Outcome models
An Outcome is the result a client wants, such as a completed inquiry form, together with the rule for counting it. An Outcome model learns from one client’s campaign results for one Outcome, and has one job: to suggest the next audience.
What it produces
Section titled “What it produces”The output is a starting audience in Search, people most like the ones who inquired, with the signals that drove it listed as picks the team can read and change. It does not print a per-person prediction, a ratio or a model statistic to a client.
Before it trains
Section titled “Before it trains”- A campaign produces inquiries. Nothing trains until a campaign’s audience groups have inquiries from the ads.
- A person approves. Training never starts on its own. A person approves it before any model builds an audience.
- It is judged by the next flight. The suggested audience is judged by the next campaign’s cost per inquiry, against the audience groups the team built by hand in the same campaign.
What it learns from, and what it does not
Section titled “What it learns from, and what it does not”- Campaign inquiries. The people who inquired after being reached by a campaign are what it learns from.
- Not site traffic alone. People who did it on the client’s site without coming through an ad are context on the Outcome page, not what the model learns from.
- Comparisons within the campaign. Results are compared within the flight that produced them, never against the general population.
Boundaries
Section titled “Boundaries”- An Outcome model is stored with the client’s organization and used only for that client.
- No model is trained across clients.
- A client’s results never set a weight in a shared catalog model.
- A client’s site tag data never feeds the catalog and is never pooled.
- An organization can be flagged so that no model trains on its data at all.
The one way a result can reach the shared catalog is a request to re-judge a single model, filed under a rule that was registered before any data existed. A request does not change the model. It asks for the model to be tested again.