> For the complete documentation index, see [llms.txt](https://docs.churned.io/churned-docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.churned.io/churned-docs/measure-impact.md).

# Measure impact

See churn, health and revenue move per audience, per test and per Co-pilot

Everything in Churned ends here: did the audience, the workflow and the Co-pilot change what customers do? This is step 5 of the [Churned journey](/churned-docs/getting-started-with-churned.md). Three views answer the question at different levels.

## Three levels of measurement

| Level      | Question                                                                                                                                  | Where                                                                |
| ---------- | ----------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------- |
| Base       | Is the customer base healthier than a period ago? Health level movements, churn per health level, sentiment, engagement and ARR over time | [Impact Dashboard](/churned-docs/measure-impact/impact-dashboard.md) |
| Experiment | Did treated customers churn less than the control group? Churn rate, churn loss and average loss per group, with a winner                 | [A/B Testing](/churned-docs/measure-impact/a-b-testing.md)           |
| Co-pilot   | Which action works for which customers, and how much uplift does the model create over the default action?                                | Copilots page and [Co-pilot](/churned-docs/co-pilot.md)              |

## How uplift is measured

* **Random splits** create the control. A workflow branch that receives no action, or the current default action, is the baseline every other branch is compared with.
* **The Co-pilot keeps a permanent exploration group.** A small share of each audience, typically 5 percent, is always assigned at random. It acts as a rolling control group and as the source of unbiased data for retraining. The uplift of the Co-pilot is the difference in outcome between customers who received the model's chosen action and this exploration group.
* **Outcomes are the same metrics as elsewhere.** Churn is the [churner definition](/churned-docs/metrics-and-definitions/metrics/subscription-kpis.md#churners) with the one-week grace period; value is [ARR](/churned-docs/metrics-and-definitions/metrics/subscription-kpis.md#arr); reactivation is a new subscription after churn. A test's churn rate is churners divided by contacts in the group.
* **Read results per segment.** Both the A/B view and the Co-pilot view break results down by channel, product, client type or engagement, because an action rarely works equally for everyone.

The page [How A/B tests train the next best action](/churned-docs/measure-impact/how-a-b-tests-train-the-next-best-action.md) describes the full loop from random split to model retraining, and [Churn prediction goals](/churned-docs/measure-impact/churn-prediction-goals-its-not-about-accuracy-its-about-impact.md) explains why impact, not model accuracy, is the goal.

## Per vertical

The outcome metric changes with the business model: churn and ARR for subscriptions and structural donors, repeat orders and retention rate for e-commerce and non-structural donors, reactivations for win-back audiences. The [Metrics vocabulary](/churned-docs/metrics-and-definitions/metrics.md#vocabulary) maps the names.