GA4 + AI in 2026: Gemini insights, predictive metrics, and what actually works

Google has been stuffing AI into every product it owns for the past two years. GA4 got its share. Gemini-powered insights, natural language queries, predictive metrics, automated recommendations. The marketing was aggressive. The reality is mixed.

I’ve been testing every AI feature in GA4 since they started rolling out in late 2024. Some are genuinely useful. Some are actively misleading. And a few are interesting ideas that Google shipped before they were ready. Here’s my honest breakdown after running these features across dozens of client properties with real traffic and real money on the line.

The AI features Google added to GA4

Let me list what’s actually in the product now, because Google’s documentation makes it hard to tell what’s live versus what’s in beta versus what’s just a blog post announcement.

Natural language queries. You can type questions in plain English in the search bar. “How many users visited from Germany last week” or “What was my conversion rate for mobile in March.” GA4 tries to interpret the question and shows you the relevant report or data.

Gemini insights. The insights panel on the home screen now uses Gemini to generate summaries and explanations. It tells you things like “Your traffic from social increased 34% this week, driven primarily by a spike in Instagram referrals.”

Predictive metrics. Three machine learning metrics: purchase probability, churn probability, and predicted revenue. These have actually been around since 2021, but Google keeps improving the models and adding new ways to use them.

Automated recommendations. GA4 suggests actions: “Create an audience of likely purchasers” or “Set up a Google Ads remarketing campaign for users who abandoned checkout.”

AI-generated report summaries. When you open certain reports, a Gemini summary appears at the top explaining trends in your data.

What actually works

I’ll start with the good news. Three features are worth your attention.

Predictive audiences are the real winner

Predictive audiences are the most useful AI feature in GA4, and it’s not close. Here’s why: they let you target users based on their predicted future behavior, not just what they’ve already done.

GA4 builds these audiences using on-device machine learning models trained on your own data. When your property has enough conversions (roughly 1,000 positive examples and 1,000 negative examples over 28 days), the models activate. You get three predictions:

• Purchase probability: likelihood a user will make a purchase in the next 7 days
• Churn probability: likelihood an active user won’t visit in the next 7 days
• Revenue prediction: expected revenue from a user in the next 28 days

The real value is connecting these to Google Ads. You build an audience of “likely purchasers” and use it as a bid signal or targeting segment. I’ve seen this improve ROAS by 15-30% for e-commerce clients who had enough conversion volume. That’s a real, measurable improvement. I wrote about audience building strategies that most people overlook, and predictive audiences are the biggest missed opportunity in that list.

One important caveat: you need volume. If your site gets fewer than 1,000 purchases per month, the predictive models either won’t activate or won’t be accurate enough to be useful. Google doesn’t publish exact thresholds, but in my experience, anything under 500 monthly conversions produces unreliable predictions.

Anomaly detection catches problems you’d miss

The anomaly detection in GA4 is surprisingly competent. It uses statistical models to identify when a metric deviates significantly from its expected range. I’ve had it flag issues I would have missed for days: a tracking tag that broke after a CMS update, a sudden drop in mobile conversions after a site redesign, a referral spam attack inflating traffic numbers.

The key is to actually check it. GA4 surfaces anomalies in the insights panel and through email alerts if you set them up. Most people ignore both. Set up email notifications for anomalies on your key metrics (sessions, conversions, revenue) and actually read them.

Natural language for quick exploration

Typing “what was my bounce rate on mobile last month” is genuinely faster than navigating to the right report, applying a segment, and setting the date range. For quick, simple questions, the natural language interface saves time.

It works well for straightforward queries. Single metrics, simple filters, standard dimensions. Think of it as a faster way to navigate to the data you already know exists.

What doesn’t work yet

Here’s where I get frustrated.

Gemini insights are generic to the point of being useless

The AI-generated insights on the GA4 home screen sound impressive if you’ve never looked at your data before. “Traffic increased 12% this week.” “Users from the United States are your largest audience.” “Most conversions happen on desktop.”

These are observations, not insights. An insight would be: “Your conversion rate dropped on mobile because the checkout form breaks on iOS 18 Safari.” GA4’s AI doesn’t do that. It describes what happened. It rarely tells you why. And it never tells you what to do about it in a way that reflects your specific business context.

I’ve watched clients get excited about these summaries, only to realize they could have gotten the same information by glancing at the overview report. The AI wrapping doesn’t add analytical value. It adds readability to data that was already accessible.

