Why do pages that dropped often recover on their own, and how much does that fool a measurement?

Regression to the mean is one reason pages that dropped often recover on their own: part of a low month is bad luck, and next month brings new luck. The effect fools a measurement when pages are picked for a bad month: in Penlid's simulations, unchanged pages picked that way were called "improved" 5.50% to 9.20% of the time.

• After an unusually low or high reading, the next reading of the same quantity tends to land closer to the average (Wikipedia, Regression toward the mean, read 30 October 2026).
• For a small drop in position, Google advises against radical changes to a page that is already performing well (Google, Search Central, read 30 October 2026).
• In simulations, unchanged pages picked because their last 28 days sat at least one standard deviation below their own average were called "improved" 5.50% to 9.20% of the time, against at most 3.10% for pages not picked that way (Penlid, Methodology, read 30 October 2026).
• In the same simulations, the pre-trend warning appeared on only 10.9% to 22.7% of the pages picked on decline (Penlid, Methodology, read 30 October 2026).

What is regression to the mean?

Regression to the mean is the tendency for an unusually low or high reading to be followed by one nearer the average, because part of every reading is chance. Wikipedia notes that the effect is statistical, not causal.

Think of a page's clicks in a 28-day window as three parts: its normal level, counting noise and drift. A page that normally earns 144 clicks does not earn exactly 144 every month; that is counting noise. Drift is what moves pages on their own: a season hits one more than another, a ranking wobbles, a link lands on one and not its neighbour (our methodology page). Pick the pages with the weakest recent window and you pick pages where noise and drift happened to push down. Noise is drawn fresh each window, and drift that was short-lived is gone, so on average those pages move back toward normal without any help from you.

Why do pages that dropped often recover on their own?

Pages that dropped often recover on their own because part of the low window was noise or short-lived drift, and neither carries into the next window. A dip made of counting noise reverses on average, because the next window draws a fresh count. A dip with a temporary cause, such as a news event or a competitor's page that moved up for a while, reverses when the cause passes.

Google makes a related point about small drops. Its page on debugging traffic drops treats them as normal movement: a position can slip a little and come back without you needing to do anything, and the page advises against radical changes to a page that is already performing well. It gives no recovery rate. Google treats a notable drop across a wide range of terms differently: it tells you to assess the whole website, not just individual pages. Not every drop reverses on its own, so the question is how much of a recovery after your change was coming anyway.

How much does regression to the mean fool an SEO measurement?

In Penlid's simulations, unchanged pages picked for a weak four weeks were falsely called "improved" 5.50% to 9.20% of the time, against at most 3.10% for pages not picked that way. We tested the estimator behind our free did-it-work tool on simulated sites where the true answer was known. Each page we treated as changed was picked because its last 28 days sat at least one standard deviation below its own average, and then nothing was done to it. The figures come from our methodology page; the SEO test calculator repeats them and says each is uncertain by about 1 point.

Picking on decline put the false-win rate at 1.8 to 3.0 times the 3.10% ceiling (5.50 ÷ 3.10 = 1.77 and 9.20 ÷ 3.10 = 2.97). In plain counts, 5.50% is about 1 unchanged page in 18 (1 ÷ 0.055 = 18.2) and 9.20% is about 1 in 11 (1 ÷ 0.092 = 10.9). More traffic did not make the problem go away: a 40-click-a-day page was falsely called "improved" 5.75% and 9.20% of the time, against 5.50% and 9.15% at 4 clicks a day. With counting noise alone, the rebound shrinks in percent as a page grows, but so does the noise it is judged against. We call refreshing pages because they were declining the largest remaining source of false "improved" results in our method, and we cannot remove it.

Why can't a comparison group remove regression to the mean?

A comparison group cannot remove regression to the mean because the comparison pages were not picked for a bad stretch, so they have no rebound to subtract. Our methodology says a page that has just had a bad four weeks is likely to have a better next four whatever you do to it, and that no comparison group can subtract it, because the comparison pages were not chosen for having had a bad four weeks.

A comparison group does absorb what moves every page alike, such as the shared part of a season, a holiday or a Google update, which is why we compare against pages left alone (see how to choose control pages and how to separate seasonality from a real SEO change). Regression to the mean acts on the pages chosen for a bad stretch, and the comparison pages were not chosen that way. The same limit applies to difference-in-differences: Wikipedia's article on the method says it can still suffer from mean regression, depending on how the treatment group is chosen.

Our pre-trend check helps a little. It repeats the calculation one stretch earlier and warns when the page was already moving differently from its comparison pages. On pages picked on decline it warned on only 10.9% to 22.7%, which leaves 77.3% to 89.1% with no warning. We call it a useful flag and not a safeguard. The did-it-work tool runs it on a day-by-day export but not on a Compare export, which is the Search Console download taken with the Compare option.

How do you stop regression to the mean from fooling a measurement?

Choose the pages to change without looking at how they were trending, and regression to the mean has nothing to act on. Our advice is to choose which pages to change without looking at how they were trending: change every page in a set, or alternate pages, and the comparison is clean. To alternate, order the candidates by something the results cannot influence, such as the URL, and change the first, third and fifth. Both halves come from the same list, so both should rebound alike, and the rebound should drop out of the comparison. Choosing pages to change without bias covers this in more detail.

If you already picked on decline, say so in the report. Our methodology tells you to read an "improved" as at least as good as the pages left alone, and possibly some recovery that was coming anyway. A decay report, meaning a list of pages losing clicks, selects on decline too, so the same reading applies to a refresh of whatever it lists. How to tell if a content update worked covers the wider method.