Validate an SEO change: did it really work?
A rise after a change is not proof that the change worked. This playbook gives a five-step test: mark the date, build a fair comparison, measure your own normal variation, let the tools find the date and the page group, and write a verdict that admits when the result is not proven.
✓ Checked against Google documentation and the workspace · 7 October 2026
Why a rise after a change is not proof
You rewrote titles on Monday and clicks went up by Friday. That may be your change, or the week’s weekday mix, a seasonal peak, a Google update or plain noise. Search Console shows what happened, not why. To say a change worked you need three things: a fixed start date, a fair comparison window and a margin of normal variation that the result must clear.
How long to wait
Google says some changes take effect in a few hours and others take several months, and suggests waiting a few weeks before judging (SEO Starter Guide). Its traffic-drop guide repeats that a change can take from days to months to show (debugging traffic drops). So set the check date before you start: for a title rewrite, four weeks after the pages were recrawled is a sensible first read; for a migration, longer.
Step 1: mark the date
Add an annotation on the day the change went live, with a label that says what changed and on which pages. Without it, in three months nobody knows which line movement belongs to which release. Note also the date you expect Google to have seen the change; the day you published is not the day it was recrawled.
Step 2: build a fair comparison
- Use equal periods. Compare the same number of days before and after, never “since the change” against “everything before”.
- Match weekdays. Turn on weekday matching so Monday is compared with Monday (see dates and comparisons); weekday mix alone can move a total by several percent.
- Leave out the fresh days. The newest data can be preliminary and may change within hours (Performance report help).
- Check the season. Compare year over year as well, and look at the queries in Google Trends, as Google advises when diagnosing a drop.
Step 3: measure normal variation first
Before the change, take eight weekly click totals and compute their average and standard deviation. A result counts as a signal only if it clears about two standard deviations and stays there.
| Step | Example (illustrative numbers) |
|---|---|
| Eight weeks before | 1,000 · 1,040 · 980 · 1,020 · 1,010 · 990 · 1,030 · 1,005 |
| Average | 1,009 clicks a week |
| Standard deviation | 20 clicks, about 2.0 percent |
| Threshold: average + 2 deviations | 1,050 clicks |
| Four weeks after | 1,090 · 1,105 · 1,080 · 1,115, average 1,098 |
| Change | +8.7 percent; every week is above 1,050 |
Here the rise clears the noise in every post-change week, so it is worth attributing to the change, subject to the checks in step 2. Had the weeks been 1,060, 1,010, 1,075 and 1,030, the average would be above the baseline but two weeks would sit inside the noise, and the honest conclusion would be “not proven”.
Step 4: let the tools find the date and the group
- Change points. The change-point report finds the day where the level moved. If it points to a date days before your release, the cause is something else.
- Page-group experiment. Put the changed pages in one group and similar untouched pages in another, then compare their click growth against an equal earlier period (forecasts and experiments). The page itself calls this an observational comparison, not a randomized test, so it strengthens the evidence without proving the cause.
- Google updates. The annotation catalogue shows Google ranking updates; a rise that starts on an update date is not yours to claim.
Step 5: write the verdict
| Evidence | Verdict |
|---|---|
| Clears two deviations every week, control group flat | likely the change |
| Clears the threshold, but the control group rose too | market or season, not the change |
| Average above baseline, some weeks inside the noise | not proven; wait or extend the window |
| Moves on an update date | confounded; check Google’s status dashboard |
Record the verdict in the annotation. A rejected change is information too.
Validating an SEO change, answered
How long should I wait before judging?
Google suggests a few weeks, and some changes take months. Fix the check date in advance.
Is a rise of five percent real?
Only if it clears your own normal variation; with a standard deviation of two percent it can be, with six percent it is noise.
Can a Search Console experiment prove the cause?
No. A group comparison is observational; it makes the case stronger, not certain.
What to read next
Use content decay to find pages worth changing and the filters and exports page to freeze the page list before you start.