The sources
Understanding Google Analytics 4
GA4 is excellent at some things and structurally incapable of others. The gap between those two lists is where most reporting arguments start.
GA4 is a behaviour and traffic analysis tool. It tracks how people move around your website: where they arrive from, what they look at, how long they stay, and where they abandon a process. It is not a revenue system, and treating it as one causes more trouble than any other single habit in marketing reporting.
What GA4 is genuinely good for #
- Understanding which channels drive the most engaged traffic, not just the most traffic
- Spotting trends over time — whether organic search is growing quarter on quarter
- Analysing user journeys and identifying where people drop out of a funnel
- Comparing the performance of content, landing pages and site sections against each other
- Identifying peak traffic periods, device mix and audience characteristics
- Diagnosing technical problems — a page that suddenly stops converting after a release
GA4 is at its strongest when you ask it about direction and proportion. It is at its weakest when you ask it for an exact count. Almost every good use of GA4 is a comparison; almost every bad use is a total.
How GA4 attribution actually works #
This is the most misunderstood part of the product, and the source of most “the numbers don't match” conversations — including the ones happening entirely inside GA4.
GA4 does not apply one attribution model across the product. Different dimensions carry different rules, and the rule depends on the scope of the dimension you are looking at.
User scope — prefixed “First user” #
For example, First user source / medium.
This records the campaign that acquired the person in the first place, and holds it for the lifetime of that user as GA4 understands them. It is a first-touch model.
One nuance is worth knowing: within that very first session, GA4 still applies last non-direct logic to decide what the acquiring source was. So it is more precisely “the source of the first session, resolved by last non-direct” than pure first click. In practice, treat it as first touch.
Session scope — prefixed “Session” #
For example, Session source / medium.
This records the source for each individual session, using last non-direct click. If someone arrives by typing your URL directly, but had clicked a LinkedIn ad within the attribution window, GA4 overwrites “direct” and credits LinkedIn for that session.
This is why direct traffic in GA4 is usually lower than you would expect, and why session-level numbers move around when nothing about your marketing has changed.
Event scope — no prefix #
For example, Source, Medium, Campaign, Default channel group.
This is the one that carries your configured attribution model. Google's documentation is explicit: for event-scoped traffic dimensions, GA4 applies the attribution model selected in your property settings, which is data-driven by default. User-scoped and session-scoped dimensions are unaffected by that setting.
An unprefixed Source / Medium is not a raw collected value, and it is not last click. It is whatever model you have configured. Build an Exploration on unprefixed dimensions believing you are looking at last click, and you are quietly reading data-driven output instead.
Where the configurable model actually lives #
The attribution model you can change — data-driven, last click, and the paid and organic variants — sits in Admin › Data display › Events › Attribution settings. You need the Marketer role or above at property level to see it.
It applies to every report and Exploration that uses an event-scoped traffic dimension, not just the Advertising section. It does not change the standard Acquisition reports, because those are built on user- and session-scoped dimensions.
This matters in both directions. It is a common mistake to open Reports › Acquisition › Traffic acquisition, look at a key event count and assume you are seeing data-driven output — you are not. It is an equally common and more expensive mistake to build an Exploration on an unprefixed Source / Medium, assume last click, and unknowingly read data-driven output.
| Where you are looking | Model applied |
|---|---|
| Reports › Acquisition › User acquisition (First user…) | First touch, user scope — not configurable |
| Reports › Acquisition › Traffic acquisition (Session…) | Last non-direct click, session scope — not configurable |
| Any report or Exploration using unprefixed Source / Medium / Campaign / Default channel group | Your configured model — data-driven by default |
| Advertising › Attribution › Conversion paths and Performance | Your configured model |
Data-driven attribution splits credit fractionally across touchpoints, so it produces decimal values. Last click never does. If you see 14.7 key events, you are looking at an event-scoped dimension under a data-driven model. If you see a whole 15, you are in a standard Acquisition report.
Lookback windows #
The attribution lookback window is also configurable, and it changes results.
GA4 defaults to 30 days for acquisition key events — which means specifically first_open (app install) and first_visit (first site visit), not first purchase or first sign-up. Seven days is the only alternative. For all other key events the default is 90 days, with 30 or 60 also available.
One consequence is routinely missed: the key event lookback window you choose also governs session attribution. Change it and your session-scoped source numbers move too. That is usually the explanation when historical channel splits appear to shift without anyone touching a campaign.
Longer B2B sales cycles routinely exceed even the 90-day maximum, which means touchpoints that genuinely contributed simply fall outside the window and receive no credit at all.
Why data goes missing, and what is actually causing it #
The common explanation is third-party cookie deprecation. That explanation is wrong in two ways, and it is worth correcting because it points at the wrong fix.
GA4 does not rely on third-party cookies #
It sets a first-party cookie (_ga) on your own domain. Third-party cookie policy affects advertising platforms and cross-site retargeting, not GA4's core measurement.
Chrome did not remove third-party cookies, and no longer intends to #
The sequence is worth knowing, because most guides stop partway through it:
- July 2024 — Google drops outright deprecation and proposes a one-time user-choice prompt instead.
- April 2025 — Google confirms it will not ship that prompt either. Third-party cookies simply stay.
- 17 October 2025 — Google announces the retirement of most Privacy Sandbox APIs, including Topics, Protected Audience and the Attribution Reporting API. Deprecated from Chrome 144 in January 2026, removal targeted for Chrome 150 in July 2026. CHIPS, FedCM and Private State Tokens survive.
