Distribution
By default a group splits the budget from its learned forecast model alone. Distribution tilts that split toward recent results — useful when the model is still young, or when the market moved faster than the model did.
It lives on Global settings → Advanced → Distribution, and it is three choices that go together.
How to use it
- Pick a basis — the recent signal to lean on.
- Pick an influence — how much modelled improvement you will trade for that lean.
- Pick a time window — how far back the signal is measured.
Leave the basis on Model only to run on the forecast alone. A basis with no influence or window blocks the save, and an influence or window with no basis blocks it too: set all three, or none.
Basis
| Basis | Leans toward |
|---|---|
| Model only (default) | Nothing — the forecast model decides |
| Recent impressions | Campaigns currently winning impression share |
| Recent clicks | Campaigns currently winning click share |
| Platform ROAS | Campaigns with stronger realized platform ROAS |
| Recent profit | Campaigns with positive realized contribution profit |
| PA-ROAS | Campaigns with stronger realized True ROAS |
Influence
Influence is a three-step slider, and each step is the floor on how much of the forecast model's expected gain you keep:
| Influence | Keeps at least | Reads as |
|---|---|---|
| Small | 90% of the modelled gain | A gentle tilt |
| Medium | 70% | A real shift |
| Large | 40% | Recent evidence dominates |
A larger shift follows recent results more, and the forecast model less. It does not make the group more aggressive with money — only more reactive about which campaign gets it.
Time window
| Window | Usable on a live group | Notes |
|---|---|---|
| Latest completed hour | No | Preview only |
| Today through latest hour | No | Preview only |
| Rolling 3 hours | Yes | Delivery signals only |
| Rolling 24 hours | Yes | Any basis |
| Last 7 days | Yes | Any basis |
Platform ROAS, Recent profit and PA-ROAS read money, which needs time to settle, so they accept 24 hours or 7 days only. Impressions and clicks are delivery signals and work on any live window.
Reading the effect
In the Allocation panel, hover a campaign. The tooltip separates:
- Model-only — what the objective and constraint alone would have given it.
- Preference wants — what the distribution basis alone would have given it.
- Proposed — the number the solve actually landed on, between the two.
If those first two are far apart, your influence setting is doing real work. If they are nearly identical, the basis is agreeing with the model and you can leave it on Model only.