> ## Documentation Index
> Fetch the complete documentation index at: https://www.pagent.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Variations and goals

> How pagent handles several variations, several goals, revenue and the traffic split check

## Which goal decides a test

Only the test's **primary goal** decides a test. You choose it when you [add goals to a test](/docs/guides/tracking/conversion-goals#add-goals-to-a-test).

* **Several primary goals.** pagent counts a visitor as converted when they complete any of them. Each visitor counts once.
* **Goal direction.** For a goal set to **Lower is better**, such as cart abandonment, a lower rate is a win. If the primary goals mix directions, the test is judged as **Higher is better**.
* **Conversions are counted per visitor.** A visitor who converts three times is one conversion for the decision.

### Other goals

Every other goal of a test is analysed and shown in the report on its own, with its own direction. These analyses help you understand the effect, but they never stop a test or change its result.

### Guardrail goals

Guardrails are safety metrics for your website, such as returns or support contacts. They are reported separately from the tested goal, with the chance that the variation is better on each. A variation can win the primary goal and still hurt a guardrail.

Guardrails never stop a test. Check them before you ship a winner.

### Exit rate

pagent also shows how often visitors leave the site from the tested page, as a diagnostic. Lower is better. It never stops a test.

## Revenue

For goals that track revenue, pagent reports **revenue per visitor** and **average order value**, with a 95 % interval and the chance that the variation is better. Very large orders are capped so a single big order does not swing the result; you can change this with **Outlier capping** in the goal's [advanced settings](/docs/guides/tracking/conversion-goals#advanced-settings). The Bayesian engine also shows the financial risk of shipping.

Revenue never decides a test. pagent stops tests on the conversion rate of the primary goal. If revenue matters most to you, compare the revenue figures before you ship, or set [Automatic stopping](/docs/statistics/settings#automatic-stopping) off and decide yourself.

## Tests with several variations

pagent compares each variation with control on its own.

The more variations a test has, the higher the chance that one of them crosses the bar by luck. pagent corrects for this automatically: it splits your error budget evenly across the variations (Bonferroni correction). This happens for both engines, whether or not sequential correction is on.

| Your setting | 1 variation | 2 variations | 3 variations |
| - | - | - | - |
| Chance to beat control threshold 98.5 % (Balanced) | 98.5 % | 99.25 % | 99.5 % |
| Chance to beat control threshold 99.5 % (Careful) | 99.5 % | 99.75 % | 99.83 % |
| Significance level 10 % (Balanced) | 10 % | 5 % | 3.33 % |

So a level's promise holds for the whole test, however many variations it has.

The adjusted value is the one shown on the results page, for example "threshold: 99.25 %". Intervals, practical equivalence, the frequentist equivalence test and the [stop losing threshold](/docs/statistics/settings#stop-losing-variations-at) use your setting as it is. A losing variation only stops the test when even the best variation is clearly losing.

Some more rules for tests with several variations:

* The [minimum total conversions](/docs/statistics/settings#minimum-total-conversions) apply to each comparison: control plus that one variation.
* The test follows its **best variation**. When that variation's comparison ends, the whole test stops for every variation. pagent does not drop single variations during a test.
* More variations need more visitors. Each one gets a smaller share of traffic and has to clear a stricter bar.

## Traffic split check

pagent checks that visitors are split between control and the variations as planned. A clear gap is called a **sample ratio mismatch** (SRM). It usually means something technical is off: bots are filtered differently, a redirect loses visitors, a cache serves one version more often, or the variation breaks on some browsers.

* The **SRM Check** column in the variations table shows **Issue Detected** when the split is off by more than chance explains (a chi-square test at p below 0.01), and **OK** otherwise.
* Smaller gaps (p below 0.05) are reported as a minor deviation worth watching.
* The check is always on and cannot be configured.

The SRM check does **not** stop a test or block a result. If it shows **Issue Detected**, find the cause before you trust the result. In chat, you can ask pagent to explain the traffic imbalance.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.