GA4 audiences and predictive metrics enable targeted analysis, remarketing, and re-engagement. Their operational value is contingent on the quality, volume, and consistency of the underlying event data. Predictive outputs should be treated as eligibility-dependent modelled signals, not substitutes for a comprehensive measurement framework.
This approach does not address incomplete tracking, incorrect event definitions, inadequate consent implementation, fragmented user identification, or insufficient conversion governance. Machine learning models cannot compensate for data that fails to represent actual user behaviour.
1. What GA4 predictive metrics measure
GA4 currently provides several predictive metrics based on behavioural event data collected from websites and applications.
The principal metrics are:
- Purchase probability
The probability that a user active within the previous 28 days will trigger a supported purchase key event during the following seven days. - Churn probability
The probability that a user active within the previous seven days will not be active during the following seven days. - Predicted revenue
The revenue expected from purchase key events during the following 28 days for a user active within the previous 28 days. - In-app purchase probability
The probability that a user active within the previous 28 days will trigger an in_app_purchase event within the prediction window.
The definitions and prediction windows are documented in Google’s GA4 predictive metrics documentation.
These metrics are not manually calculated rules. Google Analytics models generate them using behavioural event data from the relevant property. Google states that demographic dimensions available in reporting are not used to train these models. The models are based on interactions with the website or application.
Recommended action: Before planning campaigns that use predictive metrics, document the business objective, the relevant prediction window, and the supported conversion event. Each prediction should serve a defined operational purpose and should not be enabled solely due to its availability.
2. Minimum data requirements
Predictive modelling requires both positive and negative examples. A property must therefore have sufficient evidence of users who performed the relevant behaviour and users who did not.
According to Google’s documentation, the property must have, during at least one seven-day period within the previous 28 days:
- At least 1,000 returning users who triggered the relevant predictive condition, such as a purchase or churn condition.
- At least 1,000 returning users who did not trigger that condition.
These requirements apply to the relevant model. Meeting the data threshold for one predictive metric does not necessarily mean that all predictive metrics will become available.
For purchase probability and predicted revenue, the property must collect supported purchase events. Google recommends the purchase event for ecommerce implementations. For the purchase event, the value and currency parameters are also required for eligibility for purchase-related metrics.
The relevant implementation guidance is available in Google’s documentation for recommended events, GA4 ecommerce events and predictive metrics.
The data requirement is not limited to event volume. The events must also be meaningful and consistently implemented. For example, a purchase event that fires when an order is initiated rather than completed may provide a misleading training signal. Likewise, duplicate purchase events, missing transaction identifiers or inconsistent revenue values can reduce the usefulness of the resulting model.
Recommended action: Validate the entire purchase data path prior to assessing predictive eligibility. This validation should encompass event firing conditions, transaction identifiers, event deduplication, item data, value, currency, consent behaviour, and the correspondence between the data layer and the GA4 request.
3. Model quality must be sustained
A property may meet the initial volume requirements and still lose access to predictive metrics. Google requires model quality to be sustained over time.
If model quality falls below the required threshold, Analytics stops updating the corresponding predictions. The metrics may subsequently become unavailable in the property.
This creates an important operational limitation. Predictive eligibility is not a permanent property setting. It depends on continued data quality, sufficient activity and model performance.
Several factors can affect this continuity:
- A sustained reduction in traffic or returning users.
- Seasonal changes in purchasing behaviour.
- A migration that changes event names or parameters.
- Consent changes that reduce the available behavioural dataset.
- Incorrect ecommerce events or duplicated conversions.
- Changes to the website or application journey.
- A disruption to cross-domain or app-to-web measurement.
- A change in the definition of a key event.
Predictive metrics are also not real-time outputs. Google states that predictions for each eligible model are generated for each active user once per day. This makes them more suitable for audience building, exploration and campaign planning than for immediate operational decisions.
Recommended action: Incorporate predictive eligibility into the analytics monitoring process. Review changes in data volume, purchase events, and consent rates in conjunction with the availability status displayed in the GA4 audience builder.
4. Predictive audiences depend on predictive metrics
A predictive audience is an audience definition containing at least one condition based on a predictive metric.
Examples include:
- Users with purchase probability above a selected percentile.
- Users with high churn probability.
- Users predicted to generate higher revenue during the next 28 days.
- Users likely to make a first purchase within the prediction window.
Google provides suggested templates such as “Likely 7-day purchasers”, “Likely 7-day churning users” and “Predicted 28-day top spenders”. These templates use percentile thresholds rather than fixed business values. For example, a suggested audience may include users above the 90th percentile of purchase probability.
This distinction is important. A percentile identifies users who rank highly within the available modelled population. It does not guarantee that those users will complete a purchase or generate a defined level of revenue.
