- GA4 counts its headline Users metric as active users (an engaged session), not everyone who arrived, so it undercounts by design. [2][6]
- Standard reports are unsampled, but Explorations sample past 10 million events on standard properties, and up to 1 billion on 360. [1][7]
- A session is "engaged" only if it runs past 10 seconds, records a key event, or logs two pageviews. Bounce rate is just the inverse. [2]
- Consent mode backfills with modeled conversions, but no GA4 doc publishes a recovery rate. Treat modeled totals as inference, not measurement. [1]
- The 2026 default: BigQuery export plus server-side collection as the record, GA4 as a behavioral lens, BI on the warehouse. [1]
The number in the top-left of your GA4 dashboard is not the count of people who visited. GA4's primary Users metric is active users, meaning users with an engaged session, so a visitor who left before the page registered anything never enters the total. Google's own metric definition confirms it: reporting shows active users, meaning users with an engaged session or a recorded engagement signal. [2][6] That is a reasonable choice for content and product analytics. It is a quiet trap if you still read it like a Universal Analytics (UA) visit count, and it is the first of several places where GA4 reports an inference and lets you assume it is a measurement.
What it reads cleanly, and what it fudges
GA4 ships over 100 metrics, and the ones you trust every week fit on a single card. [1] The event model is genuinely strong for behavior: average engagement time counts only active interaction, not idle tabs, which makes it a far better content signal than the old session duration. [3] E-commerce and prediction are the other bright spots, with three native predictive metrics (purchase probability, churn probability, and revenue forecasting) built on Google's models. [4] The trouble starts where GA4 gets read as a ledger.
| Measurement area | Trust it for | Do not trust it for |
|---|---|---|
| Event and engagement | Content and product analytics, active engagement time [3] | UA-style benchmarks; GA4 bounce runs lower for the same content [2] |
| E-commerce and prediction | Average revenue per paying user (ARPPU), purchase and churn probability, revenue forecast [4] | Finance-grade totals under thresholding and consent gaps |
| Acquisition and attribution | Directional channel mix [3] | Precise return on ad spend (ROAS) or cost per acquisition (CPA) when conversions are modeled |
| Local (Business Profile) | Calls, directions, bookings across seven metrics, 6-month window [5] | Campaign or location attribution, or history past 6 months [5] |
Most of that "do not trust it for" column traces back to one redefinition: what GA4 now treats as an engaged session, and the bounce rate it derives from it.
A session counts as engaged if it lasts longer than 10 seconds, records a key event (what GA4 used to call a conversion), or logs at least two pageviews. Bounce rate is the inverse, one minus engagement rate, so a 65% engagement rate reports as a 35% bounce rate. The same page that looked bad in Universal Analytics can read 15-30 percentage points calmer here, in one reference's estimate, because the definition changed, not the behavior.
Why two surfaces disagree
When GA4 totals diverge, the platform's own reconciliation order tells you where the distortion enters. Work it in sequence, because an earlier layer can suppress a row before a later one ever applies. [1]
Thresholding
Suppresses rows on low user counts, so it can hide data before you ever reach a sampling limit.
Checks firstSampling
Explorations estimate from a subset once the event count for the date range crosses the limit.
Cardinality
High-cardinality dimensions collapse rare values into a single "(other)" bucket.
Consent-mode modeling
Declined-cookie traffic and conversions get backfilled with modeled numbers.
Internal-traffic filters
Office and other known addresses are removed from the view last.
That sequence is an admission. The interface applies several layers of privacy protection and estimation, and the raw event stream does not.
Consent mode is the layer most people misread. It backfills traffic and conversions for users who declined cookies with modeled numbers, which is useful for optimization. But no GA4 product doc publishes how much it recovers. The widely repeated "up to 70%" figure lives in vendor blogs, not documentation, so treat any modeled total as inference. Where money is decided, for revenue share, service-level agreements (SLAs), or board reporting, reconcile against backend data, and read the mechanics of consent-mode modeling before you quote a modeled conversion to finance.
Read the GA4 interface as a privacy-filtered view. Keep the BigQuery export as the record.
The export is unsampled, un-thresholded, and event-level, which is why Google's guidance makes it the reference when the UI and the Data API disagree. [1] Serious cohort work, lifetime value (LTV), attribution, and the aggregate feed for marketing mix modeling belong there, joined to customer relationship management (CRM) and order tables, not inside Explorations. It is the same warehouse-first move behind the choice between a packaged customer data platform (CDP) and the warehouse.
