The Google ad tech antitrust remedies decision went public on 16 September 2026: 106 pages, released unredacted, in the case the Justice Department filed in January 2023. Judge Leonie Brinkema of the Eastern District of Virginia declined to break Google up. She ordered it to build instead.
Per the Justice Department’s own summary, four things. Integrations between AdX and Prebid, and between DFP and Prebid. Real-time bids from AdX into rival publisher ad servers. Data sharing that lets publishers “access and export their own data from DFP and AdX.” And a bar on AdWords bidding preferentially into Google’s own ad tech tools, or directly into DFP.
Then a monitor — six years of one, with access to Google’s employees, systems and source code.
Now count the nouns. Across the Justice Department’s release announcing the relief, “publishers” appears five times. The word “advertisers” appears once — inside a sentence about what Google may not do to them. The word “audit” does not appear at all. The remedy produces a document for the sell side and a prohibition aimed at the buy side. Only one of those two things is something an agency can put in front of a client.
What the court actually ordered
Start with what did not happen, because for two years the loudest prediction was a breakup. On 2 September 2026 Brinkema issued a short public order confirming Google would not have to sell AdX. As AdExchanger reported, her reasoning rested on behavioural remedies being sufficient, plus concerns about the length and uncertainty of an appeal, the harm to small businesses and publishers that depend on Google’s infrastructure, and the disruption a forced sale would cause.
The full opinion followed two weeks later. Both sides had fourteen days to flag confidential business information and propose redactions. Neither did, so the document was released exactly as written.
The Justice Department’s release lists the relief under four headings, and the second one is the paragraph this post is built on:
The court also imposed meaningful data-sharing requirements on Google. Google will be required to allow publishers to access and export their own data from DFP and AdX, which will make it easier for publishers to switch ad tech providers.
Read the rest of the list for what it does not contain. The Department wanted DFP’s full auction logic open-sourced; the court said no and ordered published technical documentation instead — how ad candidates are priced, how bid adjustments are made. AdExchanger reports the court insisting the AdX integrations be “functionally equivalent” to what DFP receives, which the outlet summarised in one line worth pinning to a wall: “Building something technically compliant but deliberately worse won’t fly.”
Two more details that matter downstream. The monitor serves six years — not the fifteen the Department asked for — and the requirements apply globally rather than only in the United States. And DV360, the buy-side tool an agency would actually open, carries no restrictions, because the court found the government never drew a sufficient connection between it and the conduct at the centre of the case.
Two sides of one architecture, and only one of them shipped
The remedy is easiest to read as a build order with two columns.
Column one: what Google must produce. Integrations that let a rival publisher ad server receive AdX bids on the same terms DFP gets them. Export of publishers’ own data. Published documentation of how the auction picks a winner. Six years of a monitor with access to source code. This is a stack of artifacts — things that exist after the order and did not exist before it, addressed to the party the court found was harmed on the sell side.
Column two: what Google must stop doing. AdWords may not bid preferentially into Google’s own ad tech tools because Google owns them. It may not bid directly into DFP. That is the advertiser-side remedy in full, and it is a prohibition. A prohibition shapes the next bid. It produces no account of the last one.
“Google will be required to allow publishers to access and export their own data from DFP and AdX.”
That asymmetry is not a scandal and it is not a drafting oversight. Courts remedy the harm they find, and the harm found here was to publishers. But notice what the court’s chosen remedy implies: the fix for an opaque auction is to hand the harmed party their own data in an exportable file, and to put someone with source-code access over the machine for six years. The court did not take Google’s word for it. It hired an evidence-producing artifact.
Nothing in the order obliges Google to tell you which of your own campaigns ran through an auction the court found unlawful, when, or with what effect on price. The remedy answers the publisher’s question. It was never asked yours.
The FTC’s Amazon complaint is the same failure, one layer down
Sixteen days before the Google decision went public, the FTC and 22 state attorneys general sued Amazon over what the agency calls a “secret ad surcharge scheme” — “deceptive and unfair practices that secretly inflated prices in its online search advertising auctions.”
The mechanism is the entire point. The complaint says Amazon told advertisers for years that it ran second-price auctions, where the winner pays “one cent more than the next highest bidder.” It alleges that in practice Amazon charged Sponsored Products advertisers their own winning bid close to 80% of the time, converting a nominal second-price auction into a first-price one.
When did the rule change? The complaint says 2019, without notice, by adding an undisclosed surcharge that Amazon referred to internally as a “soft reserve price.” One internal document described auction pricing as having “a surcharge hidden in it.” The complaint quotes the executive in charge of Amazon Ads explaining that the price paid by advertisers “isn’t set by an actual bidder” but is a “proxy 2nd price that we calculate,” and quotes another document acknowledging the use of an “invented auction participant.” In a 2024 discussion, senior executives are said to have described Amazon’s “clever non-transparent way to charge first price” as an “incredibly effective way to drive revenue.”
The escalation is documented in the complaint: the percentage of the time Sponsored Products advertisers paid their own bid rose from between 30% and 40% in 2021, to 70% in 2022, to roughly 80% in 2024.