Predictive audiences are only available when the underlying predictive metrics meet the necessary prerequisites. If the property becomes ineligible and new predictions are no longer generated, predictive audiences stop accumulating new users. Exported audiences in linked advertising products also stop accumulating new users during the period of ineligibility.
The relevant requirements are set out in Google’s documentation for predictive audiences.
Access also depends on permissions. Google states that a user must have the Marketer role or above at property level to create predictive audiences.
5. Key limitations for analysis and activation
Predictive metrics and audiences can provide useful prioritisation, but several limitations should be considered before they are used in decision-making.
5.1 They are not deterministic
A high purchase probability indicates that the model identifies behavioural patterns associated with a future purchase. It does not confirm that a purchase will occur. Campaign performance, price, availability, customer service, seasonality and external conditions can all affect the outcome.
5.2 They are dependent on the property’s data
The model is trained using data from the relevant website or application property. A property with incomplete journeys, weak ecommerce tracking or inconsistent identity signals may produce limited or unreliable outputs.
5.3 Not every user receives a prediction
Google notes that some users will have prediction metrics while others will not have enough data for a prediction to be calculated. Explorations may therefore show separate groups for users with and without prediction metrics.
This means that a predictive metric should not automatically be interpreted as applying to the entire user base.
5.4 Audience membership is not a historical correction
Creating a predictive audience does not repair previously missing data or reconstruct users whose events were not collected. The audience operates from the point at which the definition is created and evaluated. It should not be used as a substitute for historical data reconciliation.
5.5 Export does not remove platform dependencies
A predictive audience exported to Google Ads, Display & Video 360 or Search Ads 360 remains dependent on the GA4 property continuing to generate predictions. An exported audience is therefore not an independent source of predictive intelligence.
Recommended action: Use predictive audiences as an activation layer. Maintain distinct reporting for observed behaviour, confirmed conversions, and modelled propensity to ensure that campaign results are not conflated with predictive estimates.
6. A practical implementation approach
A controlled implementation can be structured into five stages.
1. Define the use case
Identify whether the objective is conversion acquisition, re-engagement, churn reduction, customer value analysis or budget allocation. The use case should determine the metric and audience threshold.
2. Validate the event schema
Confirm that purchase events, key events and supporting parameters are implemented consistently across relevant websites and applications. For ecommerce, this includes transaction_id, value, currency, item data and the correct firing conditions.
3. Review consent and identity behaviour
Assess whether consent controls affect event collection and whether cross-domain or app-to-web journeys are represented consistently. A predictive model cannot use behavioural signals that it does not lawfully collect or that are not technically available.
TagDataTrust’s CMP integration services can support the relationship between consent controls, analytics collection and marketing activation.
4. Check eligibility and model status
Use the predictive section in the GA4 audience builder to review the status of each available prediction. Check eligibility after material implementation changes and during periods of significant traffic or conversion variation.
5. Test activation and governance
If the audience is exported to an advertising platform, confirm that:
- The linked account is correct.
- The audience is receiving users.
- The relevant campaign objective matches the audience definition.
- Audience size is sufficient for the planned activation.
- Reporting distinguishes modelled audiences from confirmed outcomes.
- A fallback audience or campaign approach exists if eligibility is lost.
A formal measurement plan and a documented change process are essential for organisations managing multiple websites, applications, or regional properties. TagDataTrust’s Google Analytics consulting service provides GA4 audits, event schema design, consent implementation, and data quality validation.
7. What this approach does not solve
GA4 predictive metrics do not solve:
- Missing or incorrectly implemented events.
- Inaccurate ecommerce revenue.
- Duplicate conversions.
- Incomplete consent configuration.
- Poor cross-domain measurement.
- Broken user journeys between websites and applications.
- Inconsistent key event definitions.
- Unclear ownership of analytics governance.
- Insufficient traffic or returning-user volume.
- The need for independent business forecasting.
They also do not establish causation. A user identified as likely to purchase may already have been highly engaged, and an advertising intervention may not be the reason for a subsequent conversion.
Use predictive audiences to support a broader measurement framework, not to replace testing, attribution analysis, customer research, or commercial forecasting.
Conclusion
GA4 predictive metrics can provide a useful additional signal when a property has sufficient volume, correctly implemented events and sustained model quality. Predictive audiences can then support remarketing, re-engagement and exploratory analysis focused on likely future behaviour.
The main limitation is that eligibility and output quality depend on the underlying data. Reliable implementation requires accurate event collection, appropriate consent controls, continuous monitoring, and a clear distinction between observed and modelled results.
A systematic approach should begin with data quality and governance, followed by eligibility assessment and controlled activation within a comprehensive measurement framework.
If you’d like to hear more on this topic and set up a 1:1 call with us, please use the contact form. Contact a Consultant
