She could see the number. That was the whole problem. A causal read on what the media had actually caused, presented in a quarterly review by someone who did not work for her company, on a slide she was not going to be handed. The study looked well built. The result was positive. And when she asked whether her analysts could pull that figure into the warehouse next to spend and the mix model, the answer came back as a process rather than a yes: she would need to be added to a list.
Nobody in that room did anything wrong. The person presenting was proud of the work and had every reason to be. The request was reasonable, the refusal was not really a refusal, and everyone moved on to the next slide. I have watched some version of this for years and filed it under procurement friction, which is where it belongs right up until the moment it stops being friction and starts being architecture.
The measurement argument in this industry is almost entirely about quality. Whether lift beats last click. Whether a mix model can be trusted. Whether an agent's logs prove anything at all. Those are real questions and I have spent years inside them. They are also not the question that changed this month. The change is about access: evidence regarding a system is becoming something you are granted rather than something you hold. A grant has an issuer. Increasingly the issuer is the party being measured.
The objection that nearly kills this, first
Before any of that goes further, the obvious version of it is wrong, and I would rather break it myself than have a measurement lead break it for me in the comments.
Nothing was taken away from anyone. You never self-served a conversion lift study. You did not do it in 2019 and you cannot do it now. Lift studies have always been set up inside the interface or arranged through an account team, because they involve randomised control groups, statistical minimums and a holdout that somebody has to agree to. Anyone writing this up as a platform snatching back the evidence is describing a thing that did not happen.
What is more, access has been widening rather than narrowing. Google cut the minimum budget for incrementality experiments to five thousand dollars in November 2025, down from thresholds that had been closer to six figures, and surfaced conversion lift results at product and brand level in the interface. That is a real democratisation and it deserves to be said out loud.
The permission did not change. The pipe changed.
Here is the distinction the whole essay rests on, and if it does not hold, nothing after this paragraph is worth your time. A lift number in a quarterly deck can only be a memory. The same number in an API is a standing input to an automated decision. One informs a human once, in a room, and then decays. The other can be joined against spend, reconciled against a mix model, fed to a bidding rule, read by an agent at three in the morning and re-read every night after that. Same figure. Completely different object.
What actually shipped
On August 19, 2026, Google released version 25.1 of the Google Ads API. It added read-only access to conversion lift and brand lift study data: a new LiftMeasurementConfig resource representing a study, a LiftMeasurementFlight resource carrying the measured start and end dates, twenty-four new Conversion Lift metrics alongside winner score metrics for statistical analysis, and five brand lift dimension resources covering age range, campaign, device, gender and video. Google's own framing is that v25.1 is a drop-in upgrade for v25. No breaking changes, no removed resources, no forced rewrites.
Read-only is the operative constraint. Nothing in the release lets anyone create, modify or end a study programmatically. The study still gets set up the old way. What changed is that its parameters and results can be pulled into systems you control rather than transcribed off a dashboard.
And according to the announcement, both lift features are available only to allowlisted accounts, with advertisers directed to their account representative to request access. I want to be precise here, because precision is the only thing that makes an argument like this survivable: that allowlist condition is stated in the developer blog announcement and in the trade coverage of it, not in the API release notes page, which describes the new resources without mentioning access tiering at all. Two of the six features in the release sit behind that gate. Google has not published how many accounts hold it.
So I will make the prediction rather than wait to be corrected with it. This will go generally available. Allowlists usually do. And when it does, the interesting part will not have been resolved, because general availability changes who stands inside the gate without changing that there is one, that its position is set by the seller, or that it now sits on the machine-readable layer where decisions actually get made. The gate is not the news. The gate's location is the news.
The account team is right, and it does not matter
Steelman the other side properly, because the other side is not being cynical. Allowlists exist for load-bearing reasons. Lift measurement carries statistical validity floors, and a study run too short produces a bad answer confidently: Google's own guidance recommends a minimum of fourteen days and notes drops of up to seventeen percent in absolute lift for businesses with longer conversion lags when studies run shorter than that. Demographic brand-lift dimensions carry real disclosure risk. API surfaces carry query cost. A staged rollout to accounts with account teams who can explain the caveats is, frankly, the responsible way to ship a statistical product.
All of that is true. None of it changes the consequence. The reason for a gate does not alter the effect of a gate, and the effect is that two companies buying the same product, running the same stack, now hold different evidence about it, with the difference decided by the strength of a commercial relationship. That is not a scandal. It is a structure, and structures outlive the intentions that built them.
