Data pitfalls

None of these make a chart look obviously broken. They make it look perfectly fine, while it quietly means something else entirely.

Distinct people versus a running total

Being seen and opening a page are counted as distinct visitors; a walk-in is counted as a running total.

So a conversion rate divides a total by a distinct count, and can in principle go higher than an intuitive picture of a funnel would ever allow. It isn't simply "the percentage of people who converted."

Being distinct is also per stretch of time. A visitor counted once today gets counted again tomorrow. Adding up seven separate daily distinct-visitor numbers does not give you the real weekly distinct-visitor number. It gives you a bigger number that means nothing at all. Always pull the centre's own actual figure for the exact stretch of time you actually want, rather than adding smaller ones together yourself.

Averaging hides movements that cancel each other out

One single national figure is really an average across countries, devices and traffic sources that may genuinely be moving in opposite directions at once. A flat overall total can secretly be a strong gain in one country cancelling out a real collapse in another.

Each centre splits traffic by source in its own way: browsing, search, links from other shops, links from the open web, and a lightweight preview on one side; search, browsing, and ads and referrals on the other.

Break the number down before you conclude anything at all. Search and browsing respond to completely different levers, and a change in one of them routinely gets mistakenly read as a change in the other.

Minimum thresholds before anything even shows up

One centre requires a small minimum of first-time walk-ins or pre-orders before its own numbers about being seen even appear, and a small minimum of actively used devices before usage numbers appear at all. The other groups any word searched too rarely into a catch-all bucket, and requires enough visitor volume before a specially-targeted listing's own performance shows up in reports at all.

Absence of data is not evidence that nothing happened. A specially-dressed window or a small country showing nothing might simply be sitting below a minimum threshold. This looks identical to genuine failure from the outside, and it genuinely isn't the same thing.

Opt-in samples

One centre's own usage figures are based only on visitors who agreed to share their own data.

A real visit, an actively used device, and a crash are therefore samples of a group whose exact makeup you don't actually know. A trend within that sample is usable; the absolute level isn't, and comparing it to anyone else's own numbers is meaningless.

Numbers about actually acquiring visitors aren't described as opt-in at all. Keep these two families of numbers separate, both in your own head and in anything you report.

Delay

Both centres' own back-offices publish with a delay, and both quietly revise their own numbers afterward. A number read the very next morning is provisional. Deciding something based on same-day data is the single most common cause of a reversal a week later that nobody can explain.

Wait for the numbers to settle before acting on them, and note the date you actually read them on anything you share with others.

Confusions shaped like a rebuild

A gradual, staged rollout of an automatic update can itself be adjusted over a real stretch of time.

During a gradual rollout, the group of visitors on the new version keeps growing, so any number split by version is really measuring a moving target underneath it. Writing changes go live at the moment of a rebuild, but the whole visitor base doesn't switch over all at once.

Also: on the locked side, filed writing changes ship bundled together with a rebuild, so a filing-words change and a genuine code change land at exactly the same moment, and they're confused with each other by construction.

A definition that quietly changed underneath you

One centre shifted its own store-listing reporting from actual acquisitions to clicks recently.

The reports still existed, the charts still rendered perfectly normally, and every single comparison spanning that boundary silently became wrong without anyone being told. This is the single most dangerous kind of confusion of all, and it's exactly why every claim in this whole guide carries the date it was actually last checked.

What tends to go wrong

Adding up several daily distinct-visitor numbers. Produces a number that means nothing at all.

Drawing a conclusion from one national total. An average hides movements that cancel each other out.

Reading an empty report as poor performance. It might simply be below a minimum threshold.

Quoting an opt-in usage figure as an absolute truth.

Acting on same-day data. It's provisional, not final.

Comparing across a point where a definition quietly changed. Nothing about it looks wrong from the outside.