Google Analytics 4 ships with more machine learning than most accounts ever switch on: predictive metrics that estimate which visitors will buy or churn, predictive audiences that export straight to Google Ads, a Gemini-powered assistant called Analytics Advisor that answers questions about your data in plain language, automated insights that flag anomalies before you notice them, behavioral modeling that fills the gaps consent banners create, and a new channel that finally separates traffic arriving from AI assistants. This post goes feature by feature: what each one actually does in 2026, the data thresholds that gate the predictions, which claims are marketing rather than capability, and the setup order that turns it all on.
The short version
- GA4 predicts purchase probability, churn probability, and revenue, gated by strict data thresholds.
- Predictive audiences push those forecasts into Google Ads targeting and exclusions.
- Analytics Advisor answers plain-language questions inside GA4; verify its numbers like any analyst's first draft.
- AI-assistant referrals from ChatGPT, Gemini, and Perplexity now get their own traffic channel.
- Everything here is in the free tier. The models are only as good as the event data feeding them.
- 3 predictive metrics built into GA4
- 1,000 converting users needed to train the models
- 7 days prediction window for purchase and churn
- $0 extra cost over standard GA4
The three predictive metrics and their thresholds
Purchase probability estimates the chance that a user active in the last 28 days buys within the next 7. Churn probability estimates the chance that a user active in the last 7 days does not return in the next 7. Predicted revenue estimates the purchase revenue expected from a user over the next 28 days. The models train on your event data plus pattern knowledge from Google's wider network, the same machine learning family behind Smart Bidding. One vocabulary note: what GA4 called conversions are key events since March 2024, so current menus say key events wherever older articles said conversions.
The thresholds are strict and non-negotiable. Within a 28-day window you need at least 1,000 returning users who triggered the relevant key event and 1,000 who did not, and the model has to keep passing quality checks or the metrics switch back off. There is no override and no support ticket that unlocks them. Sites below purchase volume can anchor predictions to a higher-volume key event such as add-to-cart or lead submission, which is the standard workaround for lead-gen and B2B properties. On accuracy, Google publishes no figures, so treat every probability as a ranking of users, not a forecast with error bars: the top slice really does behave differently from the bottom slice, and that difference is what you can act on.
Predictive audiences turn the forecasts into targeting
Once predictive metrics are live, GA4 offers ready-made audience templates, likely 7-day purchasers, likely 7-day churning users, and predicted top spenders, plus custom combinations with thresholds you set. Link Google Ads and the audiences sync automatically. The plays that pay are consistent across accounts. Retarget only high-probability buyers instead of every visitor, so budget stops chasing people who were never going to convert. Use low-probability users as campaign exclusions, the cheapest saving available. Trigger win-back email flows the moment users enter the churn audience, since keeping a customer costs less than reacquiring one. And give predicted top spenders early access and priority treatment while they are still deciding.
Analytics Advisor puts Gemini inside your property
Analytics Advisor is the conversational layer Google built into GA4 on its Gemini models, generally available for English-language properties since late 2025. You ask in plain language, which channels drove the most key events last month, why did traffic dip in March, build me an exploration comparing mobile and desktop buyers, and it answers from your data, explains the report it used, and links you to it. At Google Marketing Live 2026 Google went further and announced Ask Advisor, a unified agent that connects Analytics, Google Ads, and Merchant Center so one conversation can span measurement and media; it is in beta as of this writing.
The honest assessment from daily use: Advisor is genuinely good at retrieval and explanation, the where-do-I-find and what-does-this-mean work that used to eat an analyst's morning, and it lowers the wall GA4's interface built between marketers and their own numbers. It does not know your business, it answers the question you asked rather than the one you should have asked, and like every language-model product it is occasionally confidently wrong. Treat its output like a junior analyst's first draft, useful immediately, verified before it reaches a decision. The older natural language search bar still works too, so quick lookups like key events last month by channel resolve without a conversation, and for repeated questions it stays the faster path.
Generated insights watch the data so you do not have to
GA4's insights engine continuously scans your property and surfaces anomalies, trends, and shifts on the home screen and in the Insights hub: an unusual traffic spike, a revenue dip that breaks the weekly pattern, a channel quietly converting far better than the rest. The underrated half is custom insights: define your own conditions, such as key events falling sharply week over week, and GA4 emails you when they trigger. Five minutes of configuration turns the report you forget to check into an early warning system, and it is the single feature on this page we switch on for every client account without exception.
The AI assistant channel finally shows AI referrals
Visits that arrive from AI assistants used to hide inside referral traffic or vanish into direct. GA4 now groups them into their own channel, so traffic referred by ChatGPT, Gemini, Perplexity, and similar assistants is visible as a segment you can trend, compare, and build audiences from. The volumes are small for most sites and the visitors are unusually far along, because they arrive pre-briefed by the answer that sent them. Watching this channel is also the closest thing GA4 offers to measuring your AI search visibility: if the work of getting cited by these engines is succeeding, this is where it shows. Growing that number is the point of our AI SEO services, and the channel is how you hold that work accountable.
