GA4 has an AI assistant now. It's called Ask Advisor, it runs on Gemini, and if your property is eligible it sits in the top-right corner of the interface. Ask it a question in plain English and it answers, with a number, a chart, and a link to the report it pulled from. Next to it, GA4 still runs the older Analytics Intelligence: the Insights cards that flag anomalies on their own, and predictive metrics that score who's likely to buy or churn.
So the question stops being "does GA4 have AI" and turns into "is it enough." Short answer: it spots what moved fast, and it stops one step before what to do about it. Here's what it does well, and where that last step lives.
What GA4's AI actually is in 2026
The "AI" label in GA4 covers three different features, and they don't do the same job.
Analytics Intelligence is the oldest. It's the Insights panel on the Home screen and inside reports, with short auto-written cards like "Conversions from Paid Search are down 18% week over week." Some arrive on their own. Others you set yourself: tell GA4 to alert you when revenue drops more than 15%, and it watches for you. Underneath it's anomaly detection: GA4 models the expected range for a metric, flags when reality falls outside it, and surfaces that on its own, with nothing to prompt.
Then there's Ask Advisor, the new one. Google shipped it through 2026 as a Gemini chat inside GA4. Ask in plain English and it'll recap last week, pull a specific cut ("sessions by channel for April"), take a "why" ("why did purchases drop on Tuesday"), walk you through a how-to, look up your config, or point at where you're losing the most checkout traffic. It answers in words, draws the chart, and links the report. As of mid-2026 it's still in beta, English only, and gated by eligibility, so not every property has it.
The quiet third piece is predictive metrics, and it's the one that does something a chat can't. GA4 scores each user on purchase probability, churn probability, and predicted revenue. Drop those into the audience builder and you get a "likely 7-day purchasers" segment that pipes straight into Google Ads. So this one doesn't stop at reporting. It pushes a ready-made audience into your ad accounts. The price is a high bar to switch on.
How to get real work out of it
I treat it as a fast first read, not the whole job. The loop that holds up:
Read the Insights cards on Home before you touch anything. They're the cheapest signal in the product. GA4 already ran the anomaly scan, so let it point you. A card saying checkout conversions dropped is your thread for the day.
Pull on it with Ask Advisor. "Why did checkout conversions drop last week?" gets you where the loss sits, usually a device, a channel, or a date. Stay in the same chat and narrow it: "split that by device," "compare it to the prior four weeks." It keeps context, so you're refining instead of rebuilding a report each time. This is the part that used to be twenty minutes of clicking through Explorations.
Set two or three custom insights so you're not running on memory. Revenue down more than X. A key event flatlining. New users from your top channel falling off. GA4 emails you when one trips. It's the closest thing to a smoke alarm the product ships, and most people never turn it on.
And if you run a store with the volume for it, switch on predictive audiences and send "likely churners" somewhere useful, like a win-back email or a Search Ads exclusion. The catch is the volume itself. GA4 won't train the model until you have the volume for it: roughly a thousand recent converters and a thousand who didn't. Low-traffic stores never clear that bar, and there's no trick around it. The model needs the data.
Two things trip people up here. Custom insights only check once a day, so they're a next-morning alarm, not a live pager. And the assistant will summarize whatever GA4 hands it, sampled or thresholded data included, without flagging that it did, so a confident "traffic from Latvia is down 40%" can be three sessions of noise. Sanity-check the small segments before you move budget on them.
Where it stops
Every one of these features runs into the same line.
Ask Advisor will say mobile checkout conversion dropped 12%, and that most of it is on iOS. Correct, useful, and the end of what it can do. It won't tell you that, say, the shipping step is the leak, the address form is rejecting a valid postcode on Safari, and the fix is to move express pay above the fold. That's the gap between an observation and a decision, and the decision is the part you're paid for.
The reason is scope, not brains. Ask Advisor reads one property, and only what lives inside GA4. It can't open your checkout. It can't see that your iOS render runs four seconds slow. It can't read the support ticket where two customers flagged the same broken field. GA4 holds numbers, not causes, and a chat sitting on top of GA4 inherits that ceiling.
A few smaller limits are worth knowing before you lean on it. It works one property at a time, so cross-property and cross-account questions are out. Eligibility is uneven, with the beta gated, English only, and the rollout staggered, so checking whether you have it beats assuming you do. And the answer sits in the wrong place. You work in Claude, or a doc, or a Slack thread, but the number is behind a GA4 login, waiting for you to go fetch it and carry it back.
None of this makes GA4's AI bad. It's a fast, free read of your own numbers, and on most mornings that read is what you want. The decision it can't make is a different tool's job.
Getting to a decision
Closing that gap takes two things GA4's AI doesn't have: the rest of your context, and a layer that turns a number into a call.
