How a 93% from an OpenAI lawsuit became a marketing stat

One number, four days, three professional languages

On page 68 of the brief filed September 17 by the New York Times and the other publisher plaintiffs, the sentence is dry and technical: the click-through rate between Bing Chat and standard Bing search falls by 87 to 93% for Times sites. Four days later, the same number is the headline on an ad-industry outlet ("Copilot cuts publisher clicks by up to 94%"), and a section header on an acquisition-marketing blog ("the traffic collapse, in cold numbers").

No one lied, and no one misread anything. The number simply crossed three professions, and each one rewrote it in its own language. That is the trip we followed here, because it is more instructive than the number itself.

Step one: the ambiguity started in the filing

The plaintiffs' brief is document 1977-1 in case 25-md-3143, before Judge Sidney H. Stein. Its footnote 10 defines click-through rate as the percentage of clicks to a URL relative to impressions, meaning how often that URL showed up in a results page or a chatbot's answer. A rate per impression, in other words: not a visitor count, and not a change over time.

Except the document states its own result twice, and not the same way each time. Its introduction, on page 2, merges the Times and the Daily News titles into a single 83 to 93% range. Its body, on page 68, splits them apart: 87-93 for the Times, 83-91 for the Daily News, 51-94 for Ziff Davis. The same exhibit backs both passages.

Four newsrooms published the number within twenty-four hours, in four different versions. We covered it last Saturday: none of them got it wrong, each one had simply opened the document to a different page. The ambiguity was not introduced by the press. It was already in the source.

Step two: the distortion happens in the headline

This is the most interesting step, because it has nothing to do with how careful the writer was.

A body paragraph can carry "click-through rate per impression, compared across two search surfaces." It has the room, and its reader has already agreed to keep reading.

A headline cannot. It has to fit on one line, be understood with zero context, and reach for a quantity everyone already handles: clicks, traffic, lost readers. Not a ratio between two measurements.

The clearest case is PPC Land, whose body copy states exactly what the brief says: Microsoft compared Bing's click-through rates to Bing Chat's, domain by domain. Its headline, though, talks about clicks cut "by up to 94%," and that 94 is the ceiling of the widest range in the whole case, the one for Ziff Davis, whose floor sits at 51%. A headline grabs one number, not three ranges.

A rate compared between two storefronts is not a stock that is falling. A store's customers can buy less often than the ones next door without the store emptying out. But "less often than next door" does not fit in a headline, and "the store emptied out" does.

We are not well placed to lecture anyone here: we published on this same number on Saturday, and our own headline got rewritten several times before it ran. The constraint applies to us just as much as to anyone else, and that is what makes it worth studying. It does not punish carelessness. It punishes limited space.

Step three: the meaning shifts with the audience

In a legal brief, "CTR" is a term defined in a footnote, because the outcome of the case hinges on what it measures. In the tech press, it becomes a fact to summarize. In the SEO and marketing outlets, the last stop on this trip, "clicks" means traffic, because traffic is what their readers are buying. The word did not drift; it got translated into a language where it does not mean the same thing.

It does not even take a misquote. At Business Model Analyst, the number is cited accurately, then dropped into a list alongside two traffic metrics, and labeled an "industry average." It is the company it keeps that shifts the meaning: three figures sitting in the same paragraph end up measuring the same thing in a reader's mind.

And it is at the very end of the chain that the filing's two versions collide to produce an actual mistake. A piece written for software founders states that rates fell "83 to 93% for New York Times sites and 83 to 91% for Daily News sites." The 83-93 comes from the introduction, where it covers both groups; the 83-91 comes from the body, where it only covers the second one. The two versions got merged, and the Times inherits a range that was never its own.

That is the only measurable error in the entire trip, and it took four days and three professions to manufacture it out of a document that already contradicted itself.

What no one can verify, us included

The data cited in support, referenced under exhibit numbers SF1530 through 1537 and SF1539, are not attached to either brief: both sides refer to them only by number. No table, no chart, no method. They remain under seal.

Microsoft builds its main argument on that gap: the Times' and Daily News' experts never actually analyzed web traffic, even though their damages theory depends on it. The plaintiffs do not dispute that; they answer in legal terms, on page 68, arguing that proving harm empirically is not required, common sense is enough.

What we are reading, then, is an argument, not a measurement. That is a feature of the case file, not something we failed to dig up.

The next number due for the same trip

Fourteen pages earlier, the same brief slips in another one: 87.78% of ChatGPT users allegedly visit no external site during a search session, versus 26.91% for Google. Two decimal places, the kind of precision that inspires confidence. The study is not named, and the exhibit that contains it is sealed too.

It has already left the brief: one of the pieces we read cites it, with no more detail than the original. It has everything it needs to end up in a headline.

One caveat, and it matters: we found no French-language pickup of this second-generation figure. The trip described here played out in the English-language trade press, and we are not going to assume it happened the same way in French. That clicks to publishers keep thinning is well documented elsewhere, including by us, on Google's AI Overviews. This particular number simply is not the one measuring it.

A number that crosses four professions in four days is not a chain of errors. It is a chain of translations, each one faithful to the language it lands in, none of them mandated to preserve the original's meaning intact.