CATS: make an automated creative pool legible to the people funding it
The CATS case describes automated discovery across a broad gameplay pool. A human-readable concept map is still needed to understand what the system has actually learned.
The CATS case describes automated discovery across a broad gameplay pool. A human-readable concept map is still needed to understand what the system has actually learned.
What the public case establishes
TikTok's CATS: Crash Arena Turbo Stars case describes auto-selected creative covering vehicle combinations, battles and progression. It combines this supply with Smart+ purchase-value optimisation and automatic budget increases. Reported results therefore concern a package of changes rather than a standalone creative-selection intervention. TikTok for Business case study.
This is a source-led creative casefile, not a frame-by-frame viewing or a campaign we ran. The diagram below is our own schematic. It does not reproduce the advertiser’s footage, and we have not established reuse permission for that footage.
The decision worth examining
Automation can distribute attention efficiently without producing an explanation that a creative team can act on. If a report lists only asset identifiers and total return, producers cannot tell whether to create more combinations, more battle outcomes or more progression demonstrations. The missing work is a semantic map connecting the game promise to each file and the actual delivery it received.
Our proposed pool register marks each asset with a mechanic, a payoff and a production origin. Origin matters operationally because in-house, creator and generated material may have different permission and revision workflows. It is not a quality score. A creator clip and an in-house clip can test the same underlying argument and should remain comparable at that level.
A usable next check
Freeze a small, documented pool before the next creative exploration. Compare the platform's delivered mix with the mix available, then identify concepts that received little exposure. A manual follow-up can test a neglected but plausible concept under controlled conditions. Do not call an under-delivered asset a loser solely because the optimiser chose other material first.
| Test element | Proposed specification |
|---|---|
| Control | Asset-level reporting without concept families |
| Variant | The same pool with mechanic, payoff and origin labels |
| Readout | Available versus delivered concept coverage and quality by mature cohort |
| Stop or reject | Do not attribute combined budget and bidding changes solely to creative selection |
Download the populated case worksheet. Its control, variant and decision rules are a proposed experiment, not an account of results already measured. Keep the source URL with the worksheet when circulating it so reported findings cannot become unattributed team benchmarks.
Evidence boundary
We have not inspected the actual CATS ads or the campaign baseline. The vendor account's performance claims are not independently verified. This note does not repeat its developer attribution as a separately established corporate fact.
For background, see the creative brief is a hypothesis document and post campaign creative debrief process. No first-hand campaign result is asserted.
