An annual report often ends with a single line: clicks from Google fell 11.8% compared with the previous year. The line is true, but it says nothing about where it happened. The drop may be concentrated in six months and vanish in the others, it may have started in March, it may be the tail of an event in September. And the annual number, by construction, mixes everything together.
The natural route is the one we saw in the article on time series: decompose the series with decompose() or stl(), remove seasonality, look at the trend. Unfortunately, with two years of data stl() stops with an error. And two years are the most common case there is: Search Console keeps about sixteen months, many sites have had GA4 only recently, annual reports compare one year with the previous one.
The numbers in this article are simulated, but the shape of the problem is the one found in real reports. Let’s see how to remove seasonality when the seasons available are, so to speak, only two.
In short: with only 24 months, classic decomposition cannot be used, because estimating the seasonal component of January requires more than one January. The year-over-year comparison solves the problem by construction: each month is compared with the same month a year earlier, and seasonality cancels out in the ratio. The uncertainty of each ratio is estimated with a Poisson interval; the trend of the year with a regression on the log ratio, weighted by the size of the month. In our example the annual 11.8% drop is all in the first half: −25.1% from January to June, +1.3% from July to December, with an improvement of 3.7 percentage points per month (p = 0.0003).
What we will cover:
Classic decomposition splits a series into three pieces: the trend (where it is going, net of fluctuations), seasonality (the profile repeated every year) and the residual (what is left). To estimate the seasonal profile of January, the method needs to compare several Januaries with each other: with only one January per year, and two years in total, it has two observations per month and no way to distinguish «January is low because of the season» from «this January is low because something changed».
This is not a software limitation, it is a limitation of the information available. And the software says so without beating around the bush. Let’s simulate 24 months of clicks with multiplicative seasonality and a drop in the second year:
seasonal <- c(1.08, 0.94, 1.04, 0.90, 1.02, 0.86,
0.70, 0.66, 1.06, 1.16, 1.22, 1.12)
base <- 9000
month <- rep(1:12, 2)
level <- c(rep(1, 12), rep(0.75, 6), rep(1, 6))
set.seed(2026)
clicks <- rpois(24, base * seasonal[month] * level) Now try the decomposition:
stl(ts(clicks, frequency = 12), s.window = "periodic")
# Error: series is not periodic or has less than two periods The error is blunt, and it is worth reading as information: there are not enough seasons to estimate the season. With 36 or 48 months things improve, but remain fragile, and the 24-month case — the one that lands on the desk every January — stays uncovered.
There is a simpler solution than decomposition, and it works precisely because it estimates nothing: compare each month with the same month of the previous year. If December is always the highest month of the year, this December against last December does not suffer from December being high: the seasonal factor is present in both terms and cancels out in the ratio.
In formulas, for month t of the second year:
\(r_t = \frac{x_t}{x_{t-12}}\)
where x_t is the value of the month considered and x_{t-12} the same month a year earlier. A ratio of 1 means «same level», 0.90 «ten percentage points less», 1.05 «five points more». The price to pay is an assumption: that the shape of seasonality stayed the same across the two years. We will come back to that shortly, because everything rests on it.
Let’s compute the twelve ratios on our example:
ratio <- clicks[13:24] / clicks[1:12]
round(ratio, 3)
# [1] 0.751 0.761 0.758 0.754 0.745 0.724
# [7] 1.005 1.031 1.022 1.003 1.017 1.004 The reading is immediate, and it is also why this method is preferable to decomposition: the first six months are all around 0.75, the last six around 1.00. A 25% drop that lasts six months and then disappears.
The annual total, meanwhile, tells a different story: 93,101 clicks against 105,570, that is −11.8%. The average of a −25% and a +1% produces a number that describes neither period.
A ratio computed on a few thousand clicks is not an exact number. As in the comparison between two periods we discussed, each ratio must come with a confidence interval. The tool is the same: a Poisson test between two counts, applied month by month.
ci <- t(mapply(function(a, b) poisson.test(c(a, b))$conf.int,
clicks[13:24], clicks[1:12])) | Month | Ratio | 95% interval |
|---|---|---|
| January | 0.751 | 0.73–0.77 |
| February | 0.761 | 0.74–0.79 |
| March | 0.758 | 0.73–0.78 |
| April | 0.754 | 0.73–0.78 |
| May | 0.745 | 0.72–0.77 |
| June | 0.724 | 0.70–0.75 |
| July | 1.005 | 0.97–1.04 |
| August | 1.031 | 0.99–1.07 |
| September | 1.022 | 0.99–1.05 |
| October | 1.003 | 0.98–1.03 |
| November | 1.017 | 0.99–1.04 |
| December | 1.004 | 0.98–1.03 |
The message is stronger than the ratio alone: the first six intervals all sit below 1, the last six all around 1. This is not a matter of nuance: from January to June the drop is there, from July to December it is not. With these volumes the uncertainty is narrow enough to allow clear statements; with a site ten times smaller the same intervals would widen and some conclusions would have to be suspended.
The twelve ratios tell a story, but a story needs a summary. The useful question is: is the comparison with the previous year improving month after month, or is the pattern just an oscillation? To answer, we relate the logarithm of the ratio to the month, and weight each month by its size: a month with 10,000 clicks must count more than a month with 6,000.
