Chapter 7: The Optimizer (BYOM)¶
Audience: media buyers, optimization engineers
The optimizer is Cat-Scan's automated optimization engine. "BYOM" stands for Bring Your Own Model: you register an external scoring endpoint, and Cat-Scan uses it to generate config change proposals.
How it works¶
Score Propose Review Apply
────────────> ────────────> ────────────> ────────────>
Your model Cat-Scan You (human) Google AB
evaluates generates approve or config is
segments config reject updated
changes
- Score: Cat-Scan sends segment data to your model endpoint. The model returns a score for each segment (geo, size, publisher).
- Propose: Based on scores, Cat-Scan generates specific pretargeting config changes (e.g., "exclude these 5 geos", "add these 3 sizes").
- Review: You see the proposal with projected impact. You approve or reject.
- Apply: Approved proposals are applied to the pretargeting config on Google's side. The change is recorded in the history.
Model management¶
Registering a model¶
Go to /settings/system and find the Optimizer section.
- Click Register Model.
- Fill in: name, model type, endpoint URL (your scoring service).
- The endpoint must accept POST requests with segment data and return scored results.
- Save.
Validating the endpoint¶
Before activating, test your model:
- Click Validate endpoint on the model card.
- Cat-Scan sends a test payload to your endpoint.
- Results show: response time, response format validity, score distribution.
- Fix any issues before activating.
Activating and deactivating¶
- Activate: the model becomes the active scorer for this seat.
- Deactivate: the model stops being used, but its configuration is preserved. Only one model can be active per seat at a time.
Workflow presets¶
When running score-and-propose, you choose a preset:
| Preset | Behavior | When to use |
|---|---|---|
| Safe | Only proposes changes with high confidence and low risk. Smaller improvements, lower chance of mistakes. | First time using the optimizer, or conservative accounts. |
| Balanced | Moderate confidence threshold. Good trade-off between impact and safety. | Default for most usage. |
| Aggressive | Proposes larger changes with higher potential impact. More risk of over-optimization. | Experienced users who monitor daily and can roll back quickly. |
Economics¶
The optimizer also tracks the economics of optimization:
- Effective CPM: what you're actually paying per thousand impressions, accounting for waste.
- Hosting cost baseline: your bidder's infrastructure cost, configured in the optimizer setup. Used to calculate whether savings from QPS reduction offset hosting.
- Efficiency summary: overall ratio of useful QPS to total QPS.
Configure your hosting cost at /settings/system > Optimizer Setup.
Reviewing proposals¶
Each proposal shows: - Segment scores that drove the recommendation - Specific changes to pretargeting fields (adds, removes, updates) - Projected impact on QPS, waste ratio, and spend
You can: - Approve: marks the proposal as accepted - Apply: pushes the approved changes to Google - Reject: discards the proposal - Check apply status: verify the changes took effect on Google's side
Related¶
- Pretargeting Configuration: the configs the optimizer modifies
- Conversions and Attribution: conversion data feeds into scoring quality
- Reading Your Reports: tracking optimizer impact