The first is plumbing. MCP (Model Context Protocol) connects GA4 to the assistant you already use, Claude or ChatGPT, so the analysis happens where the rest of your work is. You stop switching into GA4 to ask. You ask in the same thread where you're writing the email, reading the tickets, drafting the test. The same GA4 data, read in the tool you already use, with the model you prefer. One detail under it matters: a good server answers from pre-aggregated fact tables, not live LLM guesses over raw rows, so "revenue by channel last month" is computed the same way every time instead of whatever the model pattern-matches off the rows in front of it.
The second is the diagnostic layer, which is why we built ConvRadar rather than wrap the GA4 API and call it done. Full disclosure: this site is ours. On top of the read, it runs the work GA4 won't: funnel-drop diagnosis that names the step that bleeds and how badly, traffic-quality scoring, anomaly detection across every dimension at once, and a hypothesis library that turns "checkout's down on iOS" into a ranked set of things to test. Ask "why did my conversion rate drop, and what should I test first?" and you get a diagnosis with a next action, not a chart you're left to read. When the trail ends at a page, it pulls the page and reads it, so the recommendation is about your actual checkout, not a textbook one.
Be fair about the trade, though. GA4's AI is free, it's already on, and predictive audiences feed Google Ads directly. ConvRadar only reads GA4, so it doesn't push audiences anywhere, and it's free in the open beta, then $9.99/month. If all you want is a weekly anomaly read inside Google's own stack, Ask Advisor may be the whole answer. The case for connecting GA4 to your own assistant starts the moment you want the decision rather than the number, and want it where you actually work. Setup runs about five minutes with no terminal, in Claude or ChatGPT.
Ask Advisor vs GA4 in your own assistant
| GA4 Ask Advisor | GA4 in Claude / ChatGPT (ConvRadar) | |
|---|---|---|
| Where it runs | Inside the GA4 UI | The assistant you already use |
| Setup | None, if eligible | ~5 min, browser sign-in |
| Data scope | One property, GA4 only | Your property, plus the rest of your chat |
| Numbers | Gemini-generated | Read from fact tables, repeatable |
| Output | Answer, chart, report link | Diagnosis with a ranked next action |
| Sees your actual page | No | Yes, fetches and reads it |
| Model | Gemini only | Claude or ChatGPT |
| Predictive audiences to Google Ads | Yes, built in | No |
| Cost | Free | Free in beta, then $9.99/mo |
| Availability | Beta, English, staggered | Any property, today |
The row that decides it is output. The rest is preference.
Which one you actually need
Use Ask Advisor if you live inside GA4, you want a free weekly read of your own numbers, and you're on Google's ad stack where predictive audiences pay for themselves. It's good at that, and it costs nothing.
Connect GA4 to your assistant when the answer you need is the next action, when you want it in the tool where you already work, or when "why" routinely means opening the page rather than reading the chart. The GA4 MCP server overview lays out every route, from Google's own server to the hosted ones to the self-hosted path, if you'd rather not take our word for which fits.
Plenty of people run both. Ask Advisor for the morning glance, the connected assistant when a number turns into a question worth chasing. They hand off more than they compete.
FAQ
What is GA4's AI assistant? Ask Advisor, a Gemini-powered chat inside Google Analytics, rolling out through 2026. Ask in plain English and it returns a summary, a "why," a how-to, or an optimization tip, with a chart and a link to the report behind it. It sits alongside the older Analytics Intelligence (automated and custom Insights) and predictive metrics.
Is GA4's AI free? Yes. Ask Advisor, Insights, and predictive metrics are part of standard GA4 at no extra cost. You pay in scope, not money. It reads one property, GA4 data only, and Ask Advisor is gated by eligibility.
Why don't I have Ask Advisor? It's in beta, English only, and rolling out unevenly. Check the top-right of your property, or the search box. If it isn't there, the rollout hasn't reached you yet. Predictive metrics gate separately, on volume: a low-traffic store often won't qualify no matter how long it waits.
Can Ask Advisor tell me why my conversion rate dropped? Partly. It'll find which device or channel or day the drop sits in, which is real progress. It won't name the cause on the page or what to change, because it only sees GA4 data, not your checkout. For that the analysis has to reach past the numbers, which is the conversion rate drop diagnostic in full.
Is GA4's AI the same as connecting GA4 to ChatGPT or Claude? No. Ask Advisor is Gemini inside GA4, reading one property. Connecting GA4 to your own assistant over MCP brings the same data into the tool you already use, with your choice of model, and with a diagnostic layer on top, a next action instead of a number. The Claude vs ChatGPT for analytics comparison covers the model choice.
Does Ask Advisor work across multiple properties? No. It answers at the property level, one at a time. Cross-property and cross-account questions aren't supported.
Do I need SQL or code for any of this? No. Ask Advisor is plain-English chat. And a hosted GA4 MCP server connects to Claude or ChatGPT through the browser, with no terminal, so the connected route stays point-and-click too. The GA4 MCP server overview walks the options.
By Monday morning GA4's AI will have the drop on your screen, named and charted. Tracking it to the broken step and shipping the fix is still your hour, and it's still the part that pays.