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

# A/B Testing Analytics

> Compare feature-flag variants against a goal, then measure conversion, lift over control, sample size, and statistical significance.

An experiment tells you which variant of a feature actually converts better — not which one you *think* wins. You point an experiment at a [multivariate feature flag](/docs/feature-flag-analytics) and a [goal](/docs/conversion-goal-tracking), and tinyanalytics counts conversions for each variant, computes the lift over your control, and tells you when the difference is statistically significant.

## Before you start

* A **multivariate feature flag** with the variants you want to test (for example `control` and `treatment`).
* A **goal** that defines the conversion you care about — a signup, a purchase, a key page view.
* Your app calls `window.tinyanalytics.flag()` where the visitor experiences the variant. This is what makes the measurement accurate.

## How does an experiment count conversions?

The unit of measurement is the **session**, not the pageview. For each variant, tinyanalytics counts the sessions that saw it and the sessions that then completed your goal. Attribution is **post-exposure**: a conversion only counts if it happened *at or after* the visitor first saw the variant.

* A session that converts *before* ever seeing the variant is **not** credited to it — a variant can't take credit for behavior that preceded it.
* A session that converts twice counts **once**.
* Conversions can never exceed what your goal reports for the same period, because the experiment reuses the goal's own definition.

## Where should I call flag()?

Call `window.tinyanalytics.flag()` at the moment the visitor experiences the variant. This is the one step that decides whether your results are trustworthy. The call records an **exposure** — proof that this session actually experienced the variant — and the experiment counts those exposed sessions as its primary, statistically correct measurement.

```js theme={null}
// Fire the exposure at the moment the visitor experiences the variant
const variant = window.tinyanalytics.flag('checkout-test', 'control')
render(variant === 'treatment' ? <NewCheckout /> : <OldCheckout />)
```

<Warning>
  If no session ever calls `flag()` for the experiment's flag, tinyanalytics falls back to an **assignment** measurement — everyone the flag *would* have bucketed — and flags the results as less reliable, because assignment doesn't prove anyone saw the difference. Call `flag()` at the point of experience to keep your experiment on the accurate measurement.
</Warning>

## How do I read the results?

Each variant shows its **sessions**, **conversions**, **conversion rate**, and **lift** versus the control:

* **Control** is the variant named `control`, or the first one you declared. Every variant you declared appears in the table even before it has any traffic.
* **Lift** is the variant's conversion rate relative to control. It reads as blank (not "0%") when the control itself hasn't converted yet — there's no baseline to compare against.
* **Significance** is a two-proportion z-test computed against control. tinyanalytics needs **at least 30 converting-or-exposed sessions in each arm** before it will call anything — until then it shows **"Gathering data."** A result is marked **significant at 95% confidence** or higher.

tinyanalytics never auto-picks a winner and applies no multiple-comparison correction. It surfaces which variant is leading once that lead is significant, and the call to ship stays yours.

## Frequently asked questions

<AccordionGroup>
  <Accordion title="Why does my experiment say 'Gathering data'?">
    Your arms don't have enough traffic yet. tinyanalytics needs **at least 30 converting-or-exposed sessions in each arm** before it calls significance, and shows "Gathering data" until then. Significance is a two-proportion z-test against control, marked significant at 95% confidence or higher.
  </Accordion>

  <Accordion title="Why is the lift column blank for my variant?">
    The control hasn't converted yet, so there's no baseline to compare against. tinyanalytics shows lift as blank rather than "0%" in that case. Lift is the variant's conversion rate relative to control.
  </Accordion>

  <Accordion title="Do conversions that happen before a visitor sees the variant count?">
    No. Attribution in tinyanalytics experiments is **post-exposure** — a conversion only counts if it happened at or after the visitor first saw the variant. A session that converts twice still counts once.
  </Accordion>

  <Accordion title="Does tinyanalytics pick the winning variant for me?">
    No. tinyanalytics never auto-picks a winner and applies no multiple-comparison correction. It surfaces which variant is leading once that lead is statistically significant, and the decision to ship stays yours.
  </Accordion>
</AccordionGroup>

## Related

<Columns cols={2}>
  <Card title="Feature flags" icon="toggle-on" href="/docs/feature-flag-analytics">
    Create the multivariate flag your experiment tests.
  </Card>

  <Card title="Goals" icon="bullseye" href="/docs/conversion-goal-tracking">
    Define the conversion the experiment measures.
  </Card>

  <Card title="Custom events" icon="hand-pointer" href="/docs/custom-event-tracking">
    Send the events your goal is built from.
  </Card>

  <Card title="Funnels" icon="filter-circle-dollar" href="/docs/funnel-analytics">
    See where visitors drop off on the way to converting.
  </Card>
</Columns>


## Related topics

- [Feature Flag Analytics](/docs/feature-flag-analytics.md)
- [Configure the tinyanalytics Tracking Script](/docs/configure-tinyanalytics-tracking-script.md)
- [Privacy-Friendly Web and Product Analytics](/docs/index.md)