The annual report lands on the desk with good news on the front page: the site’s average position went from 23 to below 13. Ten positions gained in a year, without a link campaign, without rewriting any content.
Then we turn the page, and clicks are the same as before: 168 a day before, 168 a day after. A site that climbs ten positions and does not get a single extra click is a strange site. Or it is a site that did not climb at all.
We have already seen, when talking about anomaly detection, how to recognise an abnormal day: a spike, a crash, a value the series did not expect. Here the problem is different, and more treacherous. There is no strange day: there is a day after which all days are different, because the way of counting them has changed. It is called a structural break, and in web data it happens much more often than we think.
In short: an anomaly is an isolated abnormal point; a structural break is a change in level (or behaviour) that stays. In SEO metrics, one cause to check every time, before interpreting the jump, is a change in the measuring instrument: in September 2025, for example, Google’s &num=100 parameter stopped working, and on many sites a good share of impressions in deep positions vanished overnight. Average position, which is a weighted average over impressions, “improved” all at once without any page moving. Finding a break requires different tools from those for anomalies: the cumulative sum of deviations (CUSUM), a formal test and the breakpoints() function of the R package strucchange, which estimates how many breaks there are and where. Once found, the break must be respected in comparisons: periods on the same side, metrics it touches less, and a line in the report that declares it.
What we will cover:
- Anomaly or break?
- The &num=100 case
- Average position is a weighted average
- Finding the break: first by eye, then with CUSUM
- Is it a break or is it chance?
- Where and how many: breakpoints()
- Handling it in comparisons
- The other yardsticks that change
- Frequently asked questions
- Try it yourself
- Further reading
Anomaly or break?
The difference is easier to grasp with two images. An anomaly is a stone dropped into water: a splash, and the level goes back to what it was. A structural break is a sluice gate opening: the level drops, and stays there.
The two raise different questions. Faced with an anomaly we ask what happened that day: a bot, a mention in a newspaper, a server problem. Faced with a break we ask what changed from that day on, and the answer, very often, is not about people’s behaviour but about the yardstick we measure it with: an instrument that counts differently, a new filter, a definition that changes.
The tools, consequently, are different. For an anomaly we compare each day with what we expected and look for deviations that are too large. For a break we need to understand whether the series is better described by a single level or by two (or more) different levels in different periods, and where the boundary falls. An anomaly is searched for day by day; a break is searched for by looking at the series as a whole.
The &num=100 case
Until the summer of 2025, adding &num=100 to the address of a Google search returned a hundred results on a single page instead of ten. Few humans used it; rank-tracking tools used it a great deal, reading the whole ranking down to position one hundred in a single request.
Around 10-12 September 2025 the parameter stopped working, and a few days later Google confirmed it is not a supported parameter. At the same moment a great many sites saw impressions fall and average position suddenly improve, especially on desktop.
The explanation most widely accepted among practitioners (an interpretation, not a statement from Google) is that part of the impressions in deep positions came precisely from those hundred-result pages loaded by the tools: without the parameter, the tools read rankings ten results at a time, and those impressions are no longer generated.
Let us take an example, with data built on purpose to reproduce the shape of the phenomenon. One year of Search Console, from 1 March 2025 to 28 February 2026, with two components:
- impressions in the top ten positions, coming from real people: about 3,000 a day (a quarter fewer at weekends), at a position around 4.5;
- impressions beyond position ten, largely generated by tools: about 2,300 a day, at a position around 45, which drop to about 700 from 11 September.
Clicks come only from the first page, and do not change. We also add an anomalous day, 20 May, on which a bot generates nine thousand extra deep impressions.