- The UK Competition and Markets Authority subsequently released Google from its Privacy Sandbox commitments and closed the investigation.
So not only did the deprecation not happen, the replacement technology was retired as well. Any guide still describing either as imminent is out of date. None of this changes your consent obligations — POPIA and GDPR requirements are exactly as they were.
The real causes #
Safari's Intelligent Tracking Prevention. Script-writable storage — which includes any cookie set through document.cookie, as GA4's is by default — is deleted after seven days without user interaction with the site. It is a seven-day idle timer, not a hard cap from creation. A tighter 24-hour cap applies specifically to cookies written on a landing page reached via a decorated link, meaning one carrying a query string or fragment, from a domain ITP has classified as a tracker. When the cookie is deleted, a returning visitor is counted as brand new: New Users inflate, returning visitors deflate, long-window attribution quietly dies.
The assumption usually runs the wrong way. Safari accounts for roughly 11% of South African browsing and iOS for about a quarter of mobile devices (Statcounter, July 2026), against Chrome at 73.5%. South Africa is an Android- and Chrome-dominant market with iOS share below the global average. ITP is a real problem, and iOS users frequently over-index on order value so the distortion lands on your best customers — but on a typical South African site, ad blockers and consent declines almost certainly cost you more signal than ITP does.
Firefox Enhanced Tracking Protection, which behaves similarly for known trackers.
Ad blockers and privacy extensions, which prevent the GA4 tag firing at all. Depending on audience this ranges from around 10% to well over 30% for technical audiences.
Consent banners. If a visitor declines analytics consent, GA4 either receives nothing or receives cookieless pings, depending on your Consent Mode configuration.
Cross-device journeys. Someone who researches on a phone and buys on a laptop is two users, unless you have User-ID or Google Signals stitching them together.
On “modelled” data #
You will see claims that a growing share of GA4 data is modelled rather than measured. This is true for some properties and not others, and the distinction matters.
Behavioural modelling #
Only activates if you have Consent Mode deployed across all pages and your property meets Google's published thresholds:
- at least 1,000 events per day carrying
analytics_storage='denied', for at least 7 days, and - at least 1,000 daily users sending
analytics_storage='granted'events on at least 7 of the previous 28 days
Meeting those numbers does not guarantee modelling — Google applies further quality criteria, including the ratio of new to returning users. Many properties model nothing at all, because they never meet the criteria. If yours does model, GA4 flags it with a data-quality icon in the reporting interface.
Conversion modelling #
Fills gaps where a click and a conversion cannot be directly joined, most often in Google Ads reporting.
Data thresholding — a different thing entirely #
When Google Signals is enabled and volumes are low, GA4 withholds rows to prevent individual identification. This looks like missing data but is a privacy suppression, not modelling. Switching reporting identity away from Signals often makes it disappear.
What GA4 is not for #
Use it for
Traffic trends. Funnel and drop-off analysis. On-site behaviour. Landing page comparison. Audience and device mix. Detecting that something broke.
Do not use it for
Revenue reporting. Exact conversion counts. Cross-channel budget decisions. Anything with a sales cycle longer than your lookback window. Matching your CRM.
Do not use it as your source of truth for revenue. Client-side ecommerce tracking drops transactions on payment redirects, network failures and consent declines. A server-side implementation with transaction IDs reconciled against the back end can be accurate — but that is a deliberate build, not the default.
Do not expect GA4 to match your CRM. It will not, and the gap is structural.
Do not evaluate campaign ROI from GA4 in isolation, particularly for anything with a sales cycle longer than the lookback window.
One correction to conventional advice #
You will often be told never to compare GA4 against your CRM. That is slightly too blunt, and it costs you a useful diagnostic.
You should not expect the numbers to match. You absolutely should compare them directionally. If GA4 records 400 form submissions and HubSpot records 180 contacts for the same period, that gap is telling you something real: a duplicate tag, bot traffic, a broken form handler, a lifecycle stage misconfiguration, or contacts landing somewhere you are not looking. That is a fixable problem, and you would never have spotted it without the comparison.
Reconcile directionally. Investigate large or widening gaps. Never present either number as the other.
Common questions #
Why do my GA4 conversions have decimal places?
Because you are looking at an event-scoped traffic dimension under data-driven attribution, which splits credit fractionally across touchpoints. Standard Acquisition reports use user or session scope and always show whole numbers. The decimals are not an error — they are a signal about which model you are reading.
Why is my direct traffic so high?
Direct is GA4's fallback bucket for anything with no referrer and no campaign tags. The usual causes are untagged campaigns, links shared in messaging apps and email clients that strip the referrer, app-to-web traffic, and redirects that drop query strings. It is nearly always a tagging problem rather than a genuine surge of people typing your URL. Start with UTMs and campaign taxonomy.
Why does GA4 show fewer users than my server logs?
Server logs record every request, including bots, crawlers and users who blocked your tags. GA4 only ever sees visitors whose browser executed its JavaScript and who did not decline consent. A 20–40% gap is normal. If server logs are dramatically higher, you are probably looking at bot traffic that GA4 correctly filtered.
Should I switch my attribution model away from data-driven?
Usually not, but understand what changing it does. Switching to last click makes event-scoped numbers whole and easier to explain, at the cost of crediting only the final touchpoint. The more important discipline is consistency: pick a model, document it, and do not change it mid-year without annotating the date, or your year-on-year comparisons become meaningless.