How GA4 counts users and sessions<\/h2>\n\n\n\n
GA4 doesn’t count users one by one. It uses a probabilistic algorithm called HyperLogLog++<\/strong><\/a> (HLL++), which estimates the number of unique values without storing every single record.<\/p>\n\n\n\n It’s a smart trade-off: instead of storing millions of unique IDs in memory, GA4 gets by with 12 KB of data and 16,384 “buckets.” The result? An estimate with a margin of error:<\/p>\n\n\n\n So 100,000 users in GA4 actually means somewhere between 98,400 and 101,600 in reality. For most decisions, that’s fine. For statistical tests on large samples (over 12,000 users per variant), it starts to matter.<\/p>\n\n\n\n Important clarification: HLL++ only affects unique counts<\/strong> \u2014 users and sessions. Event counts (pageviews, clicks, conversions) are exact. But watch out: when you calculate a conversion rate<\/strong> with users in the denominator, the HLL++ imprecision carries over.<\/p>\n\n\n\n This is the smaller deviation, though. Let’s move to the bigger ones.<\/p>\n\n\n\n Safari (ITP), Brave, Firefox (ETP), and DuckDuckGo block Google Tag Manager by default. Add uBlock Origin and other ad blockers on top.<\/p>\n\n\n\n Result: GA4 misses roughly 10% of your visitors<\/strong> due to ad blockers and browser tracking prevention. For a tech-savvy audience, it can be more. Some international sources report up to 30%, but in my experience on European projects, it’s typically around 10%.<\/p>\n\n\n\n A solution exists: server-side tracking<\/a> can recover a significant portion of this data because your server sends the data, not the user’s browser. You can also serve GTM from your own domain (careful \u2014 don’t use CNAME, use something like Cloudflare Routes instead) and send data to your own domain. This bypasses most blocking rules.<\/p>\n\n\n\n If your cookie banner is set up correctly (and a third of companies don’t have theirs right<\/a>), GA4 only measures visitors who clicked “Accept.”<\/p>\n\n\n\n Opt-in rates typically range between 40\u201370%. That means 30\u201360% of visitors<\/strong> GA4 either doesn’t see at all (Basic Consent Mode) or tries to model (Advanced Consent Mode \u2014 more on that shortly).<\/p>\n\n\n\n This is, in practice, the biggest source of inaccuracy<\/strong> in GA4 \u2014 yet many people focus on ad blockers instead of checking their opt-in rate.<\/p>\n\n\n\n By the way, I described how to calculate opt-in rate<\/a> in a separate article.<\/p>\n\n\n\n In Explorations, GA4 processes a maximum of 10 million events per query<\/strong> (free version). Anything above that gets sampled. Average error is around 5%, but for small data ranges it can be up to 30%<\/strong>.<\/p>\n\n\n\n Standard reports aren’t sampled \u2014 they’re pre-calculated. But they have a different problem.<\/p>\n\n\n\n Standard reports have a limit of 500 unique dimension values per day<\/strong>. Everything above that falls into the “(other)” row.<\/p>\n\n\n\n Running an e-commerce store with 2,000 products and want to see revenue by product in a standard report? Most of them will be in “(other).” Solution: Explorations (higher limit) or BigQuery (no limit).<\/p>\n\n\n\n If your reports use demographic dimensions or Google Signals data, GA4 may silently hide rows<\/strong>. It doesn’t want anyone to identify individual users from the data.<\/p>\n\n\n\n Google doesn’t say exactly when. From experience: typically at 30\u201350 users or events. Solution: switch reporting identity to “Device-based.”<\/p>\n\n\n\n Note: Google removed Google Signals from reporting identity in 2024 \u2014 Signals are still used for remarketing and demographic data, but they no longer affect reporting identity itself. Thresholding still applies when using demographic dimensions.<\/p>\n\n\n\n GA4 offers three modes: Blended, Observed, and Device-based. Each gives different numbers for the same data<\/strong>. Blended uses modeling, Device-based uses only cookies. Differences in user counts can be in the tens of percent.<\/p>\n\n\n\n And that brings us to modeling.<\/p>\n\n\n\n Consent Mode v2 in Advanced mode sends anonymous pings even from users who haven’t consented. GA4 uses these to model missing data.