GA4 Predictive Metrics and AI Insights

GA4 predictive metrics are scores that Google Analytics 4 calculates with machine learning to estimate what each user will do next, and there are three of them: purchase probability, churn probability and predicted revenue. They appear only on properties with enough purchase data, and they power predictive audiences for ads and remarketing.

• Purchase probability: The chance that a recently active user buys within the next 7 days.
• Churn probability: The chance that a recently active user is not active in the next 7 days.
• Predicted revenue: The revenue a recently active user is expected to bring over the next 28 days.
• Predictive audiences: Ready-made audiences built from these scores, such as likely 7-day purchasers.
• Insights: A separate feature that flags unusual changes in data that has already been collected.

This lesson follows one example: a D2C skincare brand in Bengaluru that sells sunscreen, serums and face wash on its own site. Many shoppers buy once and never return, even though sunscreen and face wash run out every few weeks. The team wants to spend its remarketing budget on people who are likely to buy again, not on everyone.

Key Characteristics of GA4 Predictive Metrics

• Built in: There is no model for you to build, because GA4 trains one on the property's own data.
• Per user: Each score belongs to one user, so you can group users by how likely they are to act.
• Short horizon: The predictions look 7 or 28 days ahead, which suits quick remarketing more than long-term planning.
• Eligibility rules: A property must have enough recent buyers and enough recent non-buyers, and the model must stay accurate.
• Commerce focus: The metrics depend on the purchase event (or in-app purchase for apps), so a site without purchases cannot use them.

How GA4 Predictive Metrics Work

1. Collect events: The site sends events such as session_start, view_item, add_to_cart and purchase, set up as in GA4 events and conversions.
2. Check eligibility: GA4 checks whether there are enough recent users who purchased and enough who did not, because a model cannot learn the difference from only one group.
3. Train the model: GA4 looks for patterns in behaviour that came before past purchases and before users went quiet, such as repeat visits to a product page.
4. Score users: Each recently active user gets a purchase probability, a churn probability and a predicted revenue figure, which update as new data arrives.
5. Build audiences: In the audience builder, you create predictive audiences from suggested templates or your own score ranges.
6. Use and measure: Shared audiences can be used in linked Google Ads accounts, and results are measured in GA4 as usual.

Example: A Bengaluru Skincare Brand

• Starting point: The brand already tracks purchases correctly, following the GA4 tutorial. After a busy summer season, the Admin screen shows that predictive metrics are available.
• Audience one: "Likely 7-day purchasers" who have not bought in the last 30 days. The team shows them a sunscreen restock ad on Google, because many of them bought a bottle a few months ago.
• Audience two: "Likely 7-day churning purchasers". These past buyers are drifting away, so the team sends them a skincare routine guide by email, only to customers who opted in.
• Audience three: "Predicted 28-day top spenders". The team offers early access to a new serum rather than a discount, to protect margin.
• Test: For each audience, a small holdout group receives no ad or message at all, and after four weeks the team compares sales between the groups before deciding to continue.

Benefits of GA4 Predictive Metrics

• Smarter spend: Remarketing budget goes to people likely to act, not to every past visitor.
• No data science team: A small brand gets a working prediction model without hiring analysts or writing any code.
• Early warning: Churn scores show which buyers are drifting before they are fully lost, which supports lifecycle marketing.
• Easy activation: Audiences flow into linked Google Ads accounts without exports, uploads or customer lists.

Limitations of GA4 Predictive Metrics

• Many sites never qualify: Small stores often do not have enough buyers to meet the eligibility rules.
• A black box: GA4 does not show which signals drove a score, so it is hard to explain to a manager.
• Depends on clean data: Duplicate purchase events or missing tags teach the model the wrong patterns.
• Consent gaps: Users who decline analytics cookies are measured less completely, which weakens the model. See Consent Mode v2.
• Prediction is not cause: A high purchase probability does not mean an ad caused the sale, and only a holdout test can show whether it did.

How AI Changes GA4 Predictive Metrics

What AI Automates Now

GA4 scores users and suggests predictive audiences automatically. Its insights feature spots unusual changes, such as a sudden fall in purchases from one city, and some properties can ask questions about their data in plain language.

What Still Needs a Human

People decide what to do with a score, whether that is a reminder, a new product, or nothing at all. They also design the holdout test and judge whether a result makes business sense. The model does not know that a sunscreen shortage caused a drop in sales.

Risk to Watch

Treating a score as fact is the main risk. Predictions can drift when the season, prices or tracking change, and the score will not warn you when that happens. Read the basics of how models can be confidently wrong in AI hallucinations explained, and apply the same doubt to any automated insight. For a wider method, see how to analyze marketing data with AI.

Do It with AI

Use this prompt to plan what to do with each predictive audience. It works in ChatGPT, Claude or Gemini.

1. Confirm in GA4 that predictive metrics are available and note which audiences exist.
2. Run the prompt with your real products and channels.
3. Build each audience with a holdout group and launch one at a time.
4. Compare sales between exposed and holdout groups before scaling any plan.

Check Before You Use It

• Facts: Check audience names and definitions in GA4 itself, not in the AI's answer.
• Brand fit: Offers and messages should match how the brand talks and what it can afford.
• Compliance: Email only customers who opted in, and follow Google's policies on sensitive audience categories.