Why the logarithm? Because ratios multiply, they do not add: going from 0.75 to 0.80 is not the same jump as going from 1.00 to 1.05, and on the log scale percentage changes become differences. The model is:
\(\log(r_t) = a + b \cdot t + \varepsilon_t\)
where t is the month of the second year (from 1 to 12), b is the monthly change in the log ratio and ε_t is the error. The slope b, once exponentiated, reads directly as a percentage change per month. I estimate it in R:
model <- lm(log(ratio) ~ I(1:12), weights = clicks[1:12])
exp(coef(model)[2]) - 1
# [1] 0.0367
exp(confint(model)[2, ]) - 1
# [1] 0.0212 0.0524
summary(model)$coefficients[2, 4]
# [1] 0.00033 The result reads like this: the comparison with the previous year improves by 3.7 percentage points per month, with an interval from 2.1 to 5.2 and a p-value of 0.0003. It is not an oscillation: it is a trend. Extending the slope over twelve months would give a 54% recovery (1.037 to the twelfth power), but that is an exercise of limited value: the slope describes the stretch observed, it does not guarantee that the recovery continues.
There is a subtlety worth keeping in mind. The slope measures how much the comparison improves, not why. A drop that stops abruptly in July produces the same slope as a gradual recovery spread over twelve months. The regression summarises, it does not tell the story: to understand the shape we have to go back to the chart of ratios.
At this point the operational conclusion is ready, and it is not «the year fell 11.8%». It is: the first half lost a quarter of the clicks, the second is in line with the previous year. In detail, the first six months go from 52,399 to 39,257 clicks (−25.1%), the last six from 53,171 to 53,844 (+1.3%).
This reading changes the questions to ask. If the drop had been uniform, we would look for a cause acting all year: a migration, an algorithm change, a market trend. A drop concentrated in the first half of the year, which then fades, points instead to an event with a date: something that started before January and stopped acting around July.
The annual report would never have said so. The year-over-year comparison does, without impossible decompositions.
There is an alternative that sometimes proves useful: working on daily data. With 730 days available, decomposition does not get stuck, because weekly seasonality has dozens of cycles:
stl(ts(daily_clicks, frequency = 7), s.window = "periodic") This decomposition removes the weekly profile — Monday different from Saturday — which is often the most annoying source of noise in daily data. It does not remove annual seasonality, which stays there and cannot be estimated from days with only two years either. It can be a good complement, not a substitute for the same-month year-over-year comparison.
A warning: the year-over-year comparison removes seasonality only if the season is the same in both years. A change in how we measure (cookie consent, a new filter, a change in the definition of a conversion) or a one-off event (a campaign, a launch, a pandemic) enter the ratio as if they were season or trend. Before attributing the drop to the site, it is worth asking whether the yardstick changed over the same period.
Yes, and with four years stl() becomes usable. But the year-over-year comparison remains more readable when the question is «how are we doing compared with last year», because it does not require choosing a window or a smoothing method. With long series the two approaches complement each other: decomposition tells the shape, the year-over-year comparison answers the operational question.
Yes, as long as we compare the same week or the same day of the previous year, keeping in mind that Easter and moving holidays fall on different dates. For daily data the noise is high and wider windows are needed: a seven-day moving average before the comparison is often necessary.
As a 12% drop: the ratio is the level of the recent period against the previous one, so 0.88 means we are at 88% of the earlier level. To go from ratio to percentage change, subtract 1. The ratio is convenient because it multiplies and can be compared on a log scale; the percentage change is convenient because it reads well in a report.
That is the main limit of the method. If the seasonal profile has shifted — for example because a piece of content that used to peak in autumn now peaks in spring — the year-over-year ratio attributes to the trend what is a change in season. A quick check: look at whether the monthly ratios follow a regular pattern or jump without a logic. Jumps are a signal to investigate.
Let’s take the simulated data and remove a month: what happens to the slope if the recovery started in September instead of July?
level <- c(rep(1, 12), rep(0.75, 8), rep(1, 4))
set.seed(2026)
clicks2 <- rpois(24, base * seasonal[month] * level)
ratio2 <- clicks2[13:24] / clicks2[1:12]
model2 <- lm(log(ratio2) ~ I(1:12), weights = clicks2[1:12])
exp(coef(model2)[2]) - 1 The slope changes, but stays positive: the regression measures the average improvement, not the date when it happened. That is why the chart of ratios remains indispensable. A second exercise: redo the comparison with 36 simulated months, check that stl() now starts, and compare the extracted trend with the year-over-year ratios.
A doubt remains that the year-over-year comparison cannot settle: the drop may not belong to the site, but to the yardstick. If over the same period Google changed the way it counts impressions, or cookie consent halved measured sessions, the year-over-year ratio records a change that does not belong to the phenomenon. That is the next step: recognising structural breaks and not mistaking them for a drop.
For a text that combines intuition and rigour on time series and regression, The Art of Statistics by David Spiegelhalter is a clear bridge between this article and more advanced models. To understand why forecasts deserve caution, and how much hidden assumptions matter in every model, The Signal and the Noise by Nate Silver is a book I recommend to anyone who writes reports.
The September report is almost ready, and there is one number that does not fit:…
The redesign is approved, the launch date is set, and sooner or later the question…
The annual report lands on the desk with good news on the front page: the…
A site's technical guides climb in the search results, show up more often, and yet…
Clicks and sessions vary more than Poisson allows: how to detect it in R and…
A site receives 612 requests this year and received 640 the year before. The arithmetic…