A Google rep called one of our ecommerce clients with an offer that sounded hard to refuse. Switch on AI Max, Google's newest automation for Search ads, in a fresh campaign. Give it a real budget. Watch it find customers your keywords never reached. The client, an online store selling professional trade tools to contractors, said yes and asked us to run it.
We did, and we ran it fairly. We followed the rep's plan, fixed the parts that would have broken the test, and let it spend for three weeks. When we stopped it, it had returned 67 cents for every dollar it spent. The rest of the same account, over the same three weeks, returned about $3.
This article walks through where that money went, what Google's plan got wrong before a dollar was spent, and when we would turn AI Max on again. If a rep has pitched you the same test, this is the readout we wish someone had handed us first.
We Ran AI Max Exactly the Way Google's Rep Drew It Up
The fastest way to dismiss a bad result is to say the test was rigged. So here is how this one was built, because we went out of our way to give AI Max its best shot.
The rep's plan called for one new Search campaign focused on a single product category, a budget of $250 a day, and broad match keywords so Google had room to explore. We kept all of that. We used the rep's own keywords, word for word. We set bidding to Maximize Conversions, which the plan called for while the campaign learned, and we counted only completed purchases as conversions, which matters more than it sounds (more on that below). Ads showed on Google Search only, to people physically in the United States.
AI Max has three parts, and we made a deliberate call on each:
- Search term matching, which lets Google show your ad for searches beyond your keyword list. On, because that is the whole point.
- Text customization, which lets Google write headlines and descriptions from your website. We wanted this off. The rep argued, reasonably, that a test of AI Max without it isn't a fair test, so we turned it on. Then we used Google's text guidelines to block phrases like "in stock," "ships today," and "next day delivery." On a catalog of thousands of products, an ad written from one page's stock status is wrong by tomorrow.
- Final URL expansion, which lets Google choose which page on your site each click lands on. Off, with the rep's agreement, so the test stayed on the product category we were testing.
Before launch, everyone agreed on a checkpoint: 30 sales or 21 days, whichever came first, at which point we would review the results and decide whether to switch the bidding to chase revenue instead of sales. The campaign never reached 30 sales. Day 21 arrived first, and the owner approved pausing it.
That checkpoint is the most useful thing in this whole story. It was set when nobody had seen a number yet, which meant nobody could move it once the numbers came in.
Three Weeks of AI Max, 67 Cents Back on Every Dollar
Here is the result in plain terms. For every dollar the AI Max campaign spent, it brought back 67 cents in sales. Not profit. Sales. The rest of the account, running the store's existing Performance Max and Search campaigns over the exact same three weeks, brought back about $3.05 per dollar.
The cost of getting there tells the same story from a different angle:
- Each click cost about six times more than the account's average click over the same period.
- Each sale cost about four times more than a sale from the rest of the account.
To be fair to Google, that comparison isn't perfectly even. The rest of the account includes Performance Max, which also buys searches for the store's own name and shows ads to people who already visited the site. Those are the cheapest, easiest sales any store has, so the gap between the two numbers is a little wider than a perfectly matched test would show.
But the gap isn't close enough for that to matter. Even the slice of the AI Max campaign that ran on the store's own name, the easiest money in the account, came in under $2 back per dollar. Every other slice lost money outright.
Almost Half the Money Went to Searches Google Wouldn't Show Us
When a campaign underperforms, the first place to look is the search terms report, the list of what people actually typed before they clicked. We split every dollar of this campaign into four buckets.
45% went to searches Google didn't show us. The search terms report only lists searches that meet Google's privacy thresholds, so a large share of spend can sit behind it, unnamed. On this campaign it was almost half. Those hidden searches brought back 82 cents per dollar, which tells you they weren't all junk. It also tells you that you can't manage what you can't see. You can't add a negative keyword for a search you were never shown.
31% went to general product searches. These are the searches the test existed to win: people looking for the product category without naming a store. They brought back 40 cents per dollar. This is the honest heart of the result. AI Max was supposed to find new buyers in exactly this pool, and at more than $4 a click, the buyers it found didn't come close to covering the cost.