I build the data in R:
library(strucchange)
set.seed(2026)
day <- seq.Date(as.Date("2025-03-01"), as.Date("2026-02-28"), by = "day")
n <- length(day)
cut <- as.Date("2025-09-11")
after <- day >= cut
weekday <- ifelse(format(day, "%u") %in% c("6", "7"), 0.75, 1)
# impressions in the top ten positions: real people, unchanged rhythm
impr_top <- rpois(n, 3000 * weekday)
pos_top <- rnorm(n, 4.5, 0.25)
# impressions beyond position ten: largely rank-tracking tools;
# after the cut few of them remain
impr_deep <- rpois(n, ifelse(after, 700, 2300) * exp(rnorm(n, 0, 0.15)))
pos_deep <- rnorm(n, 45, 3)
# an anomalous day: a bot adding nine thousand deep impressions on 20 May
spike <- day == as.Date("2025-05-20")
impr_deep[spike] <- impr_deep[spike] + 9000
# clicks come only from the first page and do not change
clicks <- rpois(n, impr_top * 0.06)
gsc <- data.frame(
day, after,
impressions = impr_top + impr_deep,
clicks,
position = (impr_top * pos_top + impr_deep * pos_deep) / (impr_top + impr_deep),
impr_top, pos_top, impr_deep
)Search Console, of course, does not give us the two components separately: it gives us total impressions, clicks and the average position of each day. I compute the means of the daily values before and after 11 September:
avg <- function(x) round(tapply(x, gsc$after, mean), 2)
avg(gsc$impressions)
avg(gsc$clicks)
avg(gsc$position)| before 11 September | from 11 September | change | |
|---|---|---|---|
| impressions per day | 5,164 | 3,500 | −32% |
| clicks per day | 168.3 | 168.1 | 0% |
| average position (mean of daily values) | 22.85 | 12.82 | ten positions “gained” |
| CTR | 3.26% | 4.80% | +47% |
Three metrics out of four tell a spectacular story (fewer impressions, but position and CTR through the roof) and one, clicks, says nothing happened. It is the metric that stays silent that tells the truth: real people see the site as much as before, in the same positions as before, and click as much as before.
Average position is a weighted average
To understand how position can “improve” without any page moving, it helps to look at how it is built. Search Console does not average the positions of our pages, but averages over impressions: every time the site appears in a search, the position of its topmost result in that search enters the calculation. In words: it is a weighted average, in which each position weighs as much as the impressions it collected.
\( \displaystyle \text{average position} = \frac{\sum_i \text{impressions}_i \cdot \text{position}_i}{\sum_i \text{impressions}_i} \)
where the sum runs over non-overlapping groups (queries, or days) and each position is multiplied by the impressions it collected. If the weights change, the average changes, even if no position changes. (Computed this way, over the two whole periods, the position in our simulation goes from 23.00 to 12.75: almost the same numbers as the table, which reports the simple mean of the daily values.)
That is exactly what happened. I compute in R the share of impressions beyond position ten and the average position of the top ten only, before and after:
tapply(gsc$impr_deep, gsc$after, sum) / tapply(gsc$impressions, gsc$after, sum)
# before 0.460 after 0.204
tapply(gsc$impr_top * gsc$pos_top, gsc$after, sum) / tapply(gsc$impr_top, gsc$after, sum)
# before 4.50 after 4.51Before the cut almost half of the impressions (46%) sat beyond position ten, around 45; afterwards, only a fifth. The average position of the top ten stayed identical, 4.50 against 4.51. The average improved because the largest numbers disappeared from the calculation, not because any number got smaller.
Readers of the article on Simpson’s paradox will recognise the mechanism: an aggregate that moves while none of the groups composing it moves, only because their weight changed. The same holds for CTR, which is a ratio of clicks to impressions: if impressions that produced no clicks vanish from the denominator, CTR rises by itself.
With real data the two components cannot be separated exactly, but an approximate check can be done. We download queries one by one (with the Search Console API or the per-query export), choose a fixed set of queries present in both periods, and recompute the average position only on those with position up to 10. The filter is not perfect: a query’s position is already an average, and a query at position 9 may have some impressions further down; moreover, the per-query export leaves out anonymised queries. But if on that set the position does not move while the overall one jumps, we have a strong hint that the jump lies in the composition.
