GA4 Predictive Audiences in 2026: Purchase Probability and Churn in Google Ads
Remarketing to “everyone who visited in the last 30 days” stopped being an edge around the time every beginner learned to do it. The difference between a visitor who will come back and buy and one who landed by accident is real — it is just invisible inside a flat list. GA4 predictive audiences exist to quantify that difference: the model scores purchase probability and churn probability per user, and you get a list you can push into Google Ads.
This guide covers how predictive metrics work in 2026, the thresholds you have to clear before the model turns on at all, how to build segments that beat the default templates, where to plug them in across Search, PMax, Demand Gen and YouTube, and — importantly — which use cases produce measurable lift and which are an expensive way to fool yourself.
What GA4 predictive metrics are and how they are calculated
Google Analytics 4 produces three predictions at the individual user level:
- Purchase probability — the likelihood that a user active in the last 28 days will complete a purchase event within the next 7 days.
- Churn probability — the likelihood that a user active in the last 7 days will not be active in the next 7 days.
- Predicted revenue — expected revenue from a user over the next 28 days.
The model trains on your own behavioural stream: event sequences, visit frequency, session depth, sources, device, purchase history. This is not one universal Google model — it is a model fitted to your property. That is exactly why the data requirements are strict.
The thresholds that decide whether you get a model at all
This is where most implementations end before they start. Google’s training requirements:
| Condition | Requirement (Google’s stated benchmark) |
|---|---|
| Positive examples | at least 1,000 returning users who triggered the target event in 28 days |
| Negative examples | at least 1,000 returning users who did not, over the same window |
| Stability | model quality must hold; if it degrades, the model switches off |
| Events | purchase and/or in_app_purchase collected correctly with value and currency |
The practical implication: predictive audiences are a tool for sites with real traffic. A shop doing 40 orders a month cannot access them, and no configuration trick changes that. If that is you, the working alternative is not to wait for the model but to build segments from explicit behavioural signals — product page views, add-to-cart, repeat visit — and feed the algorithm through micro-conversions.
Where to check availability
In GA4: Admin → Audiences → Create audience → Use predictive data. A green “Ready to use” marker next to a metric means the model is trained. “Not eligible” means it is not — hover over it and GA4 will name the specific condition you are missing. The same metrics also appear under Advertising → Predictive audiences, where the prebuilt templates live.
Templates versus your own segments
GA4 ships templates: “Likely 7-day purchasers”, “Likely 7-day churning purchasers”, “Predicted 28-day top spenders” and a few more. They work, but they are tuned for average e-commerce.
A hand-built segment is almost always sharper, because you can combine the prediction with a condition only you know about. Combinations that reliably produce cleaner lists:
- Top 10% purchase probability + add-to-cart in the last 7 days — a short, extremely warm list; the right place for aggressive remarketing bids.
- Top 20% churn probability + at least one prior purchase — leaving customers who are worth winning back. Run the margin maths first, or discounts will eat the lifetime value you are trying to protect.
- Top 5% predicted revenue + new users — future high spenders, better served with service and onboarding than with a coupon.
- Bottom 50% purchase probability + high page depth — researchers. They do not need “buy now” remarketing; they need comparison content.
Deciding where these segments matter across a quarter is easier if you also read the Google Ads Insights page and its demand forecasts: that tells you when your category peaks, while the predictive audience tells you who to reach first during the peak. The percentile threshold itself is a slider in the audience builder. A practical starting point is the top 10–20%, then check list size. An audience below 1,000 users is useless in Google Ads — it will never reach serving volume.
Pushing the audience into Google Ads
- Link the accounts. GA4: Admin → Product links → Google Ads. You need editor rights in GA4 and admin rights in Google Ads.
- Enable ad personalisation data collection in the data stream settings, or audiences will not export.
- Build the audience with the predictive condition and set a sensible membership duration — the default 30 days makes little sense for a “likely 7-day purchaser”; 7–14 days fits better.
- Wait. Audiences populate from the moment of creation; backfill is partial and not instant. Plan for 24–48 hours before the list appears in Google Ads and several more days before it reaches volume.
- Verify in Google Ads: Tools → Audience manager → Segments. Check the separate size columns for Search, YouTube and Display — they differ, and a list can be large enough for Display while too small for Search.
The serving minimums in 2026 are unchanged: roughly 1,000 active users for Search and YouTube, roughly 100 for the Display network. Below that, nothing activates.
