GA4 Predictive Metrics (Purchase Probability, Churn Probability)
Google's Models Running Inside Your Property
GA4 automatically trains machine learning models directly on your property's own data (no setup required beyond meeting eligibility thresholds) to produce three predictive metrics: Purchase probability (likelihood a user active in the last 28 days will make a purchase in the next 7 days), Churn probability (likelihood an active user will not be active in the next 7 days), and Predicted revenue (expected revenue from purchases in the next 28 days for active users). These power Predictive Audiences directly inside GA4.
Eligibility Requirements Are Stricter Than People Expect
Predictive metrics only populate if your property has enough positive and negative examples of the predicted action within the model's training window: for purchase probability and predicted revenue, GA4 requires at least 1,000 returning users who triggered purchase and 1,000 who did not, within the last 28 days. Churn probability has its own equivalent threshold. Smaller properties, or properties where purchase events aren't firing reliably, simply won't have these metrics available — there's no override.
| Metric | Predicts | Window | Key Requirement |
|---|---|---|---|
| Purchase probability | Will purchase in next 7 days | Based on last 28 days activity | ≥1,000 purchasers + ≥1,000 non-purchasers in 28 days |
| Churn probability | Will NOT be active in next 7 days | Based on last 7 days activity | Sufficient active/inactive user volume |
| Predicted revenue | Expected revenue in next 28 days | Based on last 28 days activity | Same threshold as purchase probability |
Predictive Audiences Are Immediately Actionable
-- Predictive metrics ARE NOT exported to the standard BigQuery events export.
-- They live only inside GA4's UI-based audiences and Google Ads integration.
-- You can approximate your OWN propensity model using BigQuery ML on the
-- raw event export instead, if you need the scores outside GA4's walled garden:
CREATE OR REPLACE MODEL `project.dataset.purchase_propensity_model`
OPTIONS(model_type='LOGISTIC_REG', input_label_cols=['purchased']) AS
SELECT
user_pseudo_id,
COUNT(DISTINCT CASE WHEN event_name = 'add_to_cart' THEN event_timestamp END) AS cart_adds,
COUNT(DISTINCT CASE WHEN event_name = 'view_item' THEN event_timestamp END) AS item_views,
MAX(CASE WHEN event_name = 'purchase' THEN 1 ELSE 0 END) AS purchased
FROM `project.analytics_123456789.events_*`
WHERE _TABLE_SUFFIX BETWEEN '20260801' AND '20260930'
GROUP BY user_pseudo_idWhy You Might Build Your Own Model Anyway
GA4's predictive metrics are a black box — you can't inspect feature weights, adjust the training window precisely, or export raw probability scores for arbitrary external use beyond audience membership and Google Ads sync. Teams with a data science function and BigQuery export access frequently build parallel propensity models (as shown above) when they need explainability, custom feature engineering, or activation into non-Google channels like a CRM or email platform.
Predictive Metrics Can Silently Disappear
What's Next?
Predictive metrics depend entirely on having enough raw event history and sufficiently long data retention to train against. Understanding GA4's data retention settings — and how they constrain both predictive models and historical reporting — is a natural next step.
I build these systems professionally.
Whether it's a RAG pipeline, analytics migration, or AI workflow — let's talk.