Finding the break: first by eye, then with CUSUM
The first tool is always the chart. Let us stack the impressions series on top of the average position series:

By eye the picture is clear: an isolated spike in May, which the series forgets the next day, and a step change in September that persists. The chart, however, has a limit: it works when the break is large. A 5% break, mixed with the noise of the days of the week, escapes the eye.
To see it better there is a simple trick, the cumulative sum of deviations from the mean, known as CUSUM. The idea is this: we compute the mean of the series over the whole year, and then add up, day after day, how far each day sits above or below that mean.
\( \displaystyle S_t = \sum_{i=1}^{t} \left( y_i – \bar{y} \right) \)
where \( y_i \) is the value of day \( i \) (here the logarithm of impressions, to reason in terms of percentage changes), \( \bar{y} \) is the mean of the whole period and \( S_t \) is the sum accumulated up to day \( t \).
If the series fluctuates around a single level, positive and negative deviations offset each other and the sum wanders around zero. If instead the series has two levels, in the first period the days sit almost all above the overall mean and the sum rises; in the second they sit almost all below and the sum falls (at the end it always returns to zero, by construction, because deviations from the mean add up to zero). The point where the cumulative sum stops rising and starts falling is the natural candidate for the break. It is a clue, not a proof: to verify that the break exists, and to locate it with its uncertainty, we need the tools we will see shortly.
I compute it in R:
y <- log(gsc$impressions)
cusum <- cumsum(y - mean(y))
day[which.max(abs(cusum))]
# "2025-09-10"In this example the maximum falls on 10 September, the last day before the cut, and correctly marks the boundary. The May spike, on the other hand, leaves barely a small step in the cumulative sum: a single day weighs little on a sum of three hundred.

The same figure shows the cumulative sum of clicks, which stays flat around zero all year. The two curves, side by side, suggest at a glance what the table said with numbers: impressions changed level, clicks did not. (A flat curve does not prove the absence of breaks: two opposite breaks, for example, could cancel out. That is why the next step is a test.)
Is it a break or is it chance?
Even a series with no break at all produces a cumulative sum that drifts a little away from zero: chance, by itself, wanders. The question is how far the curve must go before chance is no longer enough to explain it. It is the same logic as hypothesis testing: we start from the hypothesis that the series has a single level, and measure how surprising, under that hypothesis, a cumulative sum like the observed one would be.
The strucchange package does the calculation for us, with the OLS-CUSUM test: the cumulative sum of the residuals of a model (here the simplest one, a constant level), rescaled and compared with the limits chance can reach. I run the test in R, on impressions and then on clicks:
sctest(efp(y ~ 1, type = "OLS-CUSUM"))
# S0 = 8.2333, p-value < 2.2e-16
yc <- log(gsc$clicks)
sctest(efp(yc ~ 1, type = "OLS-CUSUM"))
# S0 = 0.39067, p-value = 0.998The impressions statistic is 8.2, far beyond the 5% critical value, which is around 1.36: if the series had a single level (and the assumptions of the test held), a deviation this large would be extremely unlikely. The printed p-value, “< 2.2e-16”, does not mean probability zero: it is just the smallest number R bothers to write. The clicks statistic is 0.39, with a p-value of 0.998: no evidence of a break.
Where and how many: breakpoints()
The test says whether there is a break, not where, nor how many. That is what breakpoints() is for, and what it does is conceptually simple: it tries every possible way of cutting the series into segments (one, two, three…), estimates each segment’s level, and looks for the cuts that make the overall error smallest. Then, to decide how many cuts, it uses the BIC, a criterion that rewards fit to the data but penalises every extra segment: one more cut earns its place only if it improves the model enough to pay its penalty.
