Synthetic monitoring

Detecting regressions. Two runs under identical conditions differ only because the site changed, which is exactly what a release check needs.

Measuring a page nobody visits yet — a new landing page, a staging build — where real-user data does not exist.

Testing a flow rather than a page: a scripted sequence through a checkout, which no passive measurement can provoke.

Availability checking, which is synthetic monitoring at its simplest and the form most sites run.

The cache is usually empty on every run, while a large share of real visits arrive with assets already stored.

The network is stable and fast, while real conditions vary by an order of magnitude.

The device is a fixed profile, while real traffic spans a range of hardware wider than most teams assume.

Third-party scripts behave differently for an automated client than for a person, and some do not run at all.

So a synthetic figure that improves is real evidence of improvement, and a synthetic figure quoted as what visitors experience is a claim the method cannot support.

Real-user measurement reports what visitors experienced and cannot tell you why, because it observes rather than controls.

Synthetic measurement can be run repeatedly with one variable changed, which is how a cause is isolated.

The practical pattern is to detect with the first and diagnose with the second.

Using only synthetic measurement produces the familiar situation where every internal number looks good and customers report the site is slow.

An external scan is synthetic measurement: one run, from one path, at one moment, with an empty cache.

Which is why its performance figures are reported with the coverage that produced them rather than as a verdict.

It is strong evidence about structure — what is served, how large it is, what blocks rendering — because those do not vary with the observer.

It is weaker evidence about timing, which is why timing findings point at causes rather than at numbers to hit.

Hold the location, the device profile and the connection profile fixed, or a change in the number means nothing.

Run several times and take a median: even under controlled conditions, a single run varies by more than most improvements are worth.

Record what the run measured and what it could not, because a measurement over half the resources is not comparable with one over all of them.

Compare a release against the release before it, not against an absolute target, which is where the method is genuinely strong.