GA4 predictive audiences: how to build and activate 'likely to purchase' segments

GA4 has a built-in machine learning system that predicts which of your website visitors are likely to purchase in the next 7 days. It also predicts which existing customers are likely to churn. This isn’t some experimental beta. It’s been production-ready since late 2023. And almost nobody uses it.

I check GA4 properties constantly. I’d estimate that fewer than 5% of ecommerce properties I audit have predictive audiences configured. The reason is a combination of strict prerequisites that most properties don’t meet and a general lack of awareness that the feature exists. If your property qualifies, predictive audiences are one of the most valuable things GA4 offers. If it doesn’t qualify, I’ll explain the alternatives.

What GA4 actually predicts

GA4 offers three predictive metrics:

Purchase probability. The likelihood that a user who was active in the last 28 days will make a purchase in the next 7 days. This is the one that gets marketers excited.

Churn probability. The likelihood that a user who was active on your site in the last 7 days will not be active in the next 7 days. Useful for subscription businesses and apps, less relevant for one-time purchase ecommerce.

Predicted revenue. The expected revenue from a user in the next 28 days, based on their purchase history and behavior patterns. This one is newer and requires even more data than the other two.

These predictions update daily. GA4 runs its models overnight and recalculates every user’s probability scores. You can build audiences based on these scores and push them directly to Google Ads.

The prerequisites most properties don’t meet

Here’s where it gets frustrating. GA4’s predictive metrics have strict data thresholds, and Google isn’t particularly transparent about them.

For purchase probability:

• You need at least 1,000 users who triggered the purchase event (returning positive examples) in the last 7 days.
• You also need at least 1,000 users who visited but did NOT purchase (returning negative examples) in the same 7-day window.
• The model needs to maintain “sustained quality” over a period. Google doesn’t define exactly what this means, but in practice, your data needs to be consistent for at least a few weeks.

For churn probability:

• At least 1,000 returning users who were active and then churned.
• At least 1,000 returning users who were active and remained active.
• Same sustained quality requirement.

Let me put this in perspective. To get 1,000 purchasers in 7 days, you need roughly 143 purchases per day. If your average conversion rate is 2%, that means you need about 7,150 daily sessions at minimum. Many ecommerce sites don’t hit this threshold.

This is a hard gate. You can’t work around it with clever configuration. Either you have the data volume or you don’t.

How to check if your property qualifies

Go to your GA4 property. Navigate to Admin > Custom Definitions > Custom Metrics. That’s not actually where you check. (Google’s UI navigation for this is confusing.)

The real check: go to Explore > start a new blank exploration. Add “Purchase probability” as a metric. If GA4 has enough data, the metric will be available and will return values. If it shows as greyed out or returns no data, your property doesn’t qualify yet.

A more reliable method: go to Admin > Audiences > New Audience > Predictive. If predictive conditions are available in the audience builder, you’re good. If the predictive section doesn’t appear, your property hasn’t met the thresholds.

You can also check by going to Advertising > Insights hub. If GA4’s predictive models are running, you’ll see automated insights about purchase probability trends. No insights there means the models aren’t active.

Building predictive audiences step by step

Assuming your property qualifies, here’s the exact setup.

Audience 1: Likely to purchase in 7 days

1. Go to Admin > Audiences > New Audience.
2. Click “Create a custom audience.”
3. Name it something clear like “Likely Purchasers - Next 7 Days.”
4. Click “Add new condition.”
5. In the condition list, scroll down to “Predictive” and select “Purchase probability.”
6. Set the condition to “greater than” the 90th percentile. GA4 gives you percentile options rather than raw probability scores. The 90th percentile means users in the top 10% of purchase likelihood.
7. Set the membership duration. I typically use 7 days since the prediction itself is 7 days forward.

For Google Ads targeting, the 90th percentile works well. It gives you a focused group with genuinely high intent. If you want a broader audience for awareness campaigns, try the 75th percentile. But don’t go below the 50th percentile. At that point, you’re not really targeting likely purchasers anymore.

Audience 2: Likely to churn

1. Same process. New audience, custom audience.
2. Add predictive condition: “Churn probability.”
3. Set to greater than the 80th percentile.
4. Membership duration: 14 days.

This audience is gold for retention campaigns. These are users who are showing signs of disengaging. Hit them with a targeted offer or reminder before they disappear.

Audience 3: High predicted revenue

1. New audience, predictive condition: “Predicted revenue.”
2. Set to greater than the 95th percentile.
3. Membership duration: 14 days.

These are your whales. The users GA4 predicts will spend the most. Use this for VIP targeting, premium product promotion, or exclusion from discount campaigns (they’ll buy at full price anyway).

As I covered in my piece on GA4 audiences nobody uses, most properties have 2-3 audiences configured. Predictive audiences should bring that number to at least 5-6. They’re the audiences with the highest direct impact on ad spend efficiency.

Connecting predictive audiences to Google Ads

Building the audience is only half the job. The real value comes from activation.

Step 1: Link GA4 to Google Ads. Go to Admin > Product Links > Google Ads. If this isn’t already set up, you should have done it months ago. The link is bidirectional. Once established, GA4 audiences automatically sync to Google Ads as remarketing lists.

Step 2: Enable personalized advertising. In the Google Ads link settings, make sure “Enable Personalized Advertising” is toggled on. Without this, audiences sync but can’t be used for targeting.

