVIDIA, Kakao; Apple later). Coverage of the launch says it flags media from Google's Gemini, Imagen, Lyria and Veo models. So the files most likely to return a hit, and most worth confirming, are the ones from those families. Sume's video docs name Gemini Omni Flash 1.1 (gemini-omni-flash-1.1) among its models, so look up each clip's model id in its job record before deciding. A hit on a clip from a model outside the partner list would not be expected, and a miss there tells you nothing. Whether a given model's output carries a watermark is the vendor's claim to make, so read the vendor page rather than assuming.

Keep your own record, because the detector is not one

A detector result cannot replace a production record. For every delivered file, store the job id, the model id and the date, and note each edit step. Sume's trim doc says the default exact mode re-encodes the clip and returns a new MP4, so the delivered file differs from the generated one; the record should name both.

When a platform asks whether content is AI-made, answer from that record. A miss on the detector is not a reason to tick no.

• Day 1: the 10 highest-risk final exports.
• Day 2 onward: only files that changed after day 1.
• Always log date, file name, hit or miss.