The most useful number about AI search this year was never published. It sat under a black bar in a motion filed on September 4, and it became readable on September 17 when the publishers refiled their response. It measures how often people clicked through to a publisher from Bing Chat compared with Bing web search, domain by domain, using Microsoft's own instrumentation. For New York Times properties the drop was between 87 and 93 percent. For the Daily News domains it was 83 to 91. For Ziff Davis it ran from 51 to 94.
The filing is docket 1977-1 in MDL 25-md-3143, Southern District of New York, before Judge Stein. TechCrunch reports the figure comes from an internal Microsoft presentation dated January 2024, which means it was almost three years old by the time anyone outside the building saw it. I should say that I have read the reporting that quotes the brief and not the brief itself. The docket number is there so you can check me, and you should.
I am not interested in the number as news about publishers. I am interested in it because it is the only measurement of AI search behaviour I can name that I trust more than the vendor's own dashboard, and the reason has nothing to do with the statistics.
Where a number came from matters more than the number
Every figure in a media review has an issuer. The conversion count belongs to the platform that sold the click. The lift study was designed by the platform that wants to be lifted, which picked the question, the arms and the window. The agent's report on its own decisions was written by whoever operates the agent. I have spent the last several essays on one version or another of this, from the seller's counterfactual to the audit you buy from the auditee, and the pattern underneath is plain enough to state in one line. In each case the party being measured decided when you would see the number, what it would cover, and how it would be described.
Microsoft decided none of those three here. Nobody at Microsoft chose to release a figure showing its product cutting clicks to the largest news publishers by nine tenths. A schedule for briefing a summary judgment motion did, and a dispute about redaction, and the plaintiffs' lawyers choosing which of Microsoft's own slides to quote. The description is theirs, not Microsoft's, and you should read it knowing whose it is. But the instrument is Microsoft's and so is the data, which closes off the one thing a defendant would otherwise say, that someone hostile reconstructed it.
That is a narrow property and I want to be careful with it. It is not accuracy. It is independence of issuance: the subject did not control the timing, the scope, or the framing.
The objection I would make myself
Litigation selects for disaster. The numbers that reach a court are the ones somebody had a reason to fight over, which makes compelled evidence the least representative sample of a company's data you could design. The three publishers in that filing are not the median website either, they are the ones with lawyers and a lawsuit. And Bing Chat is not Google's AI Overviews; nothing in that filing says what happens to your clicks on Google.
All fair. If you take one thing from the Microsoft figure, do not take 87 to 93 percent and plan a budget around it. The claim I am making is smaller, and I think it survives. Evidence a seller was made to produce has a different structure from evidence a seller chose to produce, and almost nothing in a marketing organisation's evidence stack is of the first kind.
Three tiers, and where your numbers sit
Sort what you rely on by who decided it should exist. The first tier is volunteered: platform reporting, vendor logs, seller-run lift, the optimiser's account of itself. The seller chose to produce it, which also means the seller can choose to stop, to change the definition, or to produce a different one. I wrote about what happens then in The Seller's Unit.
The second tier is commissioned. You specified the question and you own the denominator: a geo holdout, a first-party control, a matched-market test you ran on your own spend. A measurement lead at a big advertiser will tell me, correctly, that they have exactly this and that it is the best evidence in the building. I agree. It is independent of the seller. It is not independent of you, which is a smaller problem, but you also only run it on the questions you thought to ask.
The third tier is compelled: a court, a regulator, or a contractual right that obliges the seller to hand over a number it did not choose to hand over. This is the only tier that exists without anyone at the seller deciding to produce it. Count what your organisation holds in each tier and the shape is unflattering. Nearly everything is volunteered, a little is commissioned, and nothing is compelled.
That is different from the argument in Who Is Allowed to Measure?, which was about access: whether you may hold causal evidence at all. This one comes after it. Even when you hold the evidence, someone decided that it would exist, and the question is whether that someone was the party it describes.