14% went to competitor store names. People searching for other retailers by name cost more than $9 a click on average, some above $20, and brought back 8 cents per dollar. Someone typing a specific competitor's name usually wants that competitor. Our shared competitor exclusion list covered the national chains. The specialty stores that sell the same products as our client had been excluded in other campaigns, but those lists never made it onto the test. That one is on us, and it is the cheapest lesson in this article: when you attach shared lists to a test campaign, open them and check who is actually on them.
9% went to the store's own name. This was the best bucket in the test at $1.72 back per dollar, and it is also the least meaningful. These were people who already knew the store and were looking for it. The store's other campaigns were already reaching them. In the test, the store paid about $5 a click for searches on its own name, against an account-wide average click under $1.
That happens because of how Google settles a tie between two of your campaigns. When a search is eligible for both an AI Max campaign and a Performance Max campaign, Google's own help page says the one with the higher Ad Rank serves. So a new AI Max campaign can take your own-name searches away from the cheaper campaign that was already winning them, and then report those sales as its own. We covered this pattern for Performance Max in Performance Max Isn't Your Smartest Campaign. It's Google's., and AI Max does the same thing. AI Max lets you exclude your own brand at the campaign level. Use it on any test that is meant to prove new demand.
AI Max Never Climbed. No Week Came Close to Breaking Even.
The standard defense of a weak AI Max start is that the system is still learning. We have made that argument ourselves. In AI Max Filled Your Store With Junk Search Terms. Sales Went Up Anyway., we told stores to expect an ugly first few weeks and to judge AI Max on its trend, not on any single week. We still believe that. It is exactly why we looked at the trend before calling it.
On the raw report, week one looked like the best week by far, at $1.17 back per dollar. Then we opened it up. Most of that week's revenue was partial credit for a single large order. The store's own analytics gave that sale to a different campaign entirely; Google Ads gave the AI Max campaign a slice of it because the buyer had clicked one of its ads more than two weeks before buying, and Google books that credit on the day of the click, which put it in week one. Take that one order out and the three weeks read 32 cents, 56 cents, and 39 cents.
That is what a flat line looks like. A learning curve has a direction: cost per click falls as Google stops bidding on searches that don't convert, and return per dollar rises week over week. Here, cost per click barely moved. It started near $5.40 and ended near $5.00. The return bounced around inside a range that never got within reach of a dollar.
Could it have turned in week six? Possibly. Some advertisers have reported AI Max taking a month or more before it pays. But every extra week here lost about half of every dollar spent, with no sign of improvement to justify it. A rough start that is improving is the system learning. A start that sits flat for three weeks is the result.
Under 2% of the Spend Came From AI Max's Own Matching
Here is the finding that surprised us most. Google's search terms report now labels where each match came from, so you can see whether a search matched one of your own keywords or matched through AI Max's keywordless technology, the part that reads your website and finds searches your keywords never named.
Of the spend we could see, more than 98% matched our own broad match keywords. Under 2% came from AI Max's keywordless matching. In practice, this test was mostly a brand-new broad match campaign with Google writing some of the ad text. The feature the rep was selling barely got a turn.
That points at the real flaw in the plan, and it isn't a setting. It is the shape of the test. Google's help center says it plainly: AI Max "is not a new campaign type. It is an optimization layer that you activate within your existing Search campaigns." The ecommerce specialist Marco Rodriguez makes the same point from the field: AI Max earns its keep on a Search campaign that is already profitable and has run out of room to grow, where your keywords have stopped finding new searches.
A new campaign with no history isn't that. It has no conversion data for Smart Bidding to learn from, no proven keywords, and no baseline to beat. Put AI Max on top of it and you aren't testing whether AI Max adds anything. You are testing whether a cold broad match campaign can make money in a crowded category in three weeks, which is a different, much harder question.
Three Claims in Google's Test Plan Didn't Hold Up
The rep sent a written implementation guide. We checked it against the account and against Google's own documentation before building anything, and three of its claims didn't survive.
"Spend at least $250 a day, or the algorithm throttles." We found no such minimum in Google's AI Max help pages. The only budget warning there is that AI Max won't be effective on a campaign that is limited by budget. That is a reason to give it enough room, not a published threshold. The $250 figure was the rep's number, not Google's rule.