Where to plug it in, by campaign type
| Campaign | Use | Mode |
|---|---|---|
| Search (brand and generic) | Positive bid adjustment on top 10% purchase probability | Observation |
| Performance Max | Audience signal — likely purchasers and predicted top spenders | Signal |
| Demand Gen | Target likely purchasers plus lookalikes built from them | Targeting |
| YouTube | Retention: high churn probability segment | Targeting |
| Display / remarketing | Different creative for “ready” versus “researching” | Targeting |
| All campaigns | Exclusion: bottom 50% purchase probability in expensive formats | Exclusion |
One decision saves more budget than the rest combined: in Search campaigns, add predictive audiences in observation, not targeting. Search already carries expressed intent; narrowing it by a prediction throws away conversions from users the model never got to score — new visitors, infrequent visitors, cross-device visitors. The mechanics of the two modes are covered in audience observation without narrowing reach.
Three use cases that actually pay
1. Bid adjustments in Search
Add “top 10% purchase probability” in observation across your Search campaigns. After two to four weeks, read the segment breakdown: if in-segment CPA is 25% or more below account average, apply a positive adjustment. Start moderate — +15 to +20% — then watch volume. The classic beginner error is jumping to +50% and then wondering why blended CPA rose. Note also that with tCPA and tROAS, audience bid adjustments are ignored entirely; the algorithm consumes the signal itself. If you want to explicitly pay more for a predictive segment under Smart Bidding, split it into its own campaign with its own target instead.
2. Retention via churn probability
Audience: top 20% churn probability AND at least one purchase in 180 days. Run it on YouTube or Demand Gen, with a message built around a reason to return — a new range, a service upgrade, a personalised selection — rather than a blanket discount. Measure it with incrementality only; otherwise you will happily claim credit for people who were coming back anyway. How to run that measurement is in the guide to incrementality and geo-experiments.
3. Saving money through exclusions
The most underrated use case. Exclude “bottom 40% purchase probability” from your expensive remarketing formats — YouTube and Demand Gen. You lose almost nothing in conversions and stop paying for impressions served to people the model confidently classifies as non-converting. At scale this frees a double-digit percentage of remarketing budget that can move up the funnel.
Verifying that the model actually discriminates
A predictive audience is a hypothesis, not a fact. Before you build budget around it, spend an evening validating it.
The retrospective check
Create the “top 10% purchase probability” audience, wait 14 days, then compare that segment’s actual conversion rate in GA4 against all users over the same window. A working model produces at least a 2–3× gap. If the difference is 20–30%, the model barely discriminates on your property, and building bid management on it is premature.
The “is this just returning visitors?” check
The most common illusion is that the top percentile is simply people who would have returned anyway. Test it: build a control segment of “users with 3+ sessions in 7 days” and compare its conversion rate to the predictive segment. If they are close, the model is adding nothing beyond a visit counter, and a plain rule would do the same job for free.
The volume check
A list of 1,200 users technically clears the Search minimum, but inside a campaign serving 3,000 impressions a day it will take a month to produce readable statistics. Before building adjustments, estimate how many clicks the segment can realistically deliver. Below roughly 30–50 conversions in the evaluation window, every conclusion is noise.
Keep the segment breakdown visible in your regular reporting rather than digging into GA4 each time; how to build that without manual exports is covered in custom columns and the Report Editor.
Limitations that rarely get mentioned
- Tracking changes break the model. Migrating to a server container, renaming events, restructuring the
purchasepayload — training restarts. Schedule that work deliberately and out of season; the migration itself is covered in server-side tagging with sGTM. - Consent Mode shrinks the training set. Users without consent never enter the model. The stricter your region and banner, the less data and the weaker the prediction. Signal-preservation mechanics are in Consent Mode v2 and privacy.
- It predicts behaviour, not means. The model sees clicks, not bank balances. In long offline-cycle verticals — property, automotive, complex B2B — it is weaker because the decisive events happen off-site.
- It is not portable. The model belongs to a GA4 property. Change domains or properties and training starts over.
- Percentiles float. “Top 10%” is a relative threshold. As traffic grows, membership shifts even if you changed nothing — so check list size weekly rather than setting it and forgetting it.
A one-week implementation checklist
- Day 1. Confirm model eligibility in GA4 and verify the
purchaseevent fires with a correctvalue. If the event is broken, fix that first — everything downstream is meaningless otherwise. - Day 1. Confirm the GA4 ↔ Google Ads link and that ad personalisation data collection is on.
