The Complete GA4 + AI Setup Guide for Marketers

GA4 and AI now intersect twice. Traffic from AI assistants is arriving at your site and needs to be measured honestly, and GA4 itself ships machine learning - predictive metrics, anomaly detection, automated insights - that most marketers have never switched on or looked at. This guide is the one-pass setup for both directions: what to configure, in what order, entirely in the GA4 interface, and what each piece can and cannot tell you once it is running.

I wrote it because the two topics are always covered separately, and in practice they are one afternoon of work on the same property. Everything here is doable by a marketer with Editor access. Nothing needs a developer, though I will flag the one place where asking for ten minutes of engineering help pays off.

Step 1: get key events right, because every AI feature reads them

All of GA4’s intelligence is downstream of what you tell it success looks like. Key events (the artist formerly known as conversions) feed the predictive models, the anomaly alerts and every channel comparison you will run later, so I start any GA4 work by auditing them: one key event per genuine business outcome - a purchase, a signup, a qualified lead - and nothing decorative. A property where “scrolled 90%” is a key event has taught its own AI that scrolling is revenue, and every automated insight downstream inherits the confusion.

For commerce specifically, Google’s predictive metrics prerequisites are explicit: the models need purchase events carrying value and currency parameters. If your purchase event is missing those, this is the ten-minute developer request worth making, in one sentence: “our GA4 purchase event needs value and currency parameters populated.”

Step 2: know the new AI Assistant channel, and what it still misses

GA4 now classifies assistant traffic out of the box. Google’s default channel group documentation defines an AI Assistant channel - “the channel by which users arrive at your site from sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok” - so visits that used to scatter through the Referral haystack now collect into one reportable line with no setup at all. When I first built AI traffic tracking in GA4, a custom channel group was the only way to get this line. Now the default reports have it, and the custom group is the power-user layer on top.

Two classification decisions in Google’s rules are worth knowing before you read the number. First, Google’s own AI surfaces are excluded: the same documentation files clicks from AI Overviews and AI Mode under Organic Search, which means “our AI traffic” in GA4 is assistant traffic minus Google’s, and your Organic Search line now silently contains an AI component you cannot separate. Second, the mechanics haven’t changed for apps that strip referrers - those sessions still land in Direct. The channel is a floor, not a census, and I say so in every report that includes it.

What I still use the custom channel group for: splitting the assistants from each other. The default channel tells you AI traffic went up. The per-source split tells you Perplexity found a page ChatGPT ignores, which is the actionable version.

Step 3: switch on the predictive metrics, and check whether you qualify

GA4’s headline AI is predictive metrics, three per property, with Google’s definitions worth quoting exactly. Purchase probability: “the probability that a user who was active in the last 28 days will log a specific key event within the next 7 days.” Churn probability: “the probability that a user who was active on your app or site within the last 7 days will not be active within the next 7 days.” Predicted revenue: “the revenue expected from all purchase key events within the next 28 days from a user who was active in the last 28 days.”

The catch is the eligibility bar, and it filters out more sites than the feature pages suggest: in the last 28 days, at least 1,000 returning users must have triggered the predictive condition over a seven-day period, and at least 1,000 returning users must not have, and the model must hold its quality over time to stay eligible. A niche B2B site with 80 purchases a month will simply never see these metrics populate, and that is fine - the models need volume the same way Smart Bidding needs conversion volume, and pretending otherwise produces predictions nobody should act on.

If you qualify, the useful surface is not the metric - it is the audience built from it. Suggested audiences like “likely 7-day purchasers” and “likely 7-day churning users” can be created in a few clicks under Audiences, exported to Google Ads for bidding, or watched as their own rows in reports. My favorite unglamorous use: the churn-probability audience as an early-warning list for lifecycle email, which turns a black-box prediction into a concrete weekly send.

