Installing Conversion Tracking by Hand Is Dead. Mostly.
Six of the seven tasks in a classic GA4 and GTM conversion tracking setup can be executed by an AI agent today. PaidSync ships 37 Google Tag Manager tools and 28 GA4 tools that create the tag, the trigger, the variables, the version, the key event, and the Google Ads link, every write gated behind your approval. The official Google Analytics MCP can execute none of them because it is read-only, and the one task an agent cannot finish alone, real-browser QA, is exactly the one you should keep.
For fifteen years, conversion tracking setup has been the same ritual. Open Tag Manager. Click New Tag. Pick a tag type from a dropdown. Click into a trigger. Type an event name and hope you spelled it the same way the developer did. Preview. Squint at the debug pane. Publish. Switch to GA4. Wait for the event to appear. Toggle it to a key event. Switch to Google Ads. Link, import, bid. Seven tasks, three products, one afternoon, and a decent chance one step silently failed.
This page is the autopsy, not the manual. If you want the step-by-step build, the tutorial is How to Set Up Google Tag Manager With AI From Claude or ChatGPT. Here we score each task of the ritual: can an agent execute it today, or is it still yours?
Why This Argument Exists Now
Google itself started shipping MCP servers, which is how AI assistants get hands. The official Google Ads API MCP server was open-sourced on October 7, 2025, read-only. The official Google Analytics MCP is also read-only. Tag Manager has no official MCP server at all as of August 2026. So the official position is: your AI may look, but not touch. That makes the interesting question concrete. Of the seven tasks in a conversion setup, how many need touching?
The read-only receipt (checked live August 1, 2026)
The official Google Analytics MCP at github.com/googleanalytics/google-analytics-mcp requests one OAuth scope, analytics.readonly. Its tools are seven readers: account summaries, property details, Google Ads links, core reports, funnel reports, custom dimensions and metrics, realtime reports. No create, no update, no delete. An assistant connected to it can describe your broken conversion tracking in perfect detail. It cannot fix any of it.
The Seven Tasks, Scored
The classic setup, in the order you were taught to do it. Agent means PaidSync's MCP tools running inside Claude, ChatGPT, or any MCP client, with approval gates on writes. Official MCP means Google's own Analytics server.
| Task | By hand | PaidSync agent today | Official Google Analytics MCP |
|---|---|---|---|
| 1. Create the conversion tag | Dropdowns and typing | Executes (create_gtm_tag) | No (read-only, no GTM) |
| 2. Create the trigger | Click, type, hope names match | Executes (create_gtm_trigger) | No |
| 3. Create the variables | One at a time | Executes (create_gtm_variable) | No |
| 4. Preview and QA | Debug pane, Tag Assistant | Shared (reads status, you click) | Realtime reports only |
| 5. Publish the container | Version notes, submit | Executes (publish_gtm_version) | No |
| 6. Mark the key event in GA4 | Wait, then toggle | Executes (create_ga4_conversion_event) | No (read-only) |
| 7. Connect it to Google Ads | Link, import, confirm | Executes (create_ga4_google_ads_link) | No (lists links only) |
Six executed, one shared. Now the detail, because the detail is where the honest caveats live.
1Create the conversion tag
Agent executesBy hand, this is the dropdown maze. Tag type, measurement ID, event name, parameter rows added one click at a time. It is not hard. It is tedious, and tedium is where typos breed.
An agent does it in one instruction. PaidSync's GTM group covers the full tag lifecycle: create_gtm_tag, update_gtm_tag, delete_gtm_tag, plus list_gtm_tags and get_gtm_tag so the agent reads what exists before it writes. You say "create a GA4 event tag for purchase with transaction_id and value", it drafts the exact configuration, you approve, it lands in the workspace.
2Create the trigger
Agent executesThe classic failure point. The developer pushed an event called purchase_complete, you typed purchase-complete in the trigger, and the tag never fired. Nobody noticed for three weeks.
create_gtm_trigger handles custom event triggers, click triggers, form triggers, and page view conditions, with update_gtm_trigger and delete_gtm_trigger for changes. The agent's advantage is not speed. It is that it reads the container first and matches names against what is actually configured instead of what someone remembers.
3Create the variables
Agent executesData layer variables are the purest busywork in the stack. Seven ecommerce parameters means seven identical dialogs: name, type, data layer key, save. Repeat.
create_gtm_variable creates them in a batch from one description, and list_gtm_builtin_variables enables the built-ins your triggers depend on. The agent also catches the classic drift where the variable reads ecommerce.value but the site pushes ecommerce.revenue, because it compares both sides instead of assuming.
4Preview and QA
Shared, and the human half mattersHere the obituary stops. GTM's preview mode is a real browser session: you click the button, watch the tag fire, inspect the payload Tag Assistant shows you. No API replaces clicking your own checkout on a phone.
What the agent contributes is the bookkeeping around the QA pass. get_gtm_workspace_status shows exactly what changed before you test. run_ga4_realtime_report confirms the event is arriving in GA4 while you click. After publish, get_gtm_version_live verifies what is actually serving. The agent watches the instruments. You fly the plane.
5Publish the container
Agent executes, behind a gatePublishing was never hard. It was scary, because publish is live. That is why this step is approval-gated by design.
create_gtm_version snapshots the workspace with proper version notes (the agent writes better version notes than most humans, because it lists every change it made). publish_gtm_version pushes it live only after you confirm. PaidSync writes run with dry_run previews, destructive operations require explicit confirmation, and you can revoke a session's access at any time. The agent does the work. You keep the launch key.
