Google Analytics 4: The Complete Guide

GA4 is not a redesign of the previous Google Analytics. It is a different measurement model wearing a similar interface, and most of the confusion around it comes from expecting old definitions to survive. Sessions are derived rather than primary. Bounce rate is calculated backwards from engagement. Almost everything is an event.

This guide covers what the model does, how to set it up without creating problems you will find in six months, and where the reports mislead. If you are still deciding whether this is the right tool at all, start with how to choose a web analytics tool and come back.

The data model

Four concepts carry the whole product.

Events. Everything a user does is an event with a name: page_view, scroll, click, purchase. There is no separate concept of a “hit type” — a page view and a transaction are the same kind of object with different names.

Parameters. Details attached to an event: which page, which link, what value, which currency. Parameters are where the useful specificity lives, and they are the part teams most often forget to register (see the traps below).

User properties. Attributes of the person rather than the action — plan tier, account type, language preference. Set them once; they apply to subsequent events.

Key events. What used to be called conversions. Google renamed conversions to key events in 2024, keeping “conversions” for the advertising side. Marking an event as a key event is a toggle, not a separate object — which means any event can become one, and a badly named event becomes a badly named key event.

Events come from four places: automatically collected ones you cannot turn off, enhanced measurement (scrolls, outbound clicks, site search, file downloads, video engagement), recommended events with reserved names and expected parameters, and fully custom ones you define.

Engagement, and why bounce rate looks strange

A session counts as engaged when it lasts longer than ten seconds, includes a key event, or includes at least two page views. Engagement rate is the share of sessions that qualify. Bounce rate is one minus that.

The consequence is that a “bounce” in GA4 and a bounce in the older tool are not the same measurement, and comparing the two across the migration is meaningless. A visitor who reads a page for a minute and leaves is engaged here and a bounce there.

The ten-second threshold also makes engagement rate the closest thing the product has to a reading signal — which is the practical link to dwell time, a metric no analytics tool can measure directly.

Setting it up without future damage

Six settings, in the order that matters.

Data retention. The default for user- and event-level data is two months; fourteen is the maximum. This affects Explorations only — standard reports are unaffected — but once a period has passed, the detail is gone and cannot be restored. Decide deliberately, and note that a shorter period is easier to defend under the GDPR.

Internal traffic. Define your own IPs and mark the filter active, not testing. A developer refreshing a page all afternoon is a measurable share of a small site’s traffic.

Cross-domain measurement. If checkout lives on another hostname, configure it before launch. Without it, every purchase is attributed to a referral from your own site.

Enhanced measurement. Useful defaults, with one caution: on a single-page application, history-based page views can multiply counts. Verify what fires on a route change before trusting the numbers.

Consent mode. If you serve the EEA and use Google’s advertising products, consent signals are required, and behaviour without consent is modelled rather than observed. Know which of your numbers are which.

Custom dimensions. Register the parameters you send. This is the setting people discover too late — see below.

Reports and Explorations

Standard reports answer fixed questions quickly: acquisition by channel, pages and screens, events, key events. They are the right surface for a weekly read and for anything you will look at repeatedly.

Explorations are the ad-hoc layer — free-form tables, funnels, path analysis, segment overlap. Two things to know before relying on them. They obey the data retention setting, so a two-month property cannot explore last quarter. And large queries can be sampled; the card in the corner tells you when, and it is worth checking before quoting a number.

Link the property to Search Console and enable the two search reports. That connects impressions and queries to on-site behaviour — the pairing described in the Search Console integration guide, which also covers the API version for anything beyond the interface.

The traps that produce unreproducible numbers

The (other) row. When a dimension has too many distinct values, GA4 groups the tail into a single (other) bucket. High-cardinality parameters — full URLs with query strings, timestamps, user IDs — are the usual cause. The fix is upstream: send fewer distinct values.

Custom dimensions are not retroactive. A parameter sent for months but registered as a custom dimension today shows data only from today. Register first, send second.

Thresholding. When Google signals are enabled and a report could identify an individual, rows are withheld. Totals then exceed the sum of visible rows, and the demographics report looks half empty on small sites.

Real-time is not processed data. The real-time view uses a different pipeline. Numbers there will not match the same day’s report once processing completes, which usually takes up to a day or two.

Attribution defaults. Data-driven attribution distributes credit across the path, so channel numbers do not add up the way last-click did. Neither is wrong; mixing them in one deck is.

Definitions drift. Google renames things — conversions became key events, and metric definitions have shifted more than once. Keep your reporting definitions in a document outside the tool, or your year-on-year comparison will be comparing two different measurements.

BigQuery export

The export sends raw event rows to BigQuery on a daily schedule, and it is the single most useful thing a serious GA4 property can turn on. It removes sampling, removes the retention limit for anything you have already exported, and lets you write the query the interface will not.

Two practical notes: turn it on before you need history, since it is not retroactive, and watch storage growth on high-traffic sites. The daily export is generous, but a busy property accumulates.

Reading it week to week

A short routine beats an elaborate dashboard. Landing pages with sessions and key events, to see what search sends and what it does. Channel groups, to catch mislabelled traffic. Engagement rate on the pages you have recently changed. When a whole subject moves at once, check the season before blaming the site — Google Trends answers that in a minute. Key events against the previous four weeks, not against yesterday.

For anything that needs to be caught out of hours — a purchase event that stopped firing, a tag removed in a deploy — the interface is the wrong instrument. That job belongs to alerting on a handful of metrics.

When the audience mix shifts, the split between first-time and repeat visitors is where it shows first; what that ratio means depends entirely on the business model, which is the subject of new versus returning customers.

Where GA4 stops

It will not tell you what people searched before clicking — that is Search Console. It will not survive a badly planned URL change, which is a redirect problem covered in handling migrations without losing rankings. It cannot see the visitors who declined consent, and the gap grows as browser storage keeps shortening — the shift described in life after third-party cookies.

What it does well is answer, at no licence cost, what happened on your site and which entry points earn attention. Set retention deliberately, register your parameters, export to BigQuery if you have the volume, and write your definitions down. Most GA4 complaints are really complaints about one of those four being skipped. If the basics feel shaky, the definition of web analytics is the shorter place to start.