Making a Google Ads campaign yourself with AI: the agent builds it. Then comes the measuring.

Every piece about Google Ads and AI is about ad copy. Let ChatGPT do your keywords, your negatives, your headlines, saves you 75% of your time. Nothing wrong with any of that. It is just the answer to a question from two years ago.

Because the question has changed. Nobody asks an AI to write their ads any more. You connect it to your account and it does the work itself. And it goes well, better than most people think.

Until you have to measure it. That is a different game entirely.

Connect one of the community Google Ads MCP servers, such as FGRibreau's or isteam's, and your agent really does get to touch the controls. It creates campaigns, adds ad groups, sets bids, writes ads, pauses what is not running. Some have a lock on it: the agent proposes the change, you see what is about to happen, and it only goes through once you say yes, with a budget ceiling over the top. This is not a demo. You can genuinely work with this.

And it really is as fast as they say

That afternoon in which a full campaign structure appears: that is real. So is the 75%. Keyword research, negative lists, thirty ad variations, a tidy account structure, an agent is good at all of it and never gets bored. So if the question is whether AI can build a Google Ads campaign: yes. And not marginally.

One thing stands out before we go on. Google shipped its own Google Ads MCP and deliberately made it read-only. In their words, in the current release it cannot modify bids, pause campaigns, or create new assets. It gets to look, nothing more. Everything that does write to a live ad account comes from the community. Nothing wrong with that, some are excellent. But it does say something that Google would not hand that half to a model itself.

Your campaign is live, and from that moment it costs money. Now Google has to learn which clicks are worth it, and it learns that from one thing only: the conversions you report back. Everything above was about what your campaign looks like. From here on you decide what Google steers on.

So you set up conversion tracking. Your agent helps with that too. A tag fires, a conversion action appears, a number shows up. It works.

And then you have to tell Google what a customer is. That is where the afternoon ends.

Conversion tracking sounds like one thing: let Google know when a customer comes in. In practice it is a minefield of settings and small gotchas, and nobody shows you the way.

It starts at creation. You can only measure a conversion once it exists, but when you create it, its status sits on “unverified” until the tag has actually fired once, and no push notification tells you when that happens. So you sit there refreshing the tracking-status column until it flips. That is not a fault. That is simply how it works.

Then you get to choose how to report it, and that is when it gets really good. First: where does the conversion happen? On your site, in an app, on a phone call, or offline after a quote. Then, separately: how does the data reach Google? Through a tag on your site, deployed by hand or through Tag Manager. Through an offline import when the deal closes later. Through tagless routes like a GA4 import or a call-forwarding number. And layered on top of any of those, Enhanced Conversions sends hashed customer data to fill the gaps. Two axes, a fistful of combinations, and nowhere does it say which one is right for you.

And the obvious route just got locked for newcomers

Pick the route every tutorial and every language model still recommends, the offline import through the Google Ads API, and you run into something that changed in June. Since 15 June 2026 Google no longer allowlists new developers for that import; a token with no recent history gets `CUSTOMER_NOT_ALLOWLISTED_FOR_THIS_FEATURE` and is pointed at the Data Manager API instead. Your agent writes that code correctly and it fails, and nowhere in the error does it say that the reason is an access list, not your code.

This is the pattern, not one incident. The platform shifts under you on a Tuesday, the tutorials do not catch up for months, and the thing that breaks is never the part you were watching.

Say you found the right route. The whole structure still hangs on the gclid, the click id Google attaches to an ad click. No gclid, no match, no value. And that id is less certain than you think. In private browsing Apple strips a subset of tracking parameters out of your links by default, and anyone who turns the setting on gets that in regular browsing too. Apple never names the gclid, so whether it falls inside that set is not stated in black and white. What is stated: Safari lets parameters through that only serve campaign attribution, and not the ones that track down to a click or a person.

There is a second catch that is easier to miss: if someone arrived from another site, third-party scripts reading the URL get a version with no query parameters at all. So your id can be sitting right there and still be invisible to your tag. And that id had to be captured and stored months earlier, before you knew you would ever need it. While the conversion that actually matters often does not even happen on your site: it closes six weeks later on the phone. So your admin system now has to start talking to your ad account. Every day, forever.

And meanwhile that dashboard keeps reminding you that you have not switched on consent mode, or asking whether you want to share your lead data with a partner network. Small checkboxes, big consequences, no explanation. This is the work you skip when you do not know it exists.

If this went wrong loudly it would be half as bad. An integration that falls over, you see that. But nothing falls over here. The tag fires, a number appears, the graph goes up. No message arrives telling you that you are bidding on a signal that has nothing to do with your revenue. You find out when you go looking, and you only go looking once something starts to nag.

Because that bidder does have to learn something in the meantime. At low volume “someone became a customer” is far too small a number, so you add softer signals: someone who stayed thirty seconds and scrolled, someone who opened the form. Search Engine Land calls those micro-conversions. The example they give says it all: an account with 500 pageviews, 200 add-to-carts, 50 lead form starts and 10 purchases hands the bidder 760 signals for the 10 that actually matter. Their advice is to strip those soft signals out. But read their own threshold: below thirty real conversions a month you may need them, and only above sixty can you really do without.

Most businesses that advertise never reach thirty. So the advice “turn off your micro-conversions” simply is not for them. They need those signals just to keep the bidding running.

Search Engine Land draws the line in real conversions per month:

Sit between those two lines, as most advertisers do, and the counter in your dashboard cannot be your real result. It has to carry things that are not customers, just to keep the machine fed. That is not a fault. It is what low-volume lead generation looks like.

And on the other side Google sees too little

Because Google only credits revenue it can trace back to a click it sold you. If that customer came in on a tip, or simply picked up the phone, there is no click id, and Google never sees that revenue. For most smaller advertisers that is not a rare edge case but a large share of the business the ad account will never know about.

So your dashboard counts too many conversions and too little money. At the same time. And those two errors do not wipe each other out, they stack. What you are left with is a number that is wrong in both directions and looks exactly like a real number. And that is what you steer your budget with.

The fair part of this story: Google sees the problem too. At Google Marketing Live 2026 they announced journey-aware bidding, in their own words “currently in beta”. It lets Search campaigns steering on a Target CPA learn from both biddable and non-biddable conversion goals. Exactly the right direction: soft signals get to steer without inflating the counter.

Except: Google describes it exclusively for Search campaigns on a Target CPA. About Performance Max or bidding on value it says nothing in that context. So if that is what you run, nothing has been promised to you either. The fix is coming, just not for you yet.

Until then this is simply work. Not complicated work, just work: figuring out how to measure, along a route that just got locked, with an id you should have stored months ago, getting your admin system to talk to your ad account. And keeping all of it current while the platform shifts underneath you. An agent helps you with that. It does not make the call, and when it stops it does not notice.

This is not an argument against using AI for your Google Ads. Our own campaigns run on AI, every day, and we would not want to go back.

The split is clean. Everything before the money, the structure, the keywords, the copy, the endless variations: exactly what you should hand to an agent. It is faster than you and it does not get tired. Do it.

What you cannot hand over is the judgement about a number. Not because your agent is stupid, but because it looks at the same dashboard you do. Ask it whether the campaign is running well and it reads the same inflated counter you do, and advises you on that. It is just as wrong, and for exactly the same reason: that number never meant what it looked like it meant.

That right-hand list is short. It is also the only list where your money gets decided.