AI can build your GA4 Explore reports for you by clicking through GA4 in a browser, and IvI SEO used Codex to build 11 of them for a UK building society, about 4.5 hours faster than I'd have managed by hand. That included 36 shared report tabs and step-by-step funnels for 15 online forms. Most people don't know AI can do this. You tell it what you want to see. It does the clicking.

Can AI build GA4 Explore reports for you?

Yes, AI tools like Codex (from ChatGPT) and Claude can build GA4 Explore reports for you using browser control: the AI opens GA4 in a browser and clicks through the Explore section just like you would. You log in yourself, and it does the building. On IvI SEO's building society project, Codex built 11 explorations, set up funnels for 15 online forms, named every tab, set the date ranges and shared the finished reports. By my estimate, it did it about 4.5 hours faster than I would have.

Quick explainer if you've never opened it: Explore is the part of GA4 where you build your own custom reports, like funnels (how many people get from step 1 to step 2 to step 3) and tables. It's powerful, but fiddly. Loads of dropdowns, conditions and tabs, all set by hand.

Here's what Codex built:

  • Four reports for the client, with 36 tabs between them: savings application journeys, business and other form journeys, enquiry totals, and an application-completion check.
  • Seven private working reports: page-by-page activity for every form page (31 tabs), tracking checks, a data quality check, and a starting point for the app.

The funnels were the big job. Each of the 15 forms had its own funnel, and every numbered step of every form had to be set up as its own stage. That's 94 form steps, each needing the right event, the right page and the right step number.

That's hours of clicking if you do it yourself.

And it's the same trick I used to get AI to upload content to WordPress. Different screens, same idea.

How do you tell AI exactly what GA4 report to build?

Tell the AI what question each report needs to answer, which pages and events it should use, and what a finished report looks like, before it ever opens GA4. IvI SEO's rule is the same as for any AI job on a client's account: the less it has to guess, the less can go wrong. For the building society, that meant a list of every form, its exact page address, its numbered steps and its thank-you page.

Start with the question, not the report. The client's brief asked, in plain terms: when someone starts an application on the website, how far do they get, where do they drop off, and do they finish?

Then work backwards to what GA4 needs to answer that:

  • Which forms? Codex built this list itself. It went through the whole site and found 31 pages with forms on them, 32 forms in total.
  • Which events already exist? The site was already recording when someone started an application, moved through a step and finished. Codex checked which of those were actually coming through before building anything on top of them.
  • How is each form identified? Here, by its exact page address. That meant we didn't need the developers to add form names first.
  • What counts as finished? A visit to the right thank-you page, straight after the form. Not just any thank-you page.

A starter brief

You don't need a clever prompt. You need the job written down in plain words. Something like:

Build GA4 Explore reports for [site] in the GA4 property I'm logged into.

The question: for each application form, how many people start, how far they get through each step, and how many reach the thank-you page.

Here's the list of forms with their page addresses, their numbered steps and their thank-you pages: [file].

Build one funnel per form. Name each tab after the form in plain English. Set the date range to [dates]. Save everything as private until I've checked it.

Keep me updated. If you're unsure, ask me. If you're stuck, ask me. Don't change any tracking, settings or user access without asking me first.

What can AI set up in a GA4 exploration?

AI can set up pretty much anything you can click in GA4's Explore section: funnels and their steps, filters, rows and values, tab names, date ranges, sharing and tidying up. On IvI SEO's building society project, Codex built step-by-step funnels for each form, page-by-page activity tables, combined "steps plus thank-you page" journeys, and then renamed, re-dated, tidied and shared the lot.

Here's what that looked like in practice.

Funnels, one per form. Each step in the funnel was set with three conditions: the step event, the form's exact page, and the step number. The steps had to happen in order, but not necessarily one straight after the other. Every step got a plain-English label.

Steps plus the finish line. The first version showed steps in one report and completions in another, which was a pain to read. So Codex rebuilt nine of them to add a final stage: the matching thank-you page, reached straight from that form. One journey, one tab.

