I've built automated testing for GA4 and Google Tag Manager, and it changes what a tracking audit can prove. Give it a website address and nothing else. It shops, answers the cookie banner both ways, and reaches the checkout on its own, reading every tracking request on the way.

I've pointed it at more than 110 UK websites. First 14 well-known names in 11 sectors: retailers, a broadcaster, a newspaper, a bank, a software company, a charity, an estate agent, a car dealer and more. Then 97 more across retail, property, payments and foreign exchange, tourism and health.

On two of the retailers it found a product, typed in a postcode, chose home delivery and walked itself to the checkout without any help. Each full test takes minutes. By hand, in Preview mode, it would take days.

  • 110+UK websites tested, each given nothing but its address
  • 19,000+tracking requests read, one by one, across desktop, phone and "Reject all"
  • 1,500+problems found, each with the evidence, the owner and the fix
  • 1 in 3sent an email address typed into the site search to an ad platform

From my test runs in November 2026. No site is named.

Why I built it.

Tracking never breaks loudly. A release changes a template, an agency adds a pixel, a plugin updates, and purchases quietly stop being recorded, or a tag starts firing before consent. Nobody notices until the numbers stop making sense, or a regulator does.

Testing by hand only ever covers a few pages, on one device, in one consent state. Doing it properly means every journey, on desktop and phone, accepting and rejecting cookies, and reading hundreds of requests each time. No person does that every time. So I stopped asking a person to.

Watch it work.

Here's one of the retailer runs as the machine reported it, shortened, with anything that would identify the site taken out. The whole test took about seven minutes.

$ node watch.mjs --url https://www.[retailer].co.uk
Working out the config...
  cookie banner: found, "Reject all" on the first screen
  basket try 1: product A ...... the button did nothing, trying another
  basket try 2: product B ...... reached the first checkout page
  saved the config for review
Browse to checkout (desktop)... done
Browse to a product (mobile)... done
Reject all... done
With an ad blocker... done
Cookie banner check... done
Measuring what tracking costs on 3 pages... done

19 new findings: 7 critical, 5 high, 7 medium

Nobody told it where the basket was, which product to choose or that the shop would ask for a postcode. It worked all of that out, and kept only the steps that worked, so the next run follows the same path in seconds.

How it works.

  1. Step 1

    One address

    That's all it needs. No script to write, no tags to add.

  2. Step 2

    It explores

    It finds the banner, the search, a product list, a product and the forms.

  3. Step 3

    It shops

    On desktop and phone, accepting and rejecting cookies, all the way to checkout.

  4. Step 4

    It records

    Every data layer push, every request to every platform, every cookie.

  5. Step 5

    It checks

    Consent, data quality, Tag Manager, speed and GA4 itself.

  6. Step 6

    It reports

    Each problem with its evidence, why it matters and exactly how to fix it.

The machine against a person.

What gets testedBy hand, in Preview modeThe machine
JourneysA few pages someone remembers to clickSearch, list, product, basket and checkout, plus video and forms
DevicesUsually desktopDesktop and phone, every run
ConsentUsually "Accept all"Before a choice, after accepting and after rejecting
What's readThe tags Preview listsEvery request each platform receives, and every cookie it sets
RegionsWherever the tester sitsUK, EU and US banners, checked separately
ProofScreenshots, if someone remembersThe request, push or cookie behind every finding
After the fix"We think that's sorted"Run it again: each finding confirmed fixed, or not

What it checks.

Consent

What happens before and after the visitor chooses.

  • What every platform sends and stores before a choice, after "Accept all" and after "Reject all"
  • The four Consent Mode v2 signals Google's tags really receive
  • Cookies missing from the site's own cookie list, or filed as "necessary" when they track
  • Whether "Reject all" is as easy to see and use as "Accept all"
  • Whether the ad platforms still receive data when the browser sends Global Privacy Control

Data

Whether what's sent is right, not just whether it's sent.

