Creative Score
A Creative Score is Klyo's reading of one finished creative against a rubric. One pass over the rendered video answers, signal by signal, what the rubric looks for, and the answers become a single number with a plain-language band. It sits beside the approve or reject decision in the Review console.
It is advice. Nothing in the approval or activation path reads it, so a low score never blocks an approval, a host, or a push. A reviewer stays the gate.
What works today
Checks run end to end: the per-campaign opt-in, the on-demand Run check button at Review, the four-pillar panel with per-signal evidence, and the score badge in the Review header. Two limits are shown on the panel rather than hidden: the four pacing signals are never evaluated, and the first-five-seconds signals are held to the timestamp the assessment cites. A separate rubric tuned to 9:16 shorts, a rubric you can edit yourself, and a bulk "check this whole campaign" action are not built.
The number and the bands
The score is the share of the signals Klyo could judge that came back detected: detected divided by evaluated, as a whole percent. A signal that could not be judged leaves the denominator entirely, so it never counts against the creative (see Not evaluated).
| Band | Score | What it says |
|---|---|---|
| Excellent | 80 and up | The creative carries most of what the rubric looks for |
| Might improve | 65 to 79 | Solid, with signals worth a second look |
| Needs review | below 65 | Read the signals before you approve it |
| Not evaluated | no number | Nothing could be judged, so there is no score to report |
Two more states appear on the badge while there is no number yet: Scoring during the first check, and Score unavailable when a run could not finish, with the reason on hover. A re-check on a creative that already has a number keeps showing that number, and says a new assessment is running on hover.
The four categories
The default rubric is Google's ABCD framework, which sorts its signals into four pillars:
- Attention: does the creative earn a look (text overlays, speech that supports what is on screen).
- Branding: are the brand and the product there, seen and heard, and early.
- Connection: are there people, faces, a face in close up.
- Direction: is there a call to action, spoken and written, and audio from the start.
The rubric holds 23 signals across those four. The panel shows one card per pillar with its detected-of-evaluated tally, and expands into the signal list.
The check is told what Klyo already knows rather than left to guess it: your Organization name as the brand, the names of the catalogue items in that combination as the products, the call to action Klyo itself wrote for the campaign, the campaign name, the format, and the campaign language. That is why the brand and product signals read as exact checks rather than guesses.
Running a check
A check happens two ways, and both do the same work.
- Every new creative of a campaign. Turn on Check every new creative in Campaign settings (the gear on the campaign row, and in the header of every builder step). Each render of that campaign is checked as it reaches Review. Off by default, because each check costs inference.
- On demand at Review. Press the run button in the Creative Score panel's header; its tooltip reads Run check, or Run again once there is a score. This is how a campaign that has not opted in gets one creative checked.
Requesting a check spends inference, so it needs write access: the Admin and Member roles can ask for one, Read only cannot. Reading a score that already exists is open to every member.
Checks are asynchronous. The panel switches to "Reading the video against the rubric" and the result appears on its own when it lands, no refresh needed. A run already in flight is not queued twice, so a second press costs nothing. The button is inert with the reason on hover when the creative has no finished video to read yet.
There is one score per creative. A re-score replaces it rather than stacking a history.
Reading the panel
The panel sits under the Approve and Reject buttons, so a reviewer sweeping the queue never has to scroll past 23 signals to reach a decision. A compact badge in the header carries the number and its band word.
- The tally line reads how many signals were detected of those evaluated, how many were not evaluated, and which rubric was used (ABCD (long form) is the one that ships).
- Collapsed, the panel keeps only the four pillar cards and their tallies. The chevron in its header expands the signal lists, and the s key does the same.
- Each signal carries a mark: detected, not detected, or not evaluated. A miss is amber rather than red, because the score is advice.
- Hover a signal name for what the assessment concluded, how confident it was, and its written explanation of the evidence it saw.
- When the creative has been re-rendered since it was scored, the old score stays visible marked From an earlier render, because a score pins the exact video it measured.
The evidence timecode
For a signal it found, the assessment cites the moment it saw it. That moment shows as m:ss beside the signal name, and clicking it jumps the video player to that second, so you can check the claim yourself instead of taking it on trust. A signal that was not found carries no timecode, because there is no evidence for an absence.
Not evaluated
A signal Klyo cannot honestly judge is recorded as not evaluated and left out of the number, rather than guessed at or counted as a failure. Two things land there.
- The four pacing signals (Dynamic Start, Quick Pacing, Quick Pacing in the first 5 seconds, Overall Pacing) all count shot changes, and Klyo does not measure shot changes. Google's own tool answers these from video shot-detection data rather than from a model, because asking a model to count cuts is not reliable. So they are always not evaluated, which leaves 19 signals as the most a check can evaluate, and the Attention pillar reads on two signals instead of five.
- A signal the assessment did not answer. A missing answer is unknown, not a miss.
The first five seconds
Eight signals ask whether something is there in the first five seconds of the video, the branding, product, people, face, and early-audio ones. Klyo holds those to the evidence: the assessment has to cite a timestamp, and the bound is checked in Klyo rather than trusted to the model. A signal claimed as detected at 0:08 is recorded as not detected, and so is one claimed with no timestamp at all. The signal's hover says which of the two happened, and the late timecode stays beside the signal, because there it is the reason for the refusal.
What a check costs
Each check reads the whole rendered video and bills one inference call, so a sixty-cell matrix is sixty video reads. That is why the campaign-wide opt-in is off by default and the per-creative button exists. Like the rest of Klyo's generation, the inference runs on your Organization's own Google Cloud project, so the spend lands there. Checks sit outside the matrix cost estimate, which prices renders, generated clips, and voiceovers only.
Where the rubric comes from
The rubric is Google's ABCD framework. Its 23 signal definitions, their literal evaluation criteria, and the instructions that keep the assessment from inventing evidence are transcribed from Google's abcds-detector, licensed under Apache-2.0, copyright 2024 Google LLC. That tool is not an officially supported Google product, Klyo's use of its rubric is not an endorsement by Google, and a Creative Score is Klyo's own reading of the framework.
What you can do today
- Turn on Check every new creative for a campaign in Campaign settings, or leave it off and request a check per creative.
- Press Run check on one creative in Review and watch the result land.
- Read the score, its band, and the four pillar tallies beside the approve or reject decision.
- Expand the signals to see what was detected, what was not, and the written evidence behind each one.
- Click a cited timecode to jump the player to that moment.
- See plainly which signals were not evaluated and why.