EVALUATE

Evaluation

What the widget sees for one user. Use it to check that an experiment buckets the way you expect before pointing real traffic at it.

Send everything you know about the user, not just the bucketing field: the service picks which attribute buckets and which filter.

No result yet. This endpoint never fails: if the platform is unreachable every flag comes back with a null value and in_experiment false, which is indistinguishable from "no experiment applies".

Simulate traffic

Evaluate many random users at once and see how they split. One evaluation cannot tell a running experiment from a pending one — the value is the same either way. A distribution can.

A random value is generated per user for the hash attribute. It must match the experiment's hashAttribute — randomising the wrong one produces identical users and a 100/0 split that looks like a broken experiment.

Sent with every user. Use these to satisfy the experiment's targeting condition — without a match, nobody is bucketed.

Run a simulation to see how traffic splits. A running experiment shows its cohort weights; a pending or stopped one returns the flag's default to every user.

Backend vs SDK direct

The same evaluation through this service and through the GrowthBook SDK loaded here. Both run the identical algorithm over the identical payload, so the difference is the network hop to the backend — which is what centralising evaluation actually costs.

Must satisfy the experiment's targeting condition, or neither path buckets anyone and both are measuring the default.

Measures what this service costs: the same evaluation through the backend and through the GrowthBook SDK loaded here. The difference is one network hop.