On All-In last week, Chamath told a story about an enterprise whose AI token costs are doubling every 45 days against productivity gains he put at "maybe 5% max." Everyone heard a story about expensive AI. I heard a story about written numbers. Somewhere in that company there's a budget line someone typed into a system, and the system is executing it faithfully, and nobody owns it. Hold that thought, because the same arrangement just surfaced somewhere much closer to home: your ad account.
The angriest sentence in paid search this month came from a consultant who wrote that he had "never been more frustrated by a Google Ads update". The calmest came from Google, explaining the update that caused it: campaigns will now deliver what advertisers asked for. Both sentences are correct. That's the essay.
Here's what sits underneath both. For a decade, the target fields in a Google Ads account were aspirations. You typed a $10 Target CPA and the system treated it as a ceiling to beat; if the auction let it acquire conversions at $5, it did, and you enjoyed the difference. The number was a hint. Writing it cost nothing, so numbers got written the way hints get written: at launch, from a gut feel, copied from the last campaign, inherited through two re-orgs. That era is ending. The written number is becoming the delivered number. A setting is no longer a dashboard about the system. It's a standing instruction to it. Configuration, in an agentic system, is contract — and most advertisers are party to contracts they wrote back when writing was free.
What actually changes
The mechanics, precisely, because the backlash coverage keeps blurring them. Starting August 17, campaigns with a "Limited by budget" status running Target CPA or Target ROAS will deliver at the number in the account rather than below it. In scope: Search, Shopping, Performance Max, Demand Gen, and Travel. Display and Hotel campaigns already behave this way; App, Video reach, and Video view keep the old behavior. Google shipped an in-account tool on July 6 that flags affected campaigns and suggests adjustments, giving advertisers a six-week runway. My other site has the account-level playbook for the Bid Target Adjustment Tool; this essay is about what the change means, not what to click.
Walk through the worked example, because the mechanism is easy to misread as a price hike. Your campaign has a $10 target and has been achieving $5. That $5 wasn't Google being generous. It was headroom — the gap between what the auction required and what you authorized — and the bidder was leaving your authorization partially unspent. Starting in August, it spends it. Volume will rise, the CPA will drift toward $10 over a few weeks, and the campaign will be doing, to the dollar, exactly what its configuration says. If your unit economics actually support $10, you just bought growth. If $10 was a number someone typed in 2023 and nobody re-priced, you're about to fund the gap between what you wrote and what you meant.
Exhibit B: the unit moved too
If the enforcement were the whole story, you could file it under bidding news and set a calendar reminder. But the same month produced a quieter companion piece. On July 15, Demand Gen campaigns optimizing toward view-through conversions on Discover switched from per-click to per-impression billing. Automatic. No action required. Google's reasoning is internally sound — if you optimize toward conversions triggered by impressions, you should be billed on impressions — and the scope is narrow, and none of that is the point. The point is the pincer: the platform can enforce the number you wrote, and the platform can change the unit your number is denominated in. Your side of the contract is binding. Their side is amendable.
I wrote two weeks ago about the July Terms of Service, where the headline was that nothing changed in your account the day the contract did. This month is the sequel with the polarity flipped: things change in your account precisely because you already wrote the instructions, years ago, without knowing they'd someday be executed as written.
Aspiration debt
Every account carrying targets from the hint era is carrying a liability that doesn't appear on any report, so it needs a name: aspiration debt — the accumulated gap between the numbers you wrote and the numbers you'd write today, across every field that now binds.
Think about how those numbers actually got there. A target set at launch from a media plan that no longer exists. A ROAS floor negotiated in a QBR three margin-structures ago. A budget cap that was really a political settlement between two VPs, one of whom has left. None of this was negligence. Writing numbers was free, the system beat them anyway, and re-pricing a hint is nobody's job. I've audited accounts where I'd guess a third of the live targets couldn't be traced to a living rationale — and I'd include some accounts I ran. The debt is ordinary. What's new is that it accrues interest now.
The uncomfortable part, and the reason this isn't a story about Google being predatory: the debt was always mispriced risk. A target is supposed to be an economic claim — this is what a conversion is worth to us, priced against a real denominator. The industry got to skip that discipline for a decade because the numbers didn't bind, the same way you can skip proofreading a contract nobody enforces. Proof has a denominator; it turns out targets do too, and August is when Google starts checking the arithmetic. How fast your organization can re-price a number once it's wrong is a latency question — and for most orgs the honest current answer is "we've never re-priced one on purpose."