Working round GA4's limits. GA4 funnels max out at 10 stages. Two forms already had 10 steps, so there was no room for a finish line. Codex put their completion tab right next to their steps tab instead.

Page activity tables. 31 tabs, one per form page, showing which events happened on that exact page and how many people triggered them. Handy for spotting whether tracking was working at all.

Can AI set up tracking in Google Tag Manager too?

Yes, AI can set up tracking in Google Tag Manager through the browser too, and it's where the guardrails matter most. On IvI SEO's building society project, Codex added step tracking to three more forms and thank-you tracking for six enquiry forms, tested it all in Tag Manager's preview mode, and only published it once I explicitly said go. After it went live, Codex checked the live site was sending the new data and that it turned up in GA4.

Why was it needed? Some forms weren't recording their steps at all, and some thank-you pages weren't being tracked. No tracking, no report. So the reports and the tracking went hand in hand.

Here's how it ran:

  1. Prepare. Codex built the new tags in a draft workspace, not the live site. Tag Manager (GTM) is where your tracking codes live, and nothing in a draft affects real visitors.
  2. Test before anything goes live. It checked, offline, that the thank-you tracking only matched the right page from the right form: 112 checks, all passing. Then it used GTM's preview mode to watch the new tags fire on the real forms. And it checked they didn't fire where they shouldn't, like on a thank-you page visited directly.
  3. Wait for the OK. Codex didn't publish anything until I explicitly asked it to. It went live as version 16 on 15 September 2026.
  4. Check it's really live. Codex loaded the live tracking code the way any visitor's browser would and confirmed the new tags were in it.
  5. Check the data arrived. A few days later, every step of the three newly tracked forms was showing up in GA4's reports.

Tracking changes affect live data, so this is the bit to be most careful with. The rule was simple: prepare, test in preview, show me, and only publish when I say.

What guardrails should you give AI in GA4 and Tag Manager?

Log in yourself, let the AI build and read reports, and make it ask before anything that changes how data is collected or who can see the account. On IvI SEO's building society project, the rules were: no tracking goes live without my say-so, client reports are shared read-only with existing users only, no new users are added, and unused report fields are removed without deleting any underlying data.

The full set I'd use:

  • You log in, not the AI. It works inside your session. It never needs your password.
  • Keep me updated. If you're unsure, ask. If you're stuck, ask. If there's a better way, tell me before you change anything. Same four rules as every AI job.
  • Nothing goes live in Tag Manager without an explicit "publish". Draft, test, show me, wait.
  • Don't touch who has access. Share with people who already have access, read-only. Don't add users or change permissions.
  • Build new reports as private first. Only share them once they've been checked.
  • Leave existing reports alone. On this project, an older unrelated exploration was left exactly as it was.
  • Label anything unproven. If something can't be verified, say so in the report, not just in the chat.
  • Don't overclaim. A thank-you page view is a website confirmation. It isn't proof the business accepted an application. The reports said exactly that.

What doesn't work well yet?

In IvI SEO's experience, the main snags are GA4's loading lag (which can make the AI read stale settings), the 10-stage limit on funnels, the wait before new data shows up in reports, and pop-up windows the AI can't fully control, like Tag Manager's test window.

In a bit more detail:

  • Loading lag. Covered above: make it wait, then check, then reload and check again.
  • The 10-stage limit. Longer forms need their completion in a separate tab. Plan for it up front.
  • Processing delay. New tracking shows up in GA4's live view quickly, but can take a while to appear in Explore reports. Don't panic if a brand-new funnel looks empty on day one.
  • Pop-up test windows. Early on, Tag Manager's test window couldn't be fully controlled through the browser, so I checked some things myself. Once it was live, the AI could run the checks on its own.
  • Fixed date ranges. A saved date range doesn't roll forward by itself. If the client wants "last 30 days", say so.

None of it stopped the job. It just needs planning for.

Want reports like this for your site?

If you've got a GA4 account full of data and no idea where people are dropping off, take a look at how IvI SEO can help. I'll tell you honestly what's worth tracking and what isn't.