Build the stack around the warehouse
The collection layer follows the same logic. GA4 assumes every meaningful interaction is an event, button clicks, form submits, video plays, downloads, and scroll depth, [3] and client-only tags under browser tracking limits and consent pop-ups produce numbers that will not reconcile once you layer business intelligence (BI) on top. Route events through a first-party, server-side endpoint, normalize the schema, and stream to the warehouse. GA4 then becomes one consumer of first-party data rather than the store, and your BI tool sits on warehouse tables keyed to customers and orders, blending GA4 behavior with CRM, Stripe, and customer-satisfaction (CSAT) data instead of reading GA4 alone. [4]
- Read standard reports for unsampled totals. Export to BigQuery the moment an Exploration's data-quality ring turns amber. [1]
- Keep board-grade revenue on backend data. Let modeled conversions drive Google Ads optimization only.
- Route collection through a server-side endpoint with first-party IDs, not client-only tags. [3]
- Model retention on real windows: GA4 exposes 1-, 7-, 28-, and 30-day active-user counts for exactly this. [1]
GA4 in 2026 is a strong behavioral instrument wearing the interface of a system of record. Read it for what it measures cleanly, engagement, product usage, and predictive signals, and stop asking the UI for a precision it never promised. Move the source of truth into the warehouse, and GA4 is finally free to do the one job it was always good at: telling you, honestly, how people behave.
Sources
- Digital Applied · GA4 Dimensions and Metrics: Complete 2026 Referencesampling thresholds, reconciliation order, BigQuery as reference, active-user windows
- Benly · GA4 Dimensions & Metrics: Complete Reference 2026active users as the primary Users metric, bounce as inverse of engagement; the 15-30pp figure is this single reference's estimate
- True Future Media · GA4 Reporting for Business Owners: Metrics That Matter in 2026
- Tatvic · Everything You Need to Know About Google Analytics 4 (GA4) in 2026
- Digital Applied · Google Links Analytics to Business Profiles in 2026
- Google · Active usersGoogle's definition: reporting shows active users, users with an engaged session or a collected engagement signal
- Google · About data samplingsampling starts past 10 million events per query on standard properties, up to 1 billion on 360
Frequently asked questions
Does GA4 sample data?
Standard reports are unsampled. Explorations start sampling once you pass 10 million events for the selected date range on standard properties, and up to 1 billion on GA4 360. The BigQuery export is unsampled and un-thresholded, which is why Google's own guidance treats it as the reference when the UI and the Data API disagree.
Why doesn't GA4's user count match my server logs?
GA4's primary Users metric is active users, meaning users with an engaged session, so it excludes visits that never engaged. Consent-mode modeling then adds a second layer of estimation on top. For precise, financial numbers, reconcile against backend or BigQuery data and treat the interface figure as directional.
How much does consent-mode modeling recover?
No GA4 product doc publishes a recovery rate. Figures like 'up to 70%' circulate in vendor blogs, not documentation, so they are not verifiable. Use modeled conversions to optimize Google Ads and experiments, not to report revenue to a board or settle a contract SLA.
What counts as an engaged session in GA4?
A session that lasts longer than 10 seconds, records a key event (what GA4 used to call a conversion), or logs at least two pageviews. Bounce rate is simply the inverse of engagement rate, which is why GA4 bounce runs lower than the old Universal Analytics number for the same content.
Should I build BI directly on GA4?
Not for serious reporting. Pull raw events through BigQuery, join them to CRM and order tables, and calculate business metrics in the warehouse or BI layer. GA4 then becomes one behavioral input to a Looker Studio or BI dashboard, not the database the whole company reports from.
Why is GA4 bounce rate lower than Universal Analytics?
Because the definition changed, not the behavior. GA4 bounce rate is simply the inverse of engagement rate, so a session that runs past 10 seconds, records a key event, or logs at least two pageviews counts as engaged rather than bounced. In one reference's estimate the same page can read 15-30 percentage points calmer than it did under Universal Analytics.
What predictive metrics does GA4 offer?
GA4 ships three native predictive metrics built on Google's models: purchase probability, churn probability, and revenue forecasting. Along with e-commerce reporting and active engagement time, prediction is one of the areas GA4 genuinely does well. The trouble only starts when the interface gets read as a finance-grade ledger rather than a behavioral instrument.
When two GA4 reports disagree, what should I check first?
Follow the platform's own reconciliation order: thresholding, then sampling, then cardinality, then consent-mode modeling, then internal-traffic filters. Thresholding comes first because it can suppress rows on low user counts before you ever hit a sampling limit. When the interface and the raw data conflict, the unsampled BigQuery export is the reference.
Mikołaj Salecki
Writes about media, tech, and AI business for people who actually run digital. Former agency lead. Skeptic of frameworks that read better than they perform.
Get the next issue by email.
One letter, once a week. Sharp coverage of media, tech, and AI business. No filler.