I estimate in R the breakpoints of impressions, requiring each segment to last at least 10% of the series (h = 0.1, that is 36 days):
bp <- breakpoints(y ~ 1, h = 0.1)
summary(bp)$RSS["BIC", ]
day[bp$breakpoints] # "2025-09-10"
day[confint(bp)$confint] # "2025-09-09" "2025-09-10" "2025-09-12"
round(exp(coef(bp))) # 5122 and 3481| breaks | 0 | 1 | 2 | 3 | 4 |
|---|---|---|---|---|---|
| BIC | −48.3 | −539.5 | −530.8 | −522.0 | −512.1 |
The lowest BIC is the one with a single break, by far: going from zero to one break drops the BIC by almost 500 points, going from one to two makes it rise again. The break falls on 10 September 2025, which in strucchange is the last day of the first segment: the new level starts on the 11th, exactly where we put it. The 95% confidence interval runs from 9 to 12 September. The two estimated levels, brought back from the logarithmic scale, are 5,122 and 3,481 impressions a day: they are geometric means, a little lower than the arithmetic means of the first table (5,164 and 3,500), as always happens when coming back from logarithms.
Two things deserve attention. The first: the May spike was not mistaken for a break. A single day is not enough to make a segment (at least 36 would be needed), and in this simulation it does not shift the cuts. This is not a general guarantee: a stronger anomaly, or one lasting a few weeks, can move the boundary or favour a segment containing it, and with real data it is worth repeating the estimate without the anomalous days to see whether the result changes.
The second: the same procedure on clicks chooses zero breaks (the lowest BIC, −305.6, is the one without cuts).
N.B. the model we used is the simplest possible, a constant level per segment, and it ignores that weekends have fewer impressions. With such a clean jump the date of the break does not change, but the uncertainty does: putting the day of the week into the model as well (breakpoints(y ~ weekday_f, h = 0.1)) narrows the interval to 9-11 September. A regular rhythm left in the residuals makes tests and intervals less reliable, and with smaller breaks it can shift the date too: it is better to model it, or to work on weekly totals.
Handling it in comparisons
Once the break is found, the real work is not to be fooled by it in the comparisons we make every day.
A warning: before interpreting a sudden change as a success or a problem of the site, check whether the measuring instrument changed on the same day. A jump that appears on the same day on many different sites, or on one metric but not on the one next to it, is very often a change of yardstick. A number that changes all at once is first of all a question about the instrument, and only then a question about the site.
In day-to-day practice, there are three rules.
The first: compare only periods on the same side of the break. A comparison of impressions between October 2025 and October 2024 risks measuring the &num=100 cut more than the site: how much it weighs depends on the property (and on the device), and must be checked case by case. The same goes for the year-over-year comparison, which for a whole year after the break compares two different yardsticks.
The second: use the metrics the break touches least. In the &num=100 case clicks are much less exposed, because rank-tracking tools do not click; impressions, average position and CTR are much more so. If a question can be rephrased in terms of clicks, it is worth doing, while using them as a check and not as a guarantee: in the same period they may have changed too, for reasons of their own.
The third: when the break cannot be avoided, put it in the model. If we must estimate something over a period that crosses it, we add a variable that is 0 before and 1 after, so the model absorbs the jump instead of attributing it to something else. I estimate it in R on impressions, together with the day of the week:
gsc$weekday_f <- factor(format(day, "%u"))
m <- lm(log(impressions) ~ weekday_f + after, data = gsc)
exp(coef(m)["afterTRUE"]) # 0.679
exp(confint(m)["afterTRUE", ]) # from 0.668 to 0.691The estimated jump is 0.679: from 11 September, 32% fewer impressions. Careful about what this number says: it is the difference between before and after, net of the day of the week. We know the change comes from the yardstick rather than the site because we built the simulation ourselves; in real data the attribution comes from external information (the date coincides with a known change, the jump appears on many sites), and an intervention on the site on the same day would be indistinguishable with this variable alone. Any other effect we want to measure on that series (a migration, an update, a campaign) must be estimated net of this one, not together with it. It is the same reasoning as the control group in the article on expected CTR before and after an event: separating what happened to the site from what happened around it.
