Nothing changed in your Google Ads account on July 2. The alarmists are wrong about that, and it matters that they're wrong, because the truth is quieter and worse. Google didn't take ownership of your decisions this month. It disclosed that you never had it.
On July 1, every Google Ads account in the world was bound to a new Terms of Service, the first substantive rewrite since April 2018. There was no login prompt. There was no acceptance step. If you did nothing, and nearly everyone did nothing, the new terms govern your account anyway. Somewhere in the revision is the sentence the trade press has spent a week circling: advertisers now authorize Google to serve ads "including through the use of automated program features to format, select, or generate targets, ads, or destinations on Customer's behalf." And a few clauses away, its quiet twin: the advertiser remains responsible for reviewing, approving, editing, or removing campaigns and ad assets, including the ones Google's automation generated.
The coverage so far comes in two flavors. The tactical flavor tells you what to check in your account settings. The alarmed flavor tells you Google just seized control of your advertising. Both are answering the wrong question, so let me start by conceding everything the calm practitioner will tell you, because the calm practitioner is right: the features this clause describes already existed. AI Max's search-term matching, text customization, and final-URL expansion were live before July. The opt-outs that existed in June still exist in July. Your campaigns ran on July 2 exactly as they ran on June 30. Anyone selling you panic is selling you a misreading of a contract they skimmed.
Now here's what the calm practitioner walks past. A contract that changes nothing about how the system behaves, and still took eight years to rewrite, is doing something other than changing behavior. It's writing down what the behavior already was. The accurate description of July 1 isn't a taking. It's a disclosure.
A default is a decision made for whoever hasn't priced the alternative
The key word in the new language isn't "generate." It's the word doing the structural work underneath the whole revision: default. Automation moved from an optional program feature, something you enrolled in, to the default contractual state of the account, something that simply governs wherever you haven't said otherwise. And a default is a specific kind of thing. A default is not a feature. A default is a policy about absence. It answers one question: what happens in every corner of this system where you never made a decision?
I made the argument two weeks ago that you own a decision class, not an agent. A decision class without an owner, an objective, a denominator, a latency budget, and a kill condition isn't owned at all, and an unowned decision is an unpriced decision: silently expensive, silently slow, invisible because it passes every audit. What I could not have told you then is that Google's lawyers were about to agree with me in writing. Because the July terms are, read at the right altitude, a formal answer to the question that essay left hanging. What happens to an unpriced decision? It doesn't stay unowned. It gets owned by whoever set the default.
That's the inversion the week's coverage missed, and it's why the outrage framing fails on contact: you can't lose ownership you never exercised. Every decision class you have priced and claimed (named, instrumented, kill-switched) is exactly as yours today as it was in June. The contract didn't touch it. What the contract formalized is the rest of the account: the audience-expansion rule nobody ever articulated an objective for, the creative rotation nobody ever priced, the query-matching behavior nobody ever put a latency budget on. Those decision classes had a de facto owner long before July: the platform's optimization layer, running them on the platform's objective. The Terms of Service just gave that arrangement a countersignature.
A default is a decision made for whoever hasn't priced the alternative. An unpriced decision doesn't stay unowned. It gets owned by whoever set the default.
Say the formal version once, because it removes the sentiment from the argument: a decision class with no explicit policy inherits the environment's default policy, and the environment's default policy optimizes the environment's objective. "We haven't decided how to handle that yet" has never been a neutral state. It's a policy assignment to a counterparty, on the counterparty's terms. The only thing new on July 1 is that the assignment now has your signature on it, supplied on your behalf, by default.
Every vendor shifts liability. This clause does something else.
The second objection worth taking seriously comes from anyone who's spent time around procurement: liability-shifting boilerplate is every vendor contract ever written. AWS disclaims; Salesforce disclaims; your payroll provider disclaims. Marketers discovering contract law is not a thesis.
Conceded, with one distinction that carries the whole essay. AWS does not generate the workload it holds you liable for. Your cloud vendor disclaims responsibility for what you build on its infrastructure; it doesn't build things on your behalf and then hand you the review duty. The July terms pair three things no ordinary vendor contract pairs: generative authority (the platform may format, select, or generate the ads, targets, and destinations), retained liability (whatever runs is legally yours), and a review duty (you are contractually the approver of record for output you didn't produce). That combination has a name in any other industry: it makes you the quality-assurance function for a factory you don't operate, running at a speed you don't control.
So let's do what the coverage hasn't and put numbers on the review duty. Rough numbers, because rough is all this needs. Take a mid-size account where AI Max and Performance Max generate or revise something like 300 reviewable artifacts a week: creative variants, expanded destinations, new query mappings. A competent reviewer at 90 seconds per artifact is seven and a half hours of qualified labor a week, call it $560 at a loaded rate, or about $1.90 per reviewed artifact. Manageable. Now notice what you reviewed. The artifacts are not the decisions. The decisions (which query matched, which variant served, which bid cleared) happened at serving time, millions of times, inside the auction. Review every artifact and you've inspected roughly 0.003 percent of the decision volume the contract makes you responsible for. The review duty is real, budgetable, and dischargeable. And discharging it perfectly still leaves the decisions themselves unexamined. A duty defined over artifacts, owed against decisions, is not a control. It's a ritual.