Automated recommendations are often wrong

GA4’s recommendations engine suggests actions based on your data. The problem is that these recommendations are generic templates, not contextual analysis. I’ve seen it recommend “enable enhanced measurement” to clients who already had it enabled. I’ve seen it suggest creating remarketing audiences for a B2B SaaS company with a 6-month sales cycle where GA4 remarketing audiences would be worthless.

The worst offenders are the Google Ads integration recommendations. They consistently push you toward more spending, broader targeting, and features that benefit Google’s ad revenue. This isn’t conspiracy thinking. It’s just how the incentives work.

”Ask Analytics” struggles with anything complex

Try asking GA4 a compound question. “Show me users who visited more than 3 times, viewed a product page, but didn’t purchase, segmented by acquisition channel.” You’ll get a confused response, a partial answer, or a redirect to a report that doesn’t answer your question.

Natural language works for simple lookups. For the analytical questions that actually matter to your business, you still need to build explorations manually or, better yet, write SQL in BigQuery. There’s a reason I keep telling people to learn basic BigQuery queries. The interface has limits that AI doesn’t fix.

How to use predictive metrics properly

Since predictive audiences are the feature worth investing in, let me walk through the practical setup.

Building a predictive audience

Go to Admin > Audiences > New audience. Under “Predictive” you’ll see the three options. Start with “Likely 7-day purchasers.” Set the probability threshold. Google defaults to the top percentiles. I usually start with the top 10% and test from there.

You can combine predictive metrics with other conditions. “Likely purchasers who came from organic search” or “Likely churners who have made at least one previous purchase.” The combinations are where the value is.

Connecting to Google Ads

Link your GA4 property to Google Ads if you haven’t already. Then enable audience sharing. Your predictive audiences will appear in Google Ads within 24-48 hours. Use them as:

• Bid adjustments in existing campaigns (increase bids for likely purchasers)
• Targeting segments for dedicated campaigns
• Exclusion lists (exclude likely churners from acquisition campaigns)

The bid adjustment approach is safest. You’re not changing your campaign structure, just telling Google to bid more aggressively for users who your own data predicts will convert.

Expected accuracy

Google doesn’t publish accuracy metrics for predictive audiences, which is frustrating. In my testing across e-commerce properties, the “likely purchasers” audience typically converts at 3-5x the rate of the general population. That’s good enough to be useful for bidding, but not accurate enough to make major budget decisions on.

The churn probability model is less reliable, in my experience. It tends to flag users as “likely to churn” who were actually one-time visitors with no intent to return. The model works better for subscription businesses or apps with regular usage patterns than for standard e-commerce.

The real AI opportunity: BigQuery ML

Here’s what I actually get excited about. If you want to do serious machine learning on your analytics data, forget GA4’s built-in features. Export your data to BigQuery and build custom models.

BigQuery ML lets you create machine learning models using SQL. No Python required. No separate ML infrastructure. You write a CREATE MODEL statement, point it at your GA4 export data, and Google trains the model.

I’ve built models that predict:

• Which blog readers will eventually become leads (for B2B clients)
• What product combinations drive the highest lifetime value
• Which traffic sources produce customers who retain vs. churn after 90 days

These models use your specific data, your specific business context, and your specific definition of success. That’s the difference. GA4’s built-in AI uses generic models that treat every business the same way. BigQuery ML lets you build something custom.

The learning curve is real. You need to understand basic ML concepts: training vs. test data, overfitting, feature selection. But if you already know SQL, you’re 70% of the way there.

And when you build these models, the outputs actually become useful in your dashboards and reporting. Instead of showing stakeholders generic GA4 insights, you’re showing them predictions built on your own data.

My honest assessment

GA4’s AI features are a step forward from having nothing. Predictive audiences work. Anomaly detection works. Natural language saves time for simple questions.

But Google is overselling AI in analytics the same way every tech company oversells AI right now. The Gemini insights aren’t insightful. The recommendations aren’t smart. The “ask analytics” feature isn’t a replacement for knowing how to analyze data.

If you’re a marketer who doesn’t have time to learn analytics deeply, the AI features will give you surface-level answers faster. That has value. But if you need actual analysis that drives business decisions, you still need human judgment, proper tracking setup, and probably BigQuery.

The companies I work with that get the most value from their analytics data aren’t the ones using more AI features. They’re the ones with clean data, clear measurement plans, and people who know how to ask the right questions. No amount of AI fixes dirty data or unclear business objectives.

Use predictive audiences. Set up anomaly alerts. Try natural language when you need a quick number. Ignore the rest until Google ships something that actually thinks, not just describes.