And then there is a rule that has little to do with statistics: declare it. One line in the report (“from mid-September 2025 impressions and average position are affected by a change in Google’s measurement: comparisons with earlier periods should be read with caution”) saves readers from celebrating ten positions that do not exist, or worrying about a third of impressions nobody ever saw.
The other yardsticks that change
&num=100 is the most striking case of recent years, but not the only one. A few examples that come up often in day-to-day work:
- The cookie banner. A stricter banner, or a different consent configuration, can halve the users GA4 manages to count, overnight. The site has the same visitors; we see half of them.
- A duplicated tag. The opposite mistake: the measurement code loaded twice doubles page views and pushes the bounce rate down to impossible values.
- Filters and settings. Excluding internal traffic, a bot filter, a domain added to measurement: every change to the GA4 configuration is a potential break.
- New conversion definitions. In Google Ads, a conversion action set as primary (for example local actions, such as requests for directions) makes conversions jump without a single extra contact.
- Migrations and URL changes. When pages change address, per-page series break apart: the old page “crashes”, the new one “is born”, and neither tells the right story.
In all these cases the remedy starts from the same gesture: keeping a change log (of the site, of measurement, of the platforms), with dates. It is the most boring document there is, and the one that saves the most analyses.
Frequently asked questions
What is a structural break in a time series?
It is a moment when the series changes level, trend or variability in a lasting way. In web data one possible cause, to check before interpreting the jump, is a change in the measuring instrument rather than in the phenomenon measured.
Why did Search Console average position improve in September 2025?
Around 10-12 September 2025 the &num=100 parameter stopped working, and on many sites impressions in deep positions fell; the most widely accepted explanation is that part of them came from rank-tracking tools. Average position is a weighted average over impressions: without those distant impressions, the average drops even if no page has moved.
How do we find a structural break in R?
With the strucchange package: efp() and sctest() to check whether there is a break, breakpoints() to estimate how many there are and where they fall, confint() for the interval on the dates.
What is the difference from an anomaly?
An anomaly is an isolated point, after which the series returns to its level; a break is a change that stays. The tools for finding them are different; a very strong or long anomaly, however, can disturb the search for breaks, and it is good practice to check that the result does not depend on it.
Try it yourself
The simulation is designed to be modified. Three directions:
- Let us make the break small. We make deep impressions drop from 2,300 to 2,100 instead of 700, and repeat chart, CUSUM and
breakpoints(). Can the eye still see it? And the algorithm? What happens to the confidence interval on the date? - Let us add a second break. We simulate a cookie banner that, from January 2026, removes 30% of clicks. How many breaks does
breakpoints()find on impressions, and how many on clicks? - Let us replace the spike with a week of bots. Instead of one anomalous day, we put seven consecutive days of inflated impressions: at what point does a long anomaly start to look like a break?
A hint: the second test is the most instructive. Two breaks on two different metrics, at two different moments, are the normal situation of a real site, and that is where the change log makes the difference.
We have treated breaks as changes of level: the yardstick counts more or less, but it always counts. There is an even more treacherous case, when for a few days or weeks the yardstick stops counting altogether: a tag removed by mistake, a disconnected property, a consent setup that does not fire. Recognising those gaps, and reconstructing with judgement what we did not measure, is the subject of the next article, on tracking gaps.
Further reading
For the underlying idea, that a number is worth as much as the way it was measured, How to Lie with Statistics by Darrell Huff remains, seventy years on, the quickest and most entertaining read on how data can deceive without anyone lying. For a broader and more modern view of how to read numbers critically, The Art of Statistics by David Spiegelhalter is a clear and honest guide, with no formulas getting in the way.