From Data to Decisions: AI for Marketing Analytics

Ask a marketing team if they have enough data and almost none will say no. GA4 tracks every session, the CRM logs every lead, the ad platforms report cost and conversions hourly. The complaint is never a shortage of numbers. It is that none of it seems to change what happens next: the same campaigns run, the same budgets get renewed, the same segments get emailed, regardless of what last week’s dashboard said.

That gap between data and decision is exactly what AI is built to close in marketing analytics right now, and it closes it in two specific, checkable mechanisms: prediction and activation, both inside tools you likely already pay for.

The ceiling of descriptive analytics

Most marketing reporting is descriptive: it tells you what happened. Harvard Business Review’s framing of advanced analytics is useful here, describing tools that help companies solve marketing, sales, and supply-chain problems by modeling customer behavior and optimizing decisions such as pricing (Harvard Business Review), rather than simply reporting a number after the fact. A dashboard that tells you last month’s churn rate is descriptive. A model that tells you which currently active customers are likely to churn next week is predictive, and it is the difference between reacting and acting.

This matters because a decision needs a forward-looking answer. "Conversion rate was 2.1% last month" does not tell a media buyer who to bid more for tomorrow. "This segment of 4,000 users has a high purchase probability in the next 7 days" does.

GA4 predictive metrics: what they track and what they require

This is not a future capability. It is available in the standard GA4 property once your data clears the eligibility threshold below.

GA4 supports predictive metrics computed by machine learning models trained on your own event data (Google Analytics Help):

• Purchase probability. the probability that a user active in the last 28 days will log a purchase event within the next 7 days.
• Churn probability. the probability that a user active in the last 7 days will not be active in the next 7 days.
• Predicted revenue. the revenue expected from purchase events within the next 28 days from a user active in the last 28 days.

They come with a hard gate, and it is worth knowing it before you go looking for them: your property needs at least 1,000 returning users who triggered the relevant event (a purchase, for instance) and at least 1,000 who did not, within a rolling 28-day window. Below that threshold, GA4 will not generate the metric at all. This is the single most common reason a team assumes "GA4 AI does not work for us": the volume gate, not the model, is what is missing.

Predictive audiences: how a score becomes a campaign action

A probability sitting in a report changes nothing by itself. GA4 turns it into something actionable through predictive audiences, defined as any audience with at least one condition based on a predictive metric (Google Analytics Help), for example the top 10% most likely 7-day purchasers, or users on the edge of churning.

The mechanism that makes this a genuine data-to-decision loop, not just a smarter segment: those predictive audiences are shared automatically with any linked advertising account, including Google Ads, Display & Video 360, and Search Ads 360. A user who crosses into "likely to churn" or "likely to purchase" does not wait for an analyst to notice. As the audience refreshes, the campaign it feeds picks up the change and the bid or the remarketing message adjusts accordingly.

The prediction is not the deliverable. The changed bid, budget, or message is.

None of it works on messy data

Prediction and activation both assume clean, unified inputs, and this is where most marketing teams actually lose the value. Salesforce’s tenth-edition State of Marketing report, based on nearly 4,500 marketers, found that teams who have satisfactorily unified their customer data are 42% more likely to regularly respond to customers than teams unsatisfied with their data, and 60% more likely to use AI agents to help scale their efforts (Salesforce). The same report found that only 58% of marketers have complete access to service data, 56% to sales data, and 51% to commerce data, which is precisely the fragmentation that keeps predictive models starved and audiences small.

The report’s sharper finding is comparative, not absolute: high-performing marketers, the ones getting the best returns on spend, are 2.8 times more likely to use customer data to create relevant experiences and 2.4 times more likely to have unified their data sources than their lower-performing peers. AI does not substitute for that unification. It rewards it, disproportionately.

In practice, this means the GA4-CRM link is not optional infrastructure to get to eventually. It is the precondition for every predictive metric and every audience above to have enough signal to be worth activating.

From reflective to reflexive

The change in how decisions get made matters as much as the tools. Harvard Business Review describes the shift plainly: decision-making in sales and marketing is accelerating, and fast, reflexive action, driven by real-time insights, is increasingly key to relevance and results (Harvard Business Review). The old model was reflective: pull a report, discuss it in a weekly meeting, adjust next month’s plan. The model AI enables is reflexive: the system notices a user’s churn probability crossed a threshold and adjusts the campaign before the weekly meeting happens.

That shift changes what a marketing analyst’s job actually is. Less time spent producing the report, more time spent deciding which thresholds and audiences deserve to trigger an automatic action, and auditing that the automation is still doing the right thing.

Attribution is getting harder, so causation matters more

Google’s own 2026 Marketing Live announcements framed measurement itself as the growth engine for the AI era, arguing that the foundation for navigating complex consumer journeys is understanding, not just tracking (Google). Two products it introduced make the point concretely: Meridian GeoX, for geographic incrementality testing, and Meridian Studio, an enterprise platform for marketing mix modeling. Both exist because attribution, crediting a single touchpoint for a conversion, gets less reliable as journeys fragment across devices, channels, and AI-mediated search. Google also cited an average 14% conversion lift for advertisers using its tag gateway infrastructure, tying better first-party data plumbing directly to measurable outcomes.

The practical implication: pair GA4’s user-level predictions with a channel-level incrementality read, at least for your top budget lines. A prediction says who is likely to convert. An incrementality test says whether your spend caused it. You need both to make a defensible budget decision.

A working path from dashboard to decision

1. Unify before you predict. connect GA4, CRM, and ad platform data so predictive models have enough unified signal, and so audiences built from them are large enough to matter. This is the step the Salesforce data says most teams skip.
2. Check the volume gate. confirm your property clears the 1,000-user threshold on both sides of the outcome (purchased and did not, churned and did not) before assuming GA4’s predictive metrics are unavailable to you.
3. Build one predictive audience, not five. start with the highest-value case, usually churn probability for high-value customers or purchase probability for a specific product line, and get it activated end to end before expanding.
4. Activate, do not just view. link the audience to the ad account it should feed and confirm the campaign is actually bidding or messaging differently against it, not just displaying it as a segment.
5. Validate with incrementality. run a holdout or geo test on the campaign the audience feeds, so you know the lift is real and not just attribution shuffling credit between channels.
6. Close the loop back to CRM. feed outcomes back into the customer record so next month’s prediction is trained on a slightly better dataset than this month’s, which is what makes the system compound instead of plateau.

Where this goes wrong

• Treating a dashboard as the deliverable. a predictive metric nobody acts on is decoration. If a probability score does not change a bid, a message, or a budget, it has not paid for itself.
• Assuming the volume gate does not apply. small or mid-size accounts often sit below the 1,000-user threshold and wrongly conclude GA4’s AI features "do not work here," when the real fix is consolidating tracking or a product line first.
• Skipping the CRM link. audiences built from GA4 alone miss the customer-value context a CRM carries, so a "high purchase probability" user might be a low-margin one you should not chase as hard.
• Confusing attribution with causation. a channel that looks efficient in a last-click report may just be capturing demand another channel created. Incrementality testing exists to catch exactly this.

AI does not make marketing analytics more impressive. It makes the step from data to decision shorter, provided the data underneath is unified enough to predict from and the organization is willing to let a prediction actually change a campaign. Most of the work is not modeling. It is plumbing, and the discipline to activate what the model already tells you.

If you want a second set of eyes on your GA4 setup, your CRM connections, or where your measurement stack is quietly leaking decisions, that is the kind of work we do at Gaveau Strategy.