- Day 2. Build three audiences: likely purchasers (top 10%), high churn risk among purchasers, and bottom 40% purchase probability for exclusions.
- Day 3–4. Let them populate; check sizes per network in Audience Manager.
- Day 5. Deploy: observation in Search, signal in PMax, targeting in Demand Gen, exclusion in expensive formats.
- Day 5. Record baseline metrics before the change — without them you cannot prove effect a month later.
- Week 3–4. Read the segment breakdown and decide on adjustments or campaign splits.
Factor in conversion lag: if your decision cycle is ten days, reading results after one week tells you nothing. The correction logic is in conversion windows and conversion lag.
A predictive audience does not create demand — it redistributes your existing budget inside it. The largest measurable gains usually come not from targeting but from exclusions and bid adjustments, where you save money without giving up reach.
How GA4 predictive audiences fit the rest of your data stack
Predictive audiences sit on top of signal quality; they do not substitute for it. If conversions are badly configured, the model trains on noise and confidently predicts nonsense. The order of priorities is: correct conversion and value collection first, then enhanced conversions and first-party data, then Customer Match, and only then predictive modelling. To find where the gap actually is, run the Google Ads account audit checklist — and separately confirm you are not counting conversions twice, which is covered in duplicate conversions and deduplication.
If the bottleneck is the ad account rather than analytics — serving limits, no account history, verification stuck — predictive audiences will not solve it. PPC Rebels provides agency Google Ads accounts for high-volume advertisers for exactly that situation.
Related reading: Customer type labeling and lifecycle goals: Google now classifies your lists
FAQ: GA4 predictive audiences
How much data do predictive audiences need?
Google’s benchmark is at least 1,000 returning users who triggered the target event and 1,000 who did not, within the last 28 days, plus sustained model quality. Low-traffic properties simply do not get the metrics, and no setting works around it.
How is purchase probability different from an “added to cart” list?
Add-to-cart is a single recorded action. Purchase probability is a model score across dozens of behavioural features including visit frequency, depth and history. In practice the intersection of the two outperforms either alone.
I created the audience but it does not appear in Google Ads. Why?
Three usual causes: ad personalisation data collection is off in the GA4 stream, the GA4–Google Ads link is missing, or the list has not reached volume yet. Audiences populate forward from creation, so day-one size can legitimately be zero.
Can I use predictive audiences in Search campaigns?
Yes, but in observation mode. Targeting a predictive segment in Search cuts out high-intent users the model has not scored. Observation gives you the breakdown and the option to adjust without losing reach.
Do bid adjustments work with automated bidding?
With tCPA and tROAS, audience bid adjustments are ignored — the algorithm factors the signal in itself. If you want to explicitly pay more for a predictive segment, split it into a separate campaign with its own target rather than trying to steer it with an adjustment.
How often is a user’s prediction updated?
Scores are recalculated regularly as new events arrive, and audience membership refreshes within about a day. Do not expect real-time reaction to an action taken minutes ago — that requires event-based audiences, not predictive ones.
What do I do if the model switches off?
Google disables a model when prediction quality falls below threshold, usually because of a traffic drop, a seasonal trough, or broken event collection. Check purchase event integrity and returning-user volume over 28 days first. Recovery is automatic once conditions are met again.
Does Consent Mode affect predictive audiences?
Yes. Users who decline consent enter neither the training set nor the remarketing lists. In strict-consent regions predictive audiences can be several times smaller than expected — that is normal behaviour, not a misconfiguration.
Should I build lookalikes from a predictive audience?
Yes — it is one of the strongest plays for Demand Gen and YouTube. Similar audiences seeded from “likely purchasers” are usually cleaner than those seeded from all purchasers, because the source list is more behaviourally homogeneous.
Do predictive audiences work for B2B and lead generation?
Only partially. The model trains on a purchase event; in lead generation the target becomes a form fill whose quality varies enormously. If you never feed lead quality back into GA4, the model learns to predict form fills, not deals. Offline results import comes first.
Do predictive audiences replace Customer Match?
No — they are different layers. Customer Match works from your known contact base; predictive audiences work from anonymous on-site behaviour. They are strongest together: Customer Match for existing customers, predictions for unknown traffic.
How long before I can judge the results?
Three to four weeks minimum, longer with a long sales cycle. Earlier readings are unreliable: lists take days to fill and conversions arrive with lag. Record your baseline before launch, or you will have nothing to compare against.