There is a version of this the trades noted and then stepped around: lift studies measure the causal contribution of a platform's inventory using that platform's own randomised control groups. Whether independent verification of those control groups gets easier as a result is, as one write-up put it, a separate question that the release does not address. It is separate. It is also the only question that matters, and the site has a name for a number generated by the party with an interest in it. It is a seller's counterfactual: possibly accurate, structurally a quote.
Your agents are the second allowlist, and it is set to allow-all
Now put the platform down entirely, because the sharper instance of this is inside your own building and it has no gate at all.
A marketing organisation delegating work to agents is accumulating an evidence layer whose two largest inputs are the platform's causal measurement and its own agents' self-reports. One is provisioned by a vendor. The other is provisioned by the thing being evaluated, which is worse, because at least the vendor's version went through a statistician.
The failure mode here is not error. It is confident completion. An agent that attaches the wrong file and reports success is not malfunctioning in any way its logs will show you; it is doing what it was asked and grading its own work. One practitioner logging north of eleven thousand agent runs this year found that the ones that lied looked the most finished, which is the detail that should keep people up at night. Meanwhile something like eighty-two percent of enterprises report discovering agents on their networks nobody knew were there, and the EU's high-risk obligations became fully enforceable on August 2, 2026, which has quietly turned audit-trail capability into a procurement gate in regulated categories.
Both halves are the same shape. The actor and the reporter are the same process. Every other engineering discipline names this and forbids it. You do not let a component under test emit its own pass signal. The prohibition has nothing to do with anyone lying. It is that a report issued by the interested party carries no independent variance, so you cannot distinguish a result that is accurate from a result that is accurate by construction. Both look identical on the slide.
This is a different claim from the one I made about the optimizer owning the evidence layer, and the distinction is worth one sentence: that essay asked whether the proof is real, and this one asks whether you are permitted to hold it, which turns out to be the prior question.
Twenty-eight essays on this site argue, and an argument cannot be scored in retrospect. So, with a date on it:
On March 31, 2027, open the Google Ads API release notes and check one thing. Whether Conversion Lift metrics have gone generally available to every account without a representative in the loop. I claim they will not have.
Ninety seconds to verify, and it can go against me in public, which is the point. If lift is fully self-serve by then, my read on how measurement access gets provisioned was wrong, and I would rather be wrong on a date than vague enough to have been neither. A smaller one already resolved: on July 17 I wrote that stale targets were about to start binding, and on August 17 budget-limited campaigns began delivering to the target you typed rather than the better number you had been getting.
What an evidence layer you actually own costs
The incrementality skeptic has the best answer available and has had it the whole time: run your own geo holdouts and stop asking. That is right, and I am not going to pretend it is free, because pretending it is free is why it does not happen.
A geo holdout works by deliberately withholding spend from matched markets. The price of the answer is the profit you gave up in the holdout. That is a real number, a finance partner will ask about it, and "we're testing" is not a sufficient reply in the fourth consecutive quarter of it. Which makes it exactly a priced exception: a worse outcome accepted on purpose, on stated terms, for something you decided in advance was worth it. There is a calculator here for writing that number down before the fact rather than discovering it afterwards and relabelling it a test budget.
The rest is unglamorous inventory. Which numbers in your stack would still exist in twelve months if nobody at the vendor took your call? A warehouse copy you control persists. A test you can rerun persists. A mix model you run yourself persists, with the caveat that it is only as independent as its inputs, which is the whole problem with modelled conversions arriving upstream of it. A dashboard does not persist. A slide does not persist. Access does not persist, by definition, because it is granted.
None of this argues for distrusting platform measurement, and I want to be careful not to be read that way. Platform lift studies are usually well constructed and frequently better instrumented than anything an in-house team could stand up alone. Use them. Believe them, mostly. The ask is narrower and much duller: hold one number that did not come from the party being measured, and know what it cost you to have it.
That number is also the only thing that makes ownership mean anything. You own a decision class, not an agent, and owning a decision class you cannot independently evaluate is a title without an asset. The org chart says the VP owns paid media performance. The evidence layer says the evaluation of paid media performance is issued by paid media. Both of those cannot be true, and the org chart is the one that will lose.
Provisioned evidence is fine until the day it is not, and the day it is not tends to arrive alongside a contract renegotiation or a rep who moved on. I do not think anyone is acting in bad faith here, and I have been wrong before about how fast these things consolidate. But a marketing organisation that cannot name one number it produced itself has not delegated its measurement. It has delegated its ability to disagree. You do not own a number you can be denied. Decide what an independent answer is worth to you before somebody else tells you what it costs.