Which is why "just review the output" is the wrong response, the same way "just assign an owner to the agent" was the wrong response last month. You cannot review your way to owning a decision class that runs at machine speed. I made that argument about management; it holds for contracts. The only version of the review duty that isn't fiction is the instrumented one: own the class's denominator, watch its cost per decision and its latency move, and hold a kill condition. Everything else is signing the QA log of someone else's factory.
The ownership card, reread against the contract
In the last essay I put a five-line instrument on the table: the decision-class ownership card. Owner, objective, denominator, latency budget, kill condition. Here's the exercise the July terms make newly concrete: run the card against the contract. For each line you've filled in, your answer stands; the terms defer to explicit decisions. For each line you've left blank, the contract now supplies the answer, and it's worth reading what the answers are.
Read the right-hand column as a whole and the July terms stop being a legal story. They're an org chart. That column is a complete, functioning ownership structure for your media spend (owner, objective, cadence, escalation path) installed wherever you didn't build one. It works. It's just not working for you.
The prediction and the paperwork
A year of this site's readers know I put a date on this. The Sealed Auction argued that Google's search machinery would go effectively opaque in Q3 2026: that AI Max, query-report dilution, and prose-input campaign briefs shared one prediction layer, and that the auction was sealing. Q3 began on July 1. Since the Lab's standard is proof, not takes, here is the honest audit.
What landed: AI Max is moving out of beta and Dynamic Search Ads are being upgraded into it, a forced migration, on schedule. Practitioner reports of query-matching drift and irrelevant-match volume through AI Max are accumulating, along the exact prose-input line the essay drew. Google's own headline claim of about 7% more conversions at similar CPA or ROAS comes with a conditionality worth reading twice: it compares the full feature suite against search-term matching alone, which is to say it measures how much better surrender-in-full performs than surrender-in-part. What I got wrong, or at least framed wrong: I expected the sealing to arrive as a product moment: a visible switch, an interface change, a launch. It arrived as paperwork. No single Q3 event sealed the auction; instead, the contractual default did quietly what I expected a feature flag to do loudly. The auction isn't sealing by announcement. It's sealing by absence, one unclaimed decision class at a time. I'd score the mechanism right, the theater wrong, and the timing uncomfortably precise.
The default owners are multiplying
If this were only a Google story, it would be worth a settings check and a shrug. It isn't. The same week the Ads terms took effect, Salesforce made its Agentforce commerce agents generally available: a Shopper Agent, a Buyer Agent, and a Merchant Agent, with native integration into ChatGPT and Google Search on the way. Meta's Advantage+ has run on the same logic for years. Every platform in the agentic stack shipping this year ships the same way: agents on, by default, optimizing the platform's objective wherever the customer hasn't specified their own. The vendors are not being sinister; they're being rational. A default is the cheapest possible acquisition of an unpriced decision class, and the market is currently full of them.
Which means the July terms are best read not as a Google policy but as a preview of the standard agentic-era contract: the platform supplies the decisions, the customer supplies the liability, and the boundary between them is drawn precisely, automatically, and in the platform's favor, along the line of whatever the customer never priced. The optimizer already grades its own work. Now it drafts its own authorization. The response to that isn't to opt out of automation, which is neither possible nor smart. The response is to move the line: every decision class you price and claim is a decision class the default can no longer reach.
What this looks like from the CEO chair
Strip the contract language and the question for a leadership team is one sentence: as of July 1, which of our marketing decisions do we own, and which does our vendor own by default? Most organizations cannot produce that list, and the absence of the list is the finding. Producing it isn't a legal exercise; legal already lost this round by design. It's the same instrument as last month, with a deadline attached. Name the decision classes your systems run. Mark which side of the default each one sits on. For the ones you want back, fill in the card: owner, objective, denominator, latency budget, kill condition. The P&L translation is direct: every line that stays blank is a cost structure someone else prices and a margin conversation you enter without a number of your own.
The terms took effect July 1 whether you read them or not. That's what a default means, and resenting it is as useful as resenting weather. The move that matters is the one you can still make and can't backfill: before the quarter closes, make the list. Every decision class in your account sits on one side of the default or the other; the contract has already assigned the blanks, and it will keep assigning them, quarter after quarter, for as long as they stay blank. You will not get a login prompt asking whether you'd like to own your decisions. You just got the opposite: a contract that assumes you won't, countersigned on your behalf, effective the first of the month. The platforms have told you, in writing, exactly what happens to every decision you leave unpriced. It would be strange to need to be told twice.