Set that beside the Google decision and the two cases share a shape. In both, as the regulators allege, the published rule and the operating rule diverged. In both, the record of the divergence existed in exactly one place — inside the company that changed the rule — and in both, the regulator’s case was built from the defendant’s own internal documents. The FTC could quote a 2024 discussion between Amazon’s head of ads and its chief digital economist. It could not quote a single advertiser’s file, because there was nothing in it.
What actually lands on your desk
The auction question is not the one your client will ask. The version that arrives is smaller and far more answerable: what changed on this client’s account in March, who approved it, was it inside the mandate we agreed, and can you send me the file.
That question has a client-shaped answer, a jurisdiction-shaped answer and an evidence-shaped answer, and none of the three is the auction. A regulator asking about a specific decision on a named client’s account is not asking about Google’s black box. They are asking where the decision came from and who signed it off.
“Functionally equivalent” is a standard worth stealing
The court’s test for Google’s integrations — functionally equivalent, not technically compliant and deliberately worse — is the sharpest one-sentence test available for the other record in this story, the one agency-side tooling sells.
Credit where it is due first. Supermetrics ships the strictest creation default in this cohort: new campaigns are created paused so a human can review before anything goes live, documented rather than implied. That is a real control, and it is the right default. The configurable half sits next to it, in a setting whose first documented option reads, verbatim, “No human approval: Changes apply immediately.” The paused-by-default rule covers creation. The option list covers mutation of live campaigns — the half where money moves — and one of its positions produces no authorisation event to record at all.
Synter prices its governance feature by the day: “Audit Logs & Governance” at 7-day history on Solo, 90-day on Scale, unlimited plus SIEM export on Custom. A log with a seven-day life is technically present and deliberately worse than the job. A regulator’s question about a budget change does not arrive on day six.
Neither vendor is being dishonest. Both published the setting; the setting is the product. That is exactly why the court’s test applies: the question is never whether a log exists. It is whether what exists is functionally equivalent to the thing being asked of it.
| Court remedy for publishers | Court remedy for advertisers | Vendor approval setting | Ott Activity Logging | |
|---|---|---|---|---|
| Question it answers | Can a publisher take its data and leave? | Will Google bid against itself? | Did anyone approve this change? | Who changed what, when, and for which client? |
| Produces a record outside the platform | ✓ | ✗ | A gate, not a document | ✓ |
| Names the actor, human or app | ✗ | ✗ | ✗ | ✓ |
| Survives the account it describes | ✗ | ✗ | ✗ | ✓ |
What to keep, and where it lives
The remedy in the Google case is a document you can export and an evidence-producing monitor. That is the pattern to copy, at the only layer an agency controls.
- Every change, with the actor the platform reports — including writes made by an AI agent or an app. Ott’s Activity Logging stores that actor against each campaign change, which matters more each month, because the actor is increasingly a tool pointed at a client’s budget rather than a person.
- Before and after values. “The budget changed” is an observation. “The budget moved from this number to that number at this timestamp, by this actor” is the thing that answers the question.
- Client scope. Ott’s agency hierarchy runs Client → Brand → Account, so a record is filed against the engagement it belongs to — because the rulebook differs by market and the client’s obligation is not the same in all of them.
- Storage independent of the platform that accepted the change. A record that lives inside the ad account dies with the account at exactly the moment you need it most.
- An export that works before you need it. The court gave publishers an export right because switching is a real option only if you can leave with your data. Test yours while nothing is on fire.
And the approvals, kept as events rather than as a workflow gate. A gate tells you the rules were followed at the moment of the click. A ledger entry tells whoever asks, eighteen months later, that a named person authorised a specific change on a specific client.
- Can you produce every change made to one client’s account last quarter, with the actor named, in under two minutes?
- Does your record name an AI agent as an actor, or does it log the human who happened to be logged in?
- Which setting in your stack has a position that produces no evidence at all? Find the option that reads like silence.
- If the client’s ad account were restricted or suspended tomorrow, would your record of what happened in it still exist?
- Does every change carry the client and the market it was made for, so you can answer an FCA question and an ESMA question from the same file?
- Have you run the export this quarter, on a real account, before anyone demanded it?
The pattern across both cases is that the production of evidence became the remedy. A court spent 106 pages ordering a company to document how its auction works and give the harmed party their data back in a file. The FTC built a case out of internal notes that said a surcharge was hidden in the price. Neither outcome depended on the platform deciding to be transparent; both depended on somebody finally holding the record.
For a finance agency, the practical version is smaller and entirely within reach. You do not need source-code access to a search giant. You need to be able to answer, for one client, on one afternoon, with a file attached: who changed what, when, under whose authority, and what the number was before. That is the record nobody is coming to build for you, which is the whole point. The same argument showed up at the transport layer — the protocol ecosystem counts launches, not records. The governed-creation split explains why the drafting half gets built first, and the SEC’s crypto-asset proposal is the worked example of ad copy becoming evidence in a second regime.