Behavioral modeling fills the consent gap
When visitors decline cookies, GA4 does not just show a hole in the data. With Consent Mode set up, it models the behavior of non-consenting users from the patterns of similar users who did consent, restoring a privacy-compliant estimate of the traffic and key events you would otherwise lose. For sites with meaningful European traffic this one feature changes reported numbers enough to alter decisions, and it costs configuration, not money.
What is real and what is marketing
| Feature | What it does | Gate | Honest verdict |
|---|---|---|---|
| Predictive metrics | Score users on purchase, churn, revenue | 1,000 converters and 1,000 non-converters in 28 days | Real, directional, gated by volume |
| Predictive audiences | Push forecasts into Google Ads | Predictive metrics live, Ads linked | The most actionable feature here |
| Analytics Advisor | Plain-language answers and explorations | English-language properties | Strong retrieval, verify before deciding |
| Generated and custom insights | Anomaly alerts, condition-based emails | None worth noting | Quiet workhorse, switch it on everywhere |
| AI assistant channel | Separates ChatGPT and Gemini referrals | None | Small numbers, high signal |
| Behavioral modeling | Models non-consenting users | Consent Mode in place | Real, decisive for EU-heavy sites |
Three claims deserve deflating. First, a chunk of what gets marketed as new GA4 AI is anomaly detection Analytics has run for years, renamed. Useful, not new. Second, some articles still describe Universal Analytics features, Smart Goals is the usual example, as if they lived in GA4; they never shipped there, and setup time spent hunting for them is wasted. Third, no AI feature repairs bad measurement. Predictions trained on a property with broken event tracking, missing values, or duplicate tags produce confident scores about garbage, and the Advisor will summarize that garbage fluently.
Most GA4 properties we open have a tracking problem before they have an analysis problem: key events that fire twice, purchase events missing revenue values, forms that never report at all. The pattern we keep meeting is a business evaluating AI features on top of data no model could rescue. Audit the events first, then switch on the intelligence. The features reward clean inputs out of all proportion to the setup effort.
How to switch the AI on in the right order
-
Verify the key events that matter
Predictions train on your key events, and for revenue models the purchase event needs transaction_id, value, and currency parameters. Broken ecommerce tracking means no models, whatever your traffic volume.
-
Turn on Google signals and link Google Ads
Signals improves cross-device coverage; the Ads link is what lets predictive audiences flow into campaign targeting and exclusions.
-
Set up Consent Mode
Required for behavioral modeling, and increasingly the default expectation for compliant measurement in any market with consent rules.
-
Check whether predictions are live
Try creating a new audience under Admin. If the predictive templates appear, your property cleared the thresholds; if not, the gap is data volume, not configuration.
-
Create custom insights for the numbers that would fail silently
Alert on key-event drops, traffic collapses, and cost spikes. This is the five-minute setup with the best payoff on the page.
-
Put the AI assistant channel on your monthly report
Trend it beside organic search. It is small, it is growing, and it is the earliest scoreboard for AI search visibility work.
What to do with the predictions once they exist
The features only pay off when they change decisions. Shift ad budget toward high-probability audiences and away from low ones. Intervene on churn before it happens instead of discounting after it does. Use insight anomalies as the trigger for investigating engagement problems, alongside the diagnostics in our post on reducing bounce rate. And when the blocker is setup rather than strategy, Egochi audits the tracking, activates every feature on this page, and builds the reporting as part of our Google Analytics service. The machine learning is only as good as the event data feeding it, and clean events are where every engagement of ours starts.
Questions people ask about GA4 AI features
Are GA4 AI features free?
Yes. Predictive metrics, predictive audiences, generated insights, natural language search, and the Analytics Advisor assistant all ship in the free standard GA4 property. The paid 360 tier changes data limits and SLAs, not the AI. The only real gate is meeting the data thresholds.
Why are my predictive metrics not working?
Almost always insufficient data. You need at least 1,000 returning users who triggered the relevant key event and 1,000 who did not, within 28 days, and the model quality has to hold over time. If you clear the thresholds, allow a week or two for training before the metrics appear.
How accurate are GA4 predictions?
Google does not publish accuracy figures. The models share machine learning infrastructure with Google Ads Smart Bidding and improve as data accumulates. Use predictions directionally, high probability versus low, rather than treating any percentage as precise.
What is Analytics Advisor in GA4?
A Gemini-powered assistant built into the GA4 interface. You ask questions about your data in plain language and it answers, explains reports, and helps build explorations. It became generally available for English-language properties in late 2025. It retrieves and explains well; it does not replace knowing what to ask.
Do GA4 AI features work for B2B websites?
Yes, with adjustments. B2B purchase volumes rarely clear the purchase thresholds, so anchor predictions to a higher-volume key event like lead form submission instead. Churn probability works well for spotting prospects going cold between long sales-cycle touches.
Do I need GA4 360 for predictive audiences?
No. Predictive audiences are available in free GA4 once your property meets the data thresholds. Link Google Ads under Admin, then Product Links, and the audiences sync automatically for campaign targeting and exclusions.
