When traffic starts to seem unusual

Sometimes, the problem does not begin when traffic falls. It begins when there is too much of it.

That is what happened when reviewing the data from a client’s website with a fairly clearly defined audience. The volume was good, even too good at times, but there were elements that did not quite add up.

A significant proportion of traffic appeared as Direct. At the same time, some countries carried a weight that was difficult to justify in the context of the project. Singapore and China, for instance, had a far greater presence than one would expect based on language, activity or target audience.

None of this proved, in itself, that there was a problem.

Direct traffic is not necessarily suspicious. Nor is receiving visits from unexpected countries. The internet does not adhere particularly well to our geographical expectations and, moreover, Analytics cannot always identify precisely where a session comes from.

But when several unusual signals coincide, it is worth stopping looking solely at the total figure.

There were too many visits not to start asking questions.

Where were they really coming from? Why was such a significant proportion being classified as Direct? Did that geographical distribution make sense? Were we seeing people interested in the content, or simply activity that Analytics was adding to the statistics?

And, above all, a question arose before all the others:

were we measuring correctly?

The banner was there, but nobody had to decide

The first clue did not appear in Analytics, but on the website itself.

The cookie banner had been implemented. Technically, it was there. But there was an important difference between displaying a notice and getting the user to make a decision.

In our case, it was possible to continue browsing without explicitly accepting or rejecting it. And that raised a much more relevant question than the banner’s position or design: what was happening to measurement while the user made no decision at all?

This also helps explain why many large websites present consent in a particularly prominent way as soon as you arrive. It often appears centred, with considerable visual weight and even taking up much of the screen. It is not merely an aesthetic decision. It is a way of turning consent into an explicit action that is difficult to ignore, and of reducing the time during which the user continues browsing without having made any decision.

In our case, we did not change just one thing as a result of that suspicion.

We reviewed the consent configuration, the banner’s position and visibility, the clarity of the available options and the implementation of Google Consent Mode, to check which signals were being sent to Google before and after the user made a decision. We also reviewed Analytics’ behaviour and the measurement implementation itself.

It was not simply a matter of “fixing the cookies”. We were reviewing the entire chain that connects a visit with the data that ultimately appears in Analytics: from the user’s interaction with consent to the way that decision is translated into signals and ends up determining what can be measured and how.

This was our response to the problem in front of us, but it does not mean consent is the only tool for investigating unusual traffic.

Google Analytics also allows you to work with filters for certain types of traffic, exclude internal and development traffic, and rely on its own mechanisms for filtering known bots. Depending on the case, it may also be necessary to review tagging, acquisition sources, referrals, events, campaigns or even the infrastructure from which certain visits arrive.

That is why we did not want to turn a correlation into an overly convenient explanation either.

That traffic changed after reviewing consent did not prove that all that unusual traffic had a single cause. Nor did it mean that any subsequent fall could automatically be attributed to the new banner or to Consent Mode.

What mattered was something else: before interpreting a data point, we needed to understand how it was being generated.

In our case, consent and Consent Mode were important enough elements to start there.

And after changing them, the figures changed too.

When less traffic seems like better traffic

After the changes, Analytics began to show less traffic. But at the same time, some of the metrics we use to understand the quality of those visits improved quite noticeably.

The difference was no longer only in how many users arrived on the website, but in how they behaved once there.

The engagement rate increased, bounce rate fell and engagement time became more consistent. More reasonable differences also emerged between new and returning users.

That last point was particularly interesting.

Before the change, Analytics tended to show an almost identical relationship between total users and new users. It was not impossible, but it was unusual when it occurred systematically: it seemed as though almost everyone was arriving for the first time.

After reviewing consent and measurement, that picture began to change. New users still carried a very high weight, but they no longer almost matched the total, and the proportion of returning users began to become visible.

It is not definitive proof that everything had previously been measured incorrectly, but it is an important signal: the audience was beginning to behave in Analytics in a more recognisable and coherent way.

The same applied to other metrics. As some of the suspicious traffic fell away, engaged sessions gained weight and engagement times improved. Overall, things began to seem less noisy.

