Google is turning paid search into a system that makes more decisions on the marketer's behalf. AI Max is expanding matching, generating creative variations, and moving budget toward the queries its models believe will perform. Microsoft is pushing a similar idea with its own AI Max rollout.
The pitch is familiar: give the machine more room and it will find demand you couldn't reach with a tidy keyword list. The catch is that the machine also gets more room to decide what your brand means, which customers are worth chasing, and what counts as a successful visit.
That changes the job. Paid search isn't disappearing, but the old discipline of managing keywords, ads, and bids one control at a time is losing its center of gravity.
The keyword list is no longer the product
Google's own documentation says AI Max can use broad match-style search term expansion, text customization, and final URL expansion to match people with pages and messages beyond the advertiser's original setup. Google also says campaigns using text customization and automatically created assets will be automatically upgraded to AI Max beginning in September 2026, according to its current Ads Help documentation.
That is not a small feature update. It moves the campaign from a set of instructions toward a set of permissions.
A keyword list used to be a practical expression of intent. A brand could decide which questions it wanted to answer and which ones it didn't. With more automated expansion, the list becomes a starting signal. The real campaign is the combination of landing-page content, conversion data, account history, exclusions, and the model's interpretation of all of it.
This creates a useful test for every account: if the model had to infer your offer from your site alone, would it get the promise right? If the answer is no, adding more automation will not solve the problem. It will scale the ambiguity.
Control moved upstream
The biggest mistake is treating AI Max as a bid-management feature. The more accurate description is an upstream decision layer. It influences who gets matched, what copy they see, which page receives the click, and how the system learns from the outcome.
That makes the website and conversion architecture part of media buying. A vague service page is no longer just a content problem. It becomes an input problem. A soft conversion event, such as a low-intent form fill, is no longer just an analytics annoyance. It becomes training data.
The practical response is not to switch every campaign on and hope the algorithm finds its feet. Start by deciding where the model is allowed to improvise. Keep tight controls around regulated claims, high-value services, brand language, geographic boundaries, and pages that were never built to receive paid traffic.
Then separate the signals that represent business value from the signals that merely represent activity. A click is activity. A qualified opportunity, booked appointment, completed purchase, or profitable repeat customer is closer to value.
The teams that benefit will not be the ones with the most permissive settings. They'll be the ones with the clearest boundaries.
The copy problem gets expensive
AI-generated assets create a second risk. A model can produce a sentence that is technically plausible, grammatically clean, and completely wrong for the business.
It may overstate a result. It may flatten a premium service into a commodity. It may use a customer phrase that sounds fine in isolation but carries a different meaning in the category. In cannabis, healthcare, finance, and legal services, a small wording error can create compliance exposure. In any category, it can attract the wrong click at the right price.
Google's text customization tools are designed to tailor ad text to a user's query and landing page. That can improve relevance, but relevance is not the same as judgment. The question isn't whether the system can write. It obviously can. The question is whether your account has a review loop strong enough to catch what it should never say.
Give the machine a real voice system to work from. Define claims it can make, claims it cannot make, terms that signal a bad fit, and examples of language that sounds like the brand. Review search terms and generated assets as editorial material, not just performance rows.
This is where the work starts to look less like campaign maintenance and more like operating a small publishing system.
Measurement gets murkier before it gets better
Automation makes performance reports look cleaner while making causality harder to see. When matching, copy, landing pages, and bids all change together, a dashboard can tell you that results moved without telling you which decision caused the movement.
Google has continued to add planning and experimentation features around AI Max, which is a good sign. Marketers should use those controls instead of treating the campaign as a black box. But even a well-designed experiment cannot rescue a sloppy success metric.
Build a measurement layer that answers three separate questions:
- Did the system find new demand?
- Did the new demand fit the business?
- Did the resulting economics justify the margin and operational cost?
This is the same measurement problem showing up across AI search. I wrote about the broader attribution failure in the measurement crisis in AI search, where visibility metrics often arrive before reliable explanations. Paid search teams should expect a similar lag between automated reach and trustworthy incrementality.
The answer is not more rows in the report. It is a smaller set of business outcomes, tracked consistently, with enough separation between tests to make the result believable.
The new operating model
The account structure will matter less than the operating rhythm around it. A modern paid search team needs a weekly review that combines query quality, asset quality, landing-page behavior, exclusions, conversion quality, and profit.
That review should end with decisions, not observations. Which expansion should stay? Which query pattern needs a negative? Which asset should be retired? Which landing page is teaching the model the wrong lesson? Which conversion event should stop counting as a success?
Keep a change log. Record the date, the permission you changed, the business reason, and the metric you expect to move. This sounds almost boring, which is exactly why it works. When several automated systems are making adjustments, memory is not a measurement method.
There is also a strategic reason to stay disciplined. Vendor automation can become a dependency before a team notices. The more a campaign relies on one platform's interpretation of intent, the harder it becomes to compare channels or move budget without losing the model's accumulated context. That is part of the vendor lock-in problem I explored in the agentic AI trap.
Use AI Max where the system has room to learn, but keep the raw ingredients portable: clean first-party customer data, clear creative rules, owned landing pages, documented exclusions, and a measurement model that doesn't belong to the ad platform.
What deserves a human veto
Not every decision needs a person. Many decisions should not wait for one. The useful line is not automation versus manual work. It is reversible versus expensive, low-risk versus reputational, and easy to audit versus hard to reconstruct.
Let the system explore low-risk query variations when the offer is clear and the conversion signal is honest. Slow it down when a match could create a compliance claim, a bad customer experience, or a large budget leak. Require approval for new landing-page destinations, sensitive audience interpretations, and copy that makes a promise on the company's behalf.
That sounds like more work. It is less work than explaining to a client why the system spent a month buying the wrong kind of attention.
Google and Microsoft will keep widening the automation layer because it improves their ability to allocate demand. Marketers should widen their strategic view at the same time. The campaign is no longer just an ad account. It is a chain of judgments that begins with the site's promise and ends with a business outcome.
The winning question is not, “How much can we automate?” It is, “Which decisions are we willing to let a platform make in our name?”