Step 3: Wait for the audience to populate. New audiences take 24-48 hours to start populating in Google Ads. Predictive audiences can take a bit longer because they depend on the daily model run. Give it 72 hours before you start using them in campaigns.

Step 4: Use them in campaigns. In Google Ads, go to your campaign settings and add the audience. You have two options:

• Observation mode. The audience is layered on top of your existing targeting. You can set bid adjustments for users in the audience (+20%, +50%, etc.) but you still show ads to everyone else too. This is the safer starting point.
• Targeting mode. Only show ads to users in the audience. This limits your reach but concentrates your spend on high-probability converters.

I recommend starting with Observation mode and a +30% bid adjustment. Run it for two weeks and compare conversion rates between the predictive audience segment and the rest. If the predictive audience converts at 2-3x the rate (which it should), gradually increase the bid adjustment.

Step 5: Smart Bidding integration. If you’re using Smart Bidding strategies like Target ROAS or Target CPA, Google Ads can use your predictive audiences as signals automatically. The algorithm factors in whether a user belongs to a “likely to purchase” audience when deciding how much to bid. You don’t need to set manual bid adjustments in this case. Just make sure the audiences exist and are linked. Smart Bidding picks them up.

How accurate are GA4’s predictions?

I’ve tested this on six ecommerce properties over the past year. Here’s what I found.

The purchase probability model performs best for properties with 5,000+ daily sessions and clear purchase funnels. In those cases, users in the 90th+ percentile converted at 4-7x the rate of the general population. That’s significant.

For properties barely meeting the threshold (just over 1,000 purchasers in 7 days), the accuracy drops. The 90th percentile group still converts at 2-3x the average, which is useful but not spectacular. The model has less data to learn from, so its predictions are less precise.

The churn model is harder to evaluate because “churn” is a soft definition. For subscription services with clear churn events, the model works reasonably well. For traditional ecommerce where “churn” just means “didn’t come back,” the predictions are noisier.

My honest assessment: if your property qualifies, predictive audiences are absolutely worth using. The lift is real. But they’re not magic. They’re an incremental improvement on top of a solid measurement foundation, not a replacement for one.

The BigQuery alternative

What if your property doesn’t meet GA4’s thresholds? You have options.

If you’re already exporting GA4 data to BigQuery (and if you’re not, read my piece on BigQuery for marketers for why you should), you can build custom predictive models using BigQuery ML.

BigQuery ML lets you create classification models using SQL. You don’t need Python or a data science degree. A basic purchase propensity model looks like this:

CREATE OR REPLACE MODEL `your_project.analytics.purchase_propensity`
OPTIONS(model_type='logistic_reg') AS
SELECT
  user_pseudo_id,
  COUNT(DISTINCT session_id) as session_count,
  SUM(CASE WHEN event_name = 'view_item' THEN 1 ELSE 0 END) as product_views,
  SUM(CASE WHEN event_name = 'add_to_cart' THEN 1 ELSE 0 END) as add_to_carts,
  MAX(CASE WHEN event_name = 'purchase' THEN 1 ELSE 0 END) as purchased
FROM `your_project.analytics_XXXXXXXX.events_*`
WHERE _TABLE_SUFFIX BETWEEN '20260101' AND '20260630'
GROUP BY user_pseudo_id

This is a simplified example, but it shows the concept. You train a model on historical behavior patterns and use it to score current users. The advantage: no minimum data thresholds. You can build models on whatever data you have. The disadvantage: you need to maintain the model yourself, and connecting the scores back to Google Ads requires more plumbing (typically through Customer Match uploads or a custom API integration).

For properties doing 500-5,000 daily sessions, this is often the practical path. You won’t get GA4’s push-button audience creation, but you get a model tuned to your specific business.

When predictive audiences work well vs. when they don’t

Works well:

• Ecommerce with 5,000+ daily sessions and consistent purchase volume
• Subscription businesses with clear engagement patterns
• Properties with good data quality (proper ecommerce tracking, no duplicate events, correct revenue values)
• Businesses running Google Ads at scale where even small efficiency gains compound

Doesn’t work well:

• Low-traffic properties (under 1,000 daily sessions). You won’t qualify, and even BigQuery models need enough data to find patterns.
• B2B with long sales cycles. GA4’s 7-day prediction window is too short for a B2B buying cycle that takes months.
• Properties with messy data. If your purchase events fire inconsistently or your revenue values are wrong, the model learns from garbage and produces garbage. Fix your ecommerce tracking first.
• Highly seasonal businesses during off-season. The model trains on recent data. If you’re a swimwear brand in December, the model doesn’t have enough recent purchase data to make good predictions.

What to do next

If your property qualifies: build the three audiences I described above. Link them to Google Ads. Start with Observation mode and measure the impact. This is a low-risk, high-potential-reward change that takes maybe 30 minutes to implement.

If your property doesn’t qualify: focus on growing your conversion volume first. Make sure your purchase event tracking is solid. Get your daily purchaser count above 143 per day (1,000 per 7 days). Once you cross that threshold, GA4 will automatically start generating predictive metrics.

If you’re in between: explore the BigQuery ML path. It’s more work upfront, but it gives you predictive capabilities regardless of your traffic volume.

Predictive audiences are one of those GA4 features that separates “we use GA4 because we have to” from “we use GA4 because it actually drives decisions.” The data is already there. The models are already running (or ready to run). You just need to activate them.