"Your existing campaigns are held back by budget." The guide named three campaigns as persistently limited by budget, which was part of the case for new spend. Google's own status column showed two of the three were not limited at all. One of those was missing a few percent of its traffic to budget, which is normal. Only the third was genuinely constrained.
"Target this commercial builders audience." The guide recommended an audience segment for commercial builders and trade specialists. We went through every affinity audience Google offers. That segment doesn't exist. The closest real one is for do-it-yourselfers, which is the opposite of a store that sells to professionals.
None of that makes the rep dishonest. Reps are salespeople with goals, and decks get assembled from templates. But it does mean a plan from Google deserves the same checking you would give a plan from anyone else.
Two more traps weren't in the guide at all, and they are worth checking on any new campaign:
- New campaigns inherit your account's conversion goals. Google's help page confirms that every account-default goal is selected automatically when you create a campaign. In this account, those defaults included checkout starts and phone calls alongside purchases. Under Maximize Conversions, a checkout start counts the same as a sale, so the 30-conversion checkpoint could have been "hit" by people who never bought. We switched the campaign to purchases only before launch.
- A setting being off doesn't guarantee its behavior. Final URL expansion stayed off for the entire test, and a handful of clicks, about 1% of the total, still landed on the store's homepage rather than the product page. The money was trivial. The lesson isn't: check the landing pages report on any AI Max campaign, even when the switch says off.
Four of Five AI Max Sales Were New Customers. The Campaign Still Lost Money.
Every test has a silver lining if you look hard enough, and Google will point you at this one. By Google Ads' own count, about four out of every five sales from the campaign came from new customers. For a store that wants to grow, that sounds like exactly what AI Max promised.
Two things keep it from changing the verdict.
First, the label depends on who is counting. Google's new-customer figure relies on its own customer matching. The store's analytics, which counts returning visitors rather than past buyers, put more than half of the same orders on people who had been to the site before. Neither number is wrong, because they measure different things. But a figure that moves that much depending on the tool shouldn't carry a budget decision.
Second, even taken at face value, new customers acquired at 67 cents of revenue per dollar are expensive. That is revenue, not profit. After the cost of the product, a typical retailer keeps a fraction of each sale. For those customers to pay off, they would need to come back and buy several more times. That can be true for a store with loyal repeat buyers, and it is worth asking. It is also a bet, not a result, and it should be priced as one before the next test starts.
AI Max Belongs on a Search Campaign That Already Works
None of this makes AI Max a bad product. Our verdict is narrower. AI Max was tested in the wrong place, and the result told us so cheaply and quickly, because the stop date was set in advance. Here is how we would set it up next time, and how we would advise any store owner whose rep calls with the same pitch.
Start from a winner. Switch AI Max on inside an existing Search campaign that is already profitable and has stopped finding new searches. That is the job Google built it for, and it is the only setup where its results mean anything.
Run Google's own experiment, not a new campaign. If your account is eligible, Google now offers a one-click AI Max experiment that splits an existing campaign 50/50, half with AI Max and half without, at the same time. Same season, same promotions, same competitors. Whatever difference shows up is actually AI Max, not the calendar.
Exclude your own name. Add your brand as a brand exclusion so the test has to find new demand instead of collecting sales your other campaigns were already making for less.
Open every list you attach. Check that your competitor exclusions include the stores that sell what you sell, not just the big national names.
Count only what you are paid for. Set conversion goals at the campaign level to purchases, or to the one lead type that actually turns into revenue.
Write the stop rule down first. A date, a minimum return, and who decides. Agree on it before launch, while nobody has a number to defend.
We would also read Google Wants You to Turn On AI Max. Most Small Businesses Shouldn't Yet. before saying yes to any of it. Its readiness checklist still holds, and this test is a working example of what happens when a test skips the foundation.
A Google rep's job is to get new products in front of advertisers, and some of those products will end up earning a permanent place in your account. Your job is to make every one of them prove it on your numbers, in a test built to answer the right question. If you would like a second set of eyes on a test your rep has proposed, or someone to run it fairly and tell you the honest answer, that is a big part of what our Google Ads management team does every day. And if you would rather run it yourself, you now have the checklist we wish we had started with.
