In a retailer’s foreign shops, three quarters of visits arrived with no source

The situation

A retailer sells in its home market and runs separate shops in neighbouring countries, with paid search and paid social there. An agency in one of those countries told the client that most analytics traffic had no known source.

What it looked like

In the foreign shops about three quarters of sessions had no traffic source, flat for as far back as the export went, at least three months; in the home market, a few percent. Our first explanation blamed a consent tool overwriting where visitors came from.

What was actually true

The visits were real, and most still carried a referrer or an ad click ID in the raw data, mostly paid search. A bare analytics tag fired at the very start of each page, without the address of the server the rest of the tracking used. Each visitor became two identities, and the one carrying the page views never had its session opened, so no source could attach to it. Separately, a setting on Google’s side, outside the tag manager, sent most page views twice.

How we found it

We queried the analytics export in the data warehouse, splitting sessions by identity, against the home market, which runs the same shop system, consent tool and server. That disproved the consent explanation, and we withdrew it. We also retracted a claim about a hard-coded tag that did not exist, and reverted a same-day fix that made things worse: pausing the basic web tag.

What changed

The first tag now goes through the server, which accepts the foreign shops' IDs; a checkout event that fired too early is replayed; the Google setting is on. We rebuilt three months of weekly traffic sources for the client.

Where it ended

Partly verified. Split identities almost disappeared. On a partly fixed day, sessions without a source fell to about three fifths; the first full day was not read. Page views arrived once in a few test loads, not yet checked in the warehouse.

What we did not claim

That the ad platforms now receive the checkout event: the test browser had consent refused.

Why it matters beyond this case

Some of it is measurement error nobody has ever checked. Not through carelessness: a number that arrives every morning stops being questioned.

All our work →