Google is right: you wrote that number. That's the indictment.
The steelman Google deserves
Take Google's position at full strength, because it's stronger than the backlash admits. A product manager in Mountain View can say, without spin: advertisers gave us a target; we're delivering the target; delivering better-than-target was the anomaly, and unpredictable delivery is its own complaint category. This is alignment, not betrayal. Honestly, if an engineer on my own team described a system that quietly under-executed its configuration, I'd call it a bug.
Concede all of it. The number was yours. And now follow the concession to where it actually leads, which is not where Google's framing stops. If the number was always the advertiser's claim, then the scandal isn't that Google started honoring it. The scandal is that the industry spent ten years writing claims nobody owned. Ask a marketing org today: who owns the $10? Not "who set it" — who owns it now, who re-prices it when margins move, who signs when it changes? The silence you'll get is the same silence I wrote about in Who Owns the Decision?, one level down the stack. An unowned target is an unpriced decision, and unpriced decisions get owned by whoever set the default. In this account, that's the platform.
The practitioner's objection: "narrow scope, fear content"
The other pushback comes from working media buyers, and it's fair on its face: the change touches budget-limited campaigns only; well-managed accounts with proper budgets barely qualify; the agencies loudest about this are selling audits. All true. Most of the panic content deserves the eye-roll.
But scope arguments about platform defaults have a short shelf life, and this year keeps proving it. The July ToS applied to everyone and changed nothing visible. The CPM flip applied to almost no one and changed the billing unit. Each arrives narrow, reasonable, defensible — and each moves the same direction: the account's written configuration gains authority, and the advertiser's unwritten intentions lose it. "It only applies to budget-limited campaigns" is what the first day of a default looks like. I don't know whether Google thinks of it in these terms; I doubt there's a roadmap slide titled "make configuration binding." There doesn't need to be. Systems built to execute written instructions drift toward treating every written thing as an instruction. That's not a conspiracy. It's a compiler.
A hint became a limit
One paragraph of systems talk, because the reclassification deserves to be named. Every configuration system distinguishes advisory parameters from binding ones — hints the engine may consider, and limits it must obey. Engineers mark the difference explicitly, because confusing the two categories is how outages happen. What Google did, in systems terms, is silently move a parameter from the first category to the second. Same field, same UI, new semantics. Nobody's account looks different today. Every account means something different than it did.
Which makes the in-account adjustment tool worth a second look. It's genuinely useful — it finds your exposed campaigns and proposes new targets. It is also the optimizer proposing the terms of its own constraint, from inside the reporting surface it controls. I made the long argument in the evidence-layer essay: the system that spends and the system that verifies can't share a brain. The sharpest recent instance was the kill condition — the take-back threshold can't live inside the thing it kills — and a re-priced target is the same instrument pointed the other way. Accept the tool's flags. Price the new number yourself, in your own evidence layer, against your own margins. The alternative is letting the counterparty draft the contract amendments.
The ledger
The Lab instrument this month is deliberately boring: a ledger. One row for every number in one account that the system will now execute as written — each target, each floor, each cap. Four columns: the number, its owner (a named human; "the team" is how debt hides), its denominator (what economic claim prices it — unit margin, cost per decision, LTV — not last quarter's platform-reported CPA), and its re-pricing condition: what triggers a review, and on what cadence. The re-pricing condition is the sibling of the kill condition's re-entry condition — one governs when delegation resumes, this one governs how the terms of delegation stay current.
Run it on one account and two things will happen. You'll find rows nobody can fill in, which is the audit finding. And you'll notice the ledger is just the ownership card from the last three essays, applied to the smallest unit there is — not an agent, not a decision class, a single written number. The ladder goes: own the decision class, know who set the default, write the take-back threshold. This is the rung below all of them. Own the numbers.
From the CEO chair
Strip the ad-platform specifics and the question that remains is the one Chamath's token story asks: how many written numbers is your organization party to, and who owns them? Budgets, thresholds, targets, caps — every one of them typed into some system that's getting better, month by month, at doing exactly what it's told. The enterprise with doubling token costs doesn't have an AI problem. It has an unowned-numbers problem that AI made expensive. Your ad account is just the first place the invoice arrived with line items.
For ten years the deal was that your numbers were opinions and the machine's numbers were real. That deal is over, one field at a time, and it isn't coming back — systems built to execute written instructions do not evolve toward taking hints. You can call it enforcement, like the practitioners, or alignment, like Google. Both miss the operative fact: from here on, the account does what the writing says. The system finally believes you. Make sure you do.