<\/p>\n\n\n\n Sounds great. The catch:<\/p>\n\n\n\n Recovery estimates: 30\u201350% of lost conversions can be recovered. That’s not nothing. But it’s not 100%.<\/p>\n\n\n\n And most importantly: modeling on vs. off dramatically changes your numbers.<\/strong> We’ll come back to this.<\/p>\n\n\n\n GA4 uses different attribution models depending on the report type. In standard reports: User Acquisition<\/strong> uses first-click, Traffic Acquisition<\/strong> uses last non-direct click. Data-Driven Attribution (DDA) only applies to Key Events reports and Explorations \u2014 and even there, only if you have enough data.<\/p>\n\n\n\n Why does GA4 show a different traffic source in User Acquisition, a different one in Traffic Acquisition, and yet another in Explorations? Exactly for this reason \u2014 each report uses a different attribution model. More on this in my article about traffic sources in GA4<\/a>.<\/p>\n\n\n\n One more thing that doesn’t get discussed enough: GA4 has a latency of 24\u201348 hours<\/strong>. Universal Analytics was significantly faster. On larger websites, I often see data appearing in reports only by the next afternoon. Some data (offline events, late pings) can arrive up to 72 hours<\/strong> after the event \u2014 anything later gets discarded by GA4. And there’s no SLA.<\/p>\n\n\n\n For analysis, this doesn’t matter. For real-time decision-making, it does.<\/p>\n\n\n\n This needs to be said out loud. GA4 isn’t perfect, but it’s still the most accessible and widely used tool<\/strong> for web measurement. Having no measurement at all is incomparably worse.<\/p>\n\n\n\n The key is to know how accurate your data is and how much you can trust it<\/strong>.<\/p>\n\n\n\n For tracking trends and comparing periods, GA4 works well. The deviation is consistent, so even though absolute numbers aren’t precise, relative changes are<\/strong>. Campaign A performs better than Campaign B? GA4 will tell you that reliably.<\/p>\n\n\n\n Once you need absolute numbers, it’s a different story. For calculating CPA or ROAS, it makes a massive difference whether you have modeling turned on or not.<\/strong><\/p>\n\n\n\n With modeling, GA4 reports tens of percent more conversions \u2014 and your ROAS suddenly looks great. Without modeling? Same campaign, but the numbers tell a different story. And you’re making budget allocation decisions based on this.<\/p>\n\n\n\n This isn’t an academic problem. This is a problem that costs real money.<\/p>\n\n\n\n If you’re comparing revenue in GA4 with your e-commerce backend numbers<\/a>, you need to know exactly which reporting method you’re using \u2014 otherwise you’re comparing apples to oranges.<\/p>\n\n\n\n BigQuery export from GA4 contains:<\/p>\n\n\n\n Even in BigQuery, though, you need to consider how much precision you actually need. Aggregation functions like If knowing “roughly 98,000 users” instead of “exactly 97,842” is good enough, you’ll save money. And as I wrote in my Lean Analytics article<\/a>, you don’t need to measure everything \u2014 just what matters.<\/p>\n\n\n\n\n
What GA4 doesn’t see at all<\/h2>\n\n\n\n
Ad blockers and tracking prevention<\/h3>\n\n\n\n
Consent and cookie banners<\/h3>\n\n\n\n
What GA4 hides or distorts<\/h2>\n\n\n\n
Sampling<\/h3>\n\n\n\n
Cardinality \u2014 the “(other)” row<\/h3>\n\n\n\n
Thresholding<\/h3>\n\n\n\n
Reporting identity<\/h3>\n\n\n\n
Data modeling \u2014 a safety net with holes<\/h2>\n\n\n\n
\n
Attribution \u2014 a black box<\/h2>\n\n\n\n
\n
Data latency<\/h2>\n\n\n\n
What to do about it<\/h2>\n\n\n\n
Having data is better than having none<\/h3>\n\n\n\n
For trends, GA4 is enough<\/h3>\n\n\n\n
For absolute numbers, watch the modeling<\/h3>\n\n\n\n
BigQuery: accurate data, but at a cost<\/h3>\n\n\n\n
\n
APPROX_COUNT_DISTINCT<\/code> (yes, the same HLL++ as in GA4) can dramatically reduce query costs<\/strong> in BQ \u2014 at the cost of that same \u00b11.6% deviation.<\/p>\n\n\n\n![\"\"](\"https:\/\/www.sabatka.net\/wp-content\/uploads\/2026\/04\/null-did-you-know.png\")
<\/figure>\n<\/div>\n\n\n\nScientific note<\/h3>\n\n\n\n