A number the seller chose to publish is a number the seller can choose to stop publishing.
What Google volunteers, and what it leaves out
Here is the counter-evidence, and it cuts against me a little. Google's own help page on ads in AI Overviews says, as of this week: "Google Ads currently doesn't offer segmented reporting when ads show within Search AI Overviews." Then on October 2, Search Engine Roundtable reported that Google is testing URL tracking parameters on links inside AI Overviews and AI Mode, so that site owners can tell a click from an AI Overview apart from a click on a regular result.
If that ships, it is a seller volunteering a number that points toward the buyer's interest, and I would be glad to be partly wrong about how reluctant the sellers are. Notice what it is, though. It is a test, and it is aimed at site owners rather than advertisers. And it counts clicks. The Microsoft figure needed two sides, the click rate on one surface against the click rate on another, and a rate has an impression underneath it. A click count with no record of how often the surface was shown and nobody clicked is what I called in Proof Is the New Flex a numerator with nothing under it. Google may hand over the click first. The part that would show what an AI surface is doing to your demand is the part still inside.
So what does a search buyer do on Monday
Nobody sent you Microsoft's slide, and no court is about to subpoena your account rep. Fair. The Monday move is not to litigate. It is to ask of every number you rely on a question you can answer without anyone's permission.
When I wrote Who Counts the Conversion? I proposed a row per seller on the Unit Ledger: the unit as that seller defines it, the attribution window, whether post-view credit sits inside or outside the total, who reviewed the definition and when. I had not yet put that row anywhere you could copy it, so here it is, with one column added.
| Seller and number | Unit as the seller defines it | Window and credit rule | Reviewed by, when | What would force it out if it were bad, and who holds that lever |
|---|---|---|---|---|
| Platform A, conversions | A conversion the platform's own model credits | Thirty days, view-through included | Nobody, never | Nothing. No data right in the contract. |
| Platform B, AI-surface clicks | A click on a link inside the AI surface | Not stated | Nobody, never | Nothing, and no impression count behind it. |
The lever column takes four kinds of answer. A regulator that can require disclosure. A court. A third-party panel whose data does not pass through the seller. Or a contract right you wrote yourself, which is the one a buyer can actually reach today, and which is the same family as the comparison-arm clause in the audit essay: a right to something the seller would not otherwise have produced. If the cell is empty, you have not found a problem with the seller. You have found a number that exists only for as long as someone else decides it should, sitting in a deck that a budget was approved on.
An argument cannot be scored later, so here is one with a date. As of October 3, 2026, Google's help page says Google Ads does not offer segmented reporting for ads that show within AI Overviews.
On May 31, 2027, check whether Google Ads gives every advertiser, in standard reporting and without an account representative involved, impressions and clicks for ads in AI-generated surfaces separately from classic results. I claim it will not. If the help page has dropped that sentence and a segment appears in the standard interface or the API for all accounts, I am wrong. A tracking parameter on organic links for site owners does not count, because it is not advertiser reporting, and a click count without impressions does not count either. I have not checked Microsoft Advertising's position and make no claim about it.
Who decides which numbers get a way out
Legal owns the contract. Finance owns the budget. Marketing owns the results. The decision about which seller numbers must have a compelled route, whether by right, by regulation or by an independent panel, is owned by none of them, and a decision class with no owner defaults to whoever is writing the terms. As ever, that is the default owner, and in this case it is the seller. The Decision Ownership hub carries the vocabulary for the rest.
I started this expecting to argue that courts are the real auditors of advertising, and I no longer believe that. They are slow, they are adversarial, and the one number I leaned on is three years old and about somebody else's product. What I do believe is narrower. A budget that rests only on figures their issuers chose to publish is priced by the issuers, and the bill arrives in whichever quarter a judge, a regulator or a competitor's lawyer publishes the number the dashboard did not. You can pay that bill early, by writing down which of your numbers have no way out and asking for one while the terms are open, or you can find out the way those three publishers did.