> For the complete documentation index, see [llms.txt](https://docs.humblytics.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.humblytics.com/split-testing-overview/how-humblytics-split-testing-works-under-the-hood.md).

# How Humblytics Split Testing Works - Under the Hood

How visitors are assigned to a version, why they don't see a flicker, and how Humblytics decides a winner.

## Showing each visitor a version

Everything runs through the tracking script you already installed. When someone lands on a page with a live test:

1. **The page is briefly hidden.** The first part of your tracking snippet hides the page for a moment, 2 seconds at most, so the visitor never sees the original flash before their version appears.
2. **The visitor is assigned a version** based on your traffic split. With **Per visit**, each visit is assigned again. With **Sticky visitor**, the same person keeps seeing the same version.
3. **The version is shown.**
   * **No-code tests:** the script applies your editor changes (text, styles, images, and so on) to the page.
   * **Page-level tests:** the script sends the visitor to the variant URL. The redirect on the original page isn't counted as a visit, so nobody is counted twice.
4. **The page appears**, and tracking carries on as normal.

Single-page apps work too: the script also runs tests when the URL changes without a full page load.

## Measuring each version

For every version, Humblytics counts **visitors** and **conversions** for your goal, then calculates:

| Number              | What it means                                                                            | Example                          |
| ------------------- | ---------------------------------------------------------------------------------------- | -------------------------------- |
| **Conversion rate** | Conversions ÷ visitors                                                                   | 60 signups ÷ 1,000 visitors = 6% |
| **Lift**            | How much better or worse the variant is than the control, as a percentage of the control | 6% vs. 5% = **+20%** lift        |
| **Confidence**      | How sure you can be that the difference is real, not luck                                | 95%                              |

The name of the "conversion" depends on your goal: **Submissions** for forms, **Clicks** for click-through rate, **Orders** for revenue, and so on. See each goal page for details.

## How confidence is calculated

Humblytics compares the control and each variant with a **two-proportion z-test**, a standard statistics test for comparing two rates.

1. Conversion rate of each version: p₁ = conversions\_A ÷ visitors\_A, p₂ = conversions\_B ÷ visitors\_B.
2. Combined rate across both: p = (conversions\_A + conversions\_B) ÷ (visitors\_A + visitors\_B).
3. Standard error: SE = √\[p × (1 − p) × (1 ÷ visitors\_A + 1 ÷ visitors\_B)].
4. Z-score: Z = (p₂ − p₁) ÷ SE.
5. The z-score converts to a **p-value** (the chance of seeing a difference this big by luck alone) and a **confidence** level. A z-score of about ±1.96 equals 95% confidence.

The results page also shows **confidence intervals**: the likely range of each version's true rate, and of the lift. If the ranges don't overlap, the difference is probably real.

## When a test is ready to call

A test is **Ready to Call** only when all three are true:

1. Confidence has reached your target (95% by default).
2. At least **80%** of the needed sample has been collected.
3. The test has run for the recommended minimum number of days for its goal.

Until then, you'll see **Keep Running**. Checking early and stopping as soon as something looks good is the most common way to pick a false winner. [How to read the verdicts](/split-testing-overview/how-to-analyze-split-test-data.md#the-verdict).