6Mark the key event in GA4
Agent executesThe step everyone forgets. The tag fires, GA4 receives the event, and it sits there as a plain event because nobody toggled it to a key event. Google Ads has nothing to import, and the campaign optimizes toward nothing.
create_ga4_conversion_event does the toggle the moment the event exists, and list_ga4_conversion_events confirms the current state first. This is also the cleanest illustration of the official-MCP gap: Google's own Analytics server can list your key events beautifully. It cannot create one. Read-only means read-only.
7Connect it to Google Ads
Agent executesThe last mile, and the one that spans two products. GA4 must be linked to the Google Ads account, and a conversion action must exist for bidding to use.
create_ga4_google_ads_link creates the link. On the Google Ads side, PaidSync's write tools create and update conversion actions directly, so the agent finishes the chain it started in task one. The official Google Ads API MCP shipped read-only, so this side of the chain is equally out of its reach. The deeper walkthrough for this task is Set Up Google Ads Conversion Tracking with AI and GTM.
Before an agent touches anything, let it read. Both audits run in one conversation, on your real container and property.
Audit your GTM container Audit your GA4 setupWhere Manual Still Wins
The title says mostly, and this section is why. Five parts of conversion tracking are alive, human, and likely to stay that way. Pretending otherwise is how you end up with confident dashboards full of wrong numbers.
The dataLayer code on your website
Every tag on this page consumes data that your site must push. Someone has to write dataLayer.push({event: 'purchase', value: 129}) into the checkout template, with the real order value, at the real moment of purchase. That is site code, not container config. An agent with Tag Manager API access does not edit your website, and a tracking setup with a perfect container and no dataLayer is a very well-organized zero.
Real-browser QA
Task four, restated as a rule. Realtime reports tell you an event arrived. They do not tell you it fired on the right click, once instead of twice, with the right value, on mobile Safari with an ad blocker installed. Twenty minutes in GTM preview mode and Tag Assistant before you trust a number is the cheapest insurance in marketing.
Consent Mode v2 and legal signoff
Which consent banner you run, which region rules apply, what fires before consent, whether modeled conversions are acceptable to your DPO: these are legal positions, not API calls. An agent can read your configuration and flag gaps. The decision belongs to a human with authority to sign it, usually one with a law degree nearby.
Server-side tagging infrastructure
A server-side GTM container needs hosting, a first-party subdomain, DNS, and someone accountable when it goes down. That is infrastructure work with budget and ownership questions attached. Out of scope for a conversation with an agent, and honestly out of scope for most afternoons.
Deciding what to track
The most expensive tracking mistakes are not broken tags. They are perfectly firing tags on the wrong definition of success. Whether a newsletter signup is worth optimizing toward, whether a lead form counts before or after qualification, what a conversion is worth: that is business judgment. An agent executes your taxonomy faster than you can. It should not be the one choosing it.
The Honest Scoreboard
Dead as manual work: tag creation, trigger creation, variable creation, versioning and publish, key event configuration, Google Ads linking. Six tasks of clicking that an agent executes today through PaidSync's 37 GTM and 28 GA4 tools, each write behind your approval.
Alive and shared: preview and QA. The agent reads workspace status and realtime reports. You click the real buttons in a real browser.
Alive and fully yours: dataLayer code, consent and legal posture, server-side infrastructure, and the decision of what deserves tracking at all.
Not an option for any of it: the official Google Analytics MCP, which is read-only by design. Good diagnosis, no hands.
The pattern is older than software. The typesetter died, the editor did not. What died in conversion tracking is the transcription work, the part where a human relays a decision to a series of dropdowns. What is left is the deciding, the verifying, and the accountability. Those were always the job.
PaidSync signed six paying agencies in its first four days in May 2026, and conversion tracking cleanup is one of the first things agencies run, because every inherited account has a container full of archaeology. The Free tier's 15 calls per month cover a first audit. Pro from $99 per month covers the setup work itself.
Connect your accounts and run the first audit free. 15 tool calls per month, no card.
Get started for free See pricingFrequently Asked Questions
Can an AI agent set up GA4 conversion tracking end to end?
Almost. Through PaidSync, an agent creates the GTM tag, trigger, and variables, publishes the container version, marks the GA4 key event, and links GA4 to Google Ads. Two things stay human: the on-site dataLayer code that feeds the tags, and the real-browser QA pass before you trust the numbers. The full walkthrough is in the GTM setup tutorial.
Is the official Google Analytics MCP read-only?
Yes. The server at github.com/googleanalytics/google-analytics-mcp requests the analytics.readonly scope and exposes reporting and metadata tools only. No create, no update, no delete. Checked live August 1, 2026. It answers questions about your data. It does not change configuration.
Can an AI publish a GTM container version?
Through PaidSync, yes. create_gtm_version snapshots the workspace and publish_gtm_version pushes it live. Writes run through approval gates with dry_run previews, and destructive operations require explicit confirmation. Google has not shipped an official Tag Manager MCP as of August 2026, so no official server does this.
Which parts of conversion tracking still need a human in 2026?
Five parts: writing the dataLayer code on your website, real-browser QA with GTM preview mode and Tag Assistant, Consent Mode v2 decisions with legal signoff, server-side tagging infrastructure, and deciding which actions deserve tracking at all. The clicks are automatable. The judgment is not.
Can an AI agent import conversions into Google Ads?
Through PaidSync, the agent creates the GA4 to Google Ads link with create_ga4_google_ads_link and creates conversion actions directly with create_conversion_action, both behind approval gates. The official Google Ads API MCP server, open-sourced by Google on October 7, 2025, shipped read-only and cannot do either.
What does an agent-run conversion setup cost?
PaidSync's Free tier includes 15 tool calls per month, enough to audit an existing container. Pro starts at $99 per month for 600 calls. A full setup pass, tag, trigger, variables, version, publish, key event, and Google Ads link, is about a dozen calls. Details at paidsync.ai/pricing.