GA4

Universal Analytics organised the world around sessions and pageviews. GA4 organises it around events and users, and applies its own identity resolution across device, user ID and modelling on top. That is not a refinement of the old model, it is a different one.

The consequence is that metrics sharing a name across the two products are not the same measurement. Bounce rate changed from single-page visits to the inverse of engagement. Sessions stopped restarting when the campaign source changes mid-visit. Users are resolved differently. Conversions count every occurrence rather than one per session.

So a chart that runs across July 2023 is two datasets joined by a line, and any year-on-year comparison spanning it is comparing instruments rather than performance. Presenting that difference as a result, in either direction, is the single most common GA4 reporting error.

The practical handling is to draw the line explicitly. Annotate the migration date on every chart that crosses it and state the comparison window in the commentary. A client who can see where the instrument changed stops reading the step as a change in the business.

The free tier retains event-level data for a maximum of fourteen months, and the default when a property is created is two months. Aggregated reports keep working past that; the underlying events, the ones needed for any analysis nobody anticipated, do not.

This is the one GA4 setting with an irreversible consequence. A client asking in year two for a three-year comparison, or for a cohort analysis nobody imagined at kickoff, cannot have it unless the export was configured in year one. There is no way to recover data that was never retained.

The BigQuery export is the answer and it is free to configure, with storage and query costs that are negligible at the volumes agency clients produce. The barrier is not cost, it is that nobody sets it up before it is needed, and by then the window has closed.

The first thing to do on inheriting a GA4 property is therefore to check two settings: retention, which is often still on the two-month default, and whether BigQuery is linked. Both take a minute and both become impossible to fix retrospectively.

GA4 withholds rows when the numbers behind them are small enough that an individual could be identified, particularly when Google Signals is enabled. The report does not error or warn prominently; the rows are simply absent.

The visible symptom is a total that is larger than the sum of the breakdown under it. A client who spots that reasonably concludes the tracking is broken, and it is not: the platform has removed the small rows on purpose.

It bites hardest exactly where agencies do their most detailed work, which is segmented reporting on smaller accounts. Demographic breakdowns, small geographies and narrow date ranges are the ones most often thresholded, and those are the slices a monthly report is built from.

Two things reduce it. Widening the date range increases the numbers behind each row, and turning Google Signals off removes the reason for much of the thresholding, at the cost of the demographic and cross-device reporting it provides. Which trade to make is a client decision rather than a technical one, and it is worth making deliberately rather than discovering it in a meeting.

A GA4 key event and a Google Ads conversion measure different things, and the gap between them is a recurring and entirely avoidable client conversation.

They differ on attribution, since GA4 applies its own model across all channels while Google Ads credits its own click. They differ on window, since the ad platform counts a conversion back to the date of the click and GA4 counts it on the date it happened. And they differ on scope, since Google Ads includes conversions it modelled for users it could not observe.

The result is two defensible numbers that never match. Neither is wrong, and trying to reconcile them to a single figure is a project that produces nothing.

What works is to decide, per client, which number is the one decisions are made on, say so once in writing, and report it consistently. Showing both without explanation invites the client to average them, and switching between them, especially when one looks better, is how an agency loses an argument it was winning.