# Measurement

`verify_change` proves a change is on the page. `get_outcomes` reports what the page's numbers did before and after it. This page says exactly what is read, over which days, how a change is matched to the numbers, and what the comparison cannot tell you.

## Sources

You connect a source for a site with `connect_analytics`, which returns a one-time link for you to open.

**Google Analytics 4.** One `runReport` call per period, with the metrics `sessions`, `engagementRate` and `conversions` (conversions only where your property has them configured), filtered on `pagePath` for the page's path with and without a trailing slash. The sample size of a period is its sessions.

**Search Console.** One `searchanalytics.query` call per period with `dataState: final`, returning `clicks`, `impressions` and `ctr`, filtered to the page with an `includingRegex` page filter. The sample size of a period is its clicks.

**Microsoft Clarity** can be connected, and its credentials are stored encrypted, but its data is not used for outcomes, because it keeps only a few days of page history, which is too short for the windows below. GA4 sessions and Search Console clicks are the outcome sources.

## Windows

For a change verified on day D (UTC):

- the baseline period is D−28 to D−1;
- the outcome period is D+1 to D+28;
- day D itself belongs to neither, because the change went live part-way through it.

Outcomes for a change are computed only once the outcome period has ended and a further 3 days have passed, because GA4 and Search Console data for a day is not final straight away.

## Sample floor

A relative change is computed only when both periods reach 500: sessions for GA4, clicks for Search Console. Below that, both values are reported side by side with their sample sizes, labelled `insufficient_sample`, and no percentage is derived. A missing value, or a baseline of zero, is labelled `not_comparable`.

## Attribution

A change is matched to the numbers by page URL. Day D is the change's `verified_at`: the moment `verify_change` ran and found the tactic present, not the moment you deployed. Run `verify_change` promptly after each release; a late verify puts days that already had the change into the baseline. Only changes verified as present are measured.

## What it cannot tell you

- Several changes to one page inside one window cannot be told apart. They share the same before and after numbers.
- There is no control group. Traffic that changed for another reason, such as a campaign, a launch or a search update, shows up as a difference.
- There is no seasonality adjustment. A holiday or a quarter-end inside either window shows up as a difference.
- It is a before and after comparison on one page, not an experiment. It describes what happened; it does not prove the change caused it.

This is why a tactic reaches the measured [evidence grade](/docs/evidence-grades) only with outcomes from several sites, and why a measured figure is never a promise for your site.

## When outcomes run

Outcomes are computed by a nightly job, which is switched on in production as an operations step. Until it is, a verified change stays "shipped, unmeasured".