  • Every ecommerce event against GA4's recommended fields, values adding up to the items
  • Broken values, double counting and personal data leaking into requests
  • An email address typed into the site search, and every platform the search term carries it to
  • What Safari's cookie limits do to returning visitors and ad clicks
  • The shape of every event, learned run by run, so a field lost in a release is flagged next time

Setup

The machinery behind the tags.

  • Tag Manager tags with no consent settings, pixels pasted in as custom HTML, paused tags
  • What changed in the container since the last run, and which problems arrived with it
  • 40 pages from the sitemap, each checked for its tag, the banner and a GA4 page view
  • Speed, the server container, and what visitors with an ad blocker take with them

GA4 itself

With read-only access to the property.

  • Data retention, streams, key events, custom dimensions and the BigQuery link
  • 28 days of reports: "(not set)" landing pages, Unassigned traffic, payment providers taking credit, orders counted twice

What a finding looks like.

Every problem comes out the same way: what was found, the evidence, why it matters, who fixes it and how. Here's a real one, shortened:

HighData · Checkout journey

Meta receives x-fb-cd-contents as "[object Object]"

Evidence
x-fb-cd-contents=[object Object], in the request sent to Meta
Why it matters
The basket was turned into text the wrong way, so Meta gets the words "[object Object]" instead of the product IDs, quantities and prices behind catalogue matching, dynamic ads and value optimisation.
Who fixes it
Whoever manages Tag Manager, with the web developers. Under an hour.
The fix
Find the variable feeding the field and send it as JSON text, or as the structure the tag expects.
How to check
The value holds real product data in the Network tab, and the next run confirms it.

That error looks perfectly normal in Meta's interface. Ads keep running. You'd only ever see it by reading the request itself.

What it found across those sites.

None of these show up in a GA4 report or a quick Preview session. All of them show up in the requests:

  • An email address typed into the search box reaching ad platforms in plain text, on one site in three, because the results page puts the search term in its address
  • Analytics and session recording sending data before the visitor had chosen anything
  • "Reject all" hidden behind a settings button, and an advertising cookie set after the visitor rejected
  • Tracking cookies a consent tool had filed as "strictly necessary", so they were set whatever the visitor chose
  • Ad pixels in Tag Manager with no consent settings at all
  • Most of a site's sampled pages loading no tag, so nothing was measured there
  • An old container full of paused tags, still loading on every page and doing nothing
  • A server container on a separate domain, so the first-party cookies it was built for were third-party after all

What 110 websites taught it.

I didn't just test it where it would work. I pointed it at some of the biggest sites in the UK to find where it would break, and every failure made it better.

  • Four of the 14 turned it away. Bot protection blocks test browsers along with bad ones. It now stops, says so, and explains how the developers can let it in. It never tries to sneak past.
  • One consent tool thought it was in Ohio. It was being shown the American banner. A UK visitor could see something completely different, so it now asks for the UK rules and checks Germany, France and the US separately.
  • Testing too fast gets you shut out. After a few hours of back-to-back tests from one address, a run of sites started refusing it. It now tests one site at a time, with a gap between each.
  • A UK view needs a UK address. It now runs from a server in London too, so cookie banners and consent behave exactly as they do for UK visitors.
  • Shops ask questions before the basket. Postcodes, delivery choices, size pickers, products not sold online. It now answers them the way a customer would.

Safe by design.

  • It never pollutes your data. Every request to GA4 and the ad platforms is recorded, then answered by the test browser itself. Nothing reaches your reports or your ad accounts.
  • It never buys anything. It stops before personal or payment details, and won't press a button that looks like placing an order.
  • It never hides anything. Personal data in the evidence is masked, and every finding shows the request that proves it.

The machine finds it. I fix it.

It finds what's broken and proves it, across every journey, in minutes. What it doesn't do is decide what your business should measure, which KPIs matter, or how a report should change a decision.

That's my job, and it's what every audit is for. The difference now is that the audit starts from evidence, not from a few hours in Preview mode.

Sources.

Checked on 11 November 2026.