Why GA4's Anomaly Detection Keeps Telling You the Wrong Story

Introduction

Your GA4 home screen has a little "Insights" panel that quietly claims to be watching your data for you. Sessions dropped. Conversion rate fell. Something happened, and Google noticed. It feels like a safety net — one less thing to manually check every Monday morning.

The trouble is, that safety net has some big holes in it. It isn't very configurable, it often flags the wrong thing for a perfectly understandable reason, and even the AI layer Google is building on top of it only knows Google Analytics in general — not your site, your history, or what actually matters to your business.

None of that makes GA4's anomaly detection useless. It just means it's a starting point, not an answer. This post walks through where it falls short, why some of its flags are actively misleading, and what we've built at Fresh Egg to close the gap.

What GA4's anomaly detection actually does

GA4's built-in anomaly detection uses a statistical model trained on your historical data to predict what a metric "should" look like, then flags anything that falls outside the expected range. It trains on two weeks of data for hourly checks, 90 days for daily checks, and 32 weeks for weekly checks. There's also a segment-level version that uses principal component analysis across dimension combinations to spot segments behaving unusually. Google Support

On top of the automatic version, you can build "custom insights" — simple rules with email notifications, where you pick a metric, a segment, and a trigger condition. Nextflywebdesign

It's a genuinely clever bit of statistics. It's also a long way short of what most teams actually need.

It isn't very configurable

Automated insights aren't yours to shape. You can't point them at specific pages, specific campaigns, or specific user segments, that granularity simply isn't available. Custom insights help, but they're limited to basic threshold conditions, and there's no way to bring in a custom dimension or a custom metric that actually reflects how your business measures success. If the thing you care about isn't one of GA4's standard metrics, GA4 isn't watching it. Anomaly AI

Delivery is limited too. Notifications default to sitting inside the product itself, which most teams never see day-to-day, so "automated monitoring" often just means a report nobody opens. Elsop Insights

It flags the wrong thing, for the right reason

Here's the one that actually causes damage: GA4 will happily tell you your conversion rate dropped, when what really happened is something completely different, and arguably good.

Say a campaign runs that pulls in a burst of low-intent traffic. None of those sessions convert. Sessions go up, conversions stay flat, and conversion rate, a simple ratio, goes down. Mathematically, that's exactly what happened. But it's the wrong headline. The real story isn't "conversion rate has dropped." It's "a specific campaign brought in traffic that was never going to convert, and it's diluting your denominator."

GA4 can't tell the difference, because it doesn't reason about why a number moved, it just measures whether the number is outside its expected range. GA4 tells you what changed but not why; you still have to investigate manually. A tool that can't distinguish "something is broken" from "something is working exactly as intended, just not converting" isn't really diagnosing anything. It's just measuring variance and handing you the wrong label.

There's no real AI spin, and even when there is one, it won't know you

At the moment, none of this has any actual reasoning layered on top. It's statistics, not intelligence, a credible interval and a flag, not an explanation.

That's changing. Google is integrating Gemini into GA4 to let people ask questions in natural language, and "generated insights" that summarise changes in plain language are already rolling out. That will make GA4 feel a lot smarter. Anomaly AIALM Corp

But it's worth being clear-eyed about what that AI actually knows. It knows Google Analytics, the schema, the metric definitions, the standard patterns. It doesn't know your metrics, custom dimensions, margins, your inventory constraints, your sales team's priorities, or your broader business strategy unless that's represented in the data and interpreted by your team. It's generic GA4 knowledge, applied to your account. It isn't knowledge of your account, your site, your campaigns, your quirks, the context of why last month's numbers looked the way they did. ALM Corp

That distinction matters more than it sounds. A model that knows GA4 well can tell you a metric moved. A model that knows your GA4 well can tell you a metric moved because of the specific campaign you ran on the 14th, that it's the third time this quarter a similar pattern has shown up, and that it isn't actually anything to worry about.

The question worth asking first

This is exactly the gap we built Inspectre Insights to close.

It's Fresh Egg's own monthly analytics briefing, and the difference starts with what it's built on: the same GA4 mastertable data engineering we've already done for every client, sitting clean and structured in BigQuery, not GA4's native, generic schema. It has knowledge on the whole setup, from the documentation already written. It knows the shape of your data, not just the shape of GA4 in general.

Every month, it doesn't just check whether a metric drifted outside a range. It runs real SQL against your actual warehouse, works through a fixed set of checks, sessions and users, channel shifts, conversions and revenue, event volumes, data freshness, and then writes up what it found in plain English, the way a colleague would explain it to you over Slack, not the way a dashboard would flag it. Every number in that write-up is checked back against the query that produced it, so it isn't guessing. And where it's genuinely relevant, it'll connect a spike or a dip to something happening in the news or your sector, clearly labelled as a possible coincidence, never dressed up as a proven cause.

The result reads like analysis, because it is analysis, not a threshold rule that fired.

The takeaway

GA4's anomaly detection is a reasonable first line of defence, and it's fine as far as it goes. But "something moved" isn't the same as "here's what happened and why it matters to you," and no amount of generic AI bolted onto GA4 will close that gap on its own. What closes it is a tool that actually knows your data, your history, and your business — not just Google Analytics in general.

If your team is still relying on GA4 to tell you when something's wrong, it's worth asking how many of those alerts you've actually acted on, and how many false alarms (or missed real ones) it's cost you along the way. That's usually the moment to talk to us about what a proper data foundation and a genuinely tailored monitoring layer could do instead.

Get Fresh Thinking and Stay Ahead of Digital Marketing Trends

Subscribe to our Fresh Thinking Newsletter and get...

✅ Exclusive invites to our helpful knowledge-share events ✅ Curated news roundup of pertinent and topical industry news ✅ Early access to our industry resources and insights

Register for Fresh Thinking and keep up to date with our helpful digital marketing news insights.

Extraction complete. All navigation, ads, and non-content elements have been removed. The content is now ready for CMS import with proper markdown formatting, preserved links, and logical heading hierarchy.