And this matters because low-quality traffic does not only inflate the number of users. It can also undermine the metrics we later use to assess a website’s performance.

If many sessions come in that last only a few seconds, do not interact and originate from sources that are difficult to explain, the problem is not only that total traffic is overstated. They can also worsen bounce rate, engagement time, recurrence and other signals we later use to make decisions.

That is why, when volume falls after a measurement review but several behavioural metrics improve at the same time, the interpretation should not be limited to “we have lost traffic”.

The right question is another one:

are we seeing fewer users, or are we seeing the users who really matter more clearly?

We cannot claim that everything that disappeared was spam, bots or automated traffic. But when certain anomalies decrease and, at the same time, indicators such as engagement, bounce rate, engagement time or recurrence improve, it is reasonable to consider that some of the previous volume was introducing noise.

And that is where an interesting paradox emerges: a website can show less traffic in Analytics and yet offer a much more useful reading of its audience.

Sometimes, measuring less does not mean knowing less.

It can mean exactly the opposite.

Two periods that no longer mean the same thing

When the way we measure changes, comparing two periods as if they were equivalent can be misleading.

The percentage may be correctly calculated and yet the conclusion may still be wrong.

If a visit could previously be recorded under certain conditions and those conditions then change, we are no longer comparing exactly the same observed reality. The metric’s name may remain the same —users, sessions, Direct traffic, engagement rate— but the way those data reach Analytics is no longer identical.

That requires us to introduce a kind of break point into the historical reading.

It does not mean that we should stop comparing. It means that we should compare with context.

A fall in traffic after reviewing consent, Consent Mode or implementation cannot automatically be interpreted as an equivalent loss of audience. Likewise, an improvement in engagement time or bounce rate should not be presented as though the entire improvement came solely from content or user experience.

Part of the change may lie in the audience’s actual behaviour.

And part of it may lie in who now enters the sample and under what conditions.

This is where consent UX becomes important again.

The way we present the banner, its visibility, the clarity of the options and the ease of accepting or rejecting them can influence how many users make a decision and when they make it. That decision also ends up affecting the data available to measure the session.

That is why consent design does not exist solely in the sphere of privacy or legal compliance. It is also part of the measurement architecture.

A small change to an interface can end up having consequences far beyond that interface: in reports, in monthly comparisons and, ultimately, in the decisions we make based on them.

Perhaps one of the most useful things we can do when we change this kind of implementation is to acknowledge it in our own reports.

Not hide the break.

Mark it.

Because from that point onwards, the “before” and “after” remain useful, but they no longer mean exactly the same thing.

Conclusion

For a long time, we have associated good measurement with having more data.

More users. More sessions. More events. More page views. More information to analyse.

But more does not always mean better.

If some of the traffic we record introduces noise, distorts the origin of visits, artificially worsens behavioural metrics or makes an audience seem homogeneous when it is not, then accumulating more data can take us further away from reality rather than closer to it.

That is probably the most interesting aspect of this whole process.

The initial fall seemed like bad news because we were looking at volume. But when we began to observe what had actually changed, the picture was different: less difficult-to-explain Direct traffic, fewer unlikely geographies, better engagement signals and an audience that was beginning to behave in a more coherent way.

It does not mean that we have found perfect measurement.

Nor does it mean that everything that disappeared was invalid traffic.

It means something more useful: the quality of a metric depends as much on the data it collects as on the conditions under which it collects it.

And that requires us to change a fairly widespread habit.

When a figure rises, we should not automatically celebrate it.

When it falls, neither should we assume that something has got worse.

First, it is worth asking what we are measuring, how we are measuring it and whether the sample before us still means what it meant before.

Because a fall can be a loss.

But it can also be a clean-up.

It can be a break in the series.

It can be a more correct implementation.

Or it can be the first time the data begins to resemble reality a little more closely.

In the end, measuring better is not about collecting everything possible.

It is about understanding better what we are seeing.

Are we making decisions about what our users do, or about what our tool manages to record about them?