Step 4: point anomaly detection at the metrics you would lose sleep over

GA4’s Analytics Intelligence watches your time series and flags outliers automatically - it fits a Bayesian state-space model to your history and flags data points that fall outside the credible interval it predicted. The training windows tell you what it can catch: two weeks of history for hourly anomalies, 90 days for daily, 32 weeks for weekly. Translation for the practitioner: it is good at “traffic cratered yesterday” and structurally incapable of “this quarter is quietly 8% soft,” because a slow drift never leaves the credible interval.

The automatic insights are fine, but the setup step people skip is custom insights: your own conditions (“daily key events drop 20%”, “AI Assistant channel sessions spike”) evaluated on your schedule, with email alerts. I set five or six per property, and at least one always watches the key event itself rather than traffic - a tag that breaks on the checkout page produces a serene traffic graph and a silent revenue hole, and the anomaly system only guards the metrics you told it to watch.

One rule keeps this feature honest: an anomaly flag is a question, not an answer. The model knows your Tuesday looked wrong. It has no idea a newsletter went out, a page went viral or a tracking tag died, and the correlation-versus-causation discipline that runs through all my measurement writing applies with extra force to alerts generated by a system optimized to find something noteworthy.

Step 5: decide how you will read the AI traffic before it grows

With the channel reporting cleanly and the alerts armed, the last setup decision is interpretive: what does a visit from an assistant mean in your reporting? The external evidence says these visitors are unusually valuable - Semrush’s study of AI search traffic put the average AI search visitor at 4.4 times the value of a traditional organic visit, measured on conversion rate - though the study does not disclose its sample size, so I treat the direction as solid and the multiple as somebody else’s number until my own data speaks. The mechanism is believable: a person who clicked out of an answer has already been pre-sold by the summary.

So the monthly read I run: AI Assistant sessions and trend, engagement and key-event rate against the site average, and landing pages by source. And the caveat that goes in the same report every month: clicks are the visible minority of AI influence. The assistant that recommends you without a link shows up in Direct or branded search or nowhere, which is the zero-click attribution problem, and no GA4 configuration fixes it. GA4 tells you what the trackable slice did. The aggregate instruments tell you whether the whole is moving.

What I deliberately leave off

Two things stay out of my GA4-AI setup. I do not let any AI - GA4’s insights or an external model - write numbers into reports. Generated insights are leads for a human to check, and the numbers-in, narrative-out rule from the measurement hub governs everything that reaches a stakeholder. And I do not chase attribution-model tinkering as an AI strategy - the models that remain all share the same blindness to AI-mediated influence, and the honest response is the aggregate measures above, not a more elaborate division of the credit GA4 can still see.

FAQ

Is any of this a paid feature? No. The AI Assistant channel, predictive metrics, anomaly detection and custom insights all ship in free GA4. The gates are data volume (for predictions) and setup attention (for everything else), not license tier.

My predictive metrics say “not eligible.” Can I fix that? Only by having more converting users - the 1,000/1,000 returning-user thresholds are hard. What sometimes works: pointing the purchase prediction at a higher-volume key event that still means something, like a trial start instead of a paid conversion. What never works is wishing, and a prediction trained on too little data would be worse than none.

Should I still build the custom AI channel group now that the default channel exists? Yes, if you want per-assistant detail - the walkthrough is in the AI referral tracking guide. The default channel answers “how much.” The custom group answers “from whom,” which is what feeds the citation work.

Does GA4 count AI crawlers reading my site? No. Bots fetching pages to train on or answer from do not execute your analytics tag. GA4 sees only humans who click through, which is one more reason the channel understates AI’s real footprint on your business.

This guide is part of the analytics and ROI hub. The per-assistant tracking layer is the AI referral traffic setup, the influence GA4 cannot see is covered in zero-click attribution, the proof framework that survives it is proving AI ROI without attribution, and the hygiene layer under all of it - naming, duplicates, sync - is the data cleaning audit. The wider practice starts at the complete guide to AI marketing.