ENTRY 027 · ANSWER-ENGINES · By Answer Engineered Research
94%vs51%
Three Click-Through Ranges in One Filing. The Widest Ran in Neither Story.
A 92-page federal filing gives three Microsoft click-through ranges. TechCrunch printed one number, TheWrap printed two ranges, and 51% to 94% ran nowhere.
What the document is
| Field | Value |
|---|---|
| Document | News Plaintiffs’ Combined Summary Judgment Brief, Document 1587-1 |
| Case | The New York Times Company v. Microsoft Corporation, 1:23-cv-11195-SHS-OTW (S.D.N.Y.), part of In re: OpenAI, Inc., Copyright Infringement Litigation, MDL No. 25-md-3143 (SHS)(OTW) |
| Filed | 09/17/26, per every page footer inside it |
| Length | 92 pages |
| Copy read | CourtListener RECAP, HTTP 200, 12,369,196 bytes, extracted with pdftotext |
Page references are given twice where they differ: the brief’s printed page number, then the docket stamp, which counts the cover sheets.
One caveat governs everything below. This is a brief, one side’s argument for why the case should be decided without a trial, and nothing in it has been ruled on. TheWrap states the point in its own coverage: “The court has not determined that either company infringed the publishers’ copyrights.”
Three ranges, one sentence
At page 14 (“Page 23 of 92”), the brief writes:
Microsoft’s representative CTR data shows that the overall click-through rate reduction between Bing Chat and Bing Web Search was 87% to 93% for The Times’s websites, 83% to 91% for DNP’s websites, and 51% to 94% for ZD’s websites.
The sentence carries the reference SF1536. We do not know what “SF” stands for. The brief never defines it in the text we extracted, and guessing is how a paraphrase becomes a claim.
The shorthand is defined at page 4 (“Page 13 of 92”). “DNP” is the Daily News Plaintiffs, eight newspapers including the New York Daily News and the Chicago Tribune. “ZD” is Ziff Davis, which operates IGN, CNET, ZDNET, PCMag and Mashable.
Here is the full set, and where each range landed:
| Plaintiff group, in the brief’s shorthand | Click-through reduction, Bing Chat against Bing Web Search | In TechCrunch | In TheWrap |
|---|---|---|---|
| The Times | 87% to 93% | Not as a range: “as much as 93%“ | Yes: “87% to 93% lower” |
| DNP, the Daily News Plaintiffs | 83% to 91% | Not mentioned | Yes: “83% to 91% lower” |
| ZD, Ziff Davis | 51% to 94% | Not mentioned | Not mentioned |
The ZD range is the one that changes the shape of the finding. Its ceiling, 94%, is the highest of the six figures in that sentence; its floor, 51%, is by far the lowest. A range running from 51% to 94% is a different claim from “as much as 93%”, and it is the range nobody printed.
The compression starts inside the document
Before any journalist touched this filing, the filing had already done the same thing to itself. At page 1 (“Page 9 of 92”) the Introduction summarises the page-14 sentence:
Defendants’ own data show this “doom loop” in motion, with Microsoft recording 83-93% drops in click-through rates for The Times and DNP’s domains, and 51% to 94% for ZD’s domains
The Introduction blends the Times range (87% to 93%) and the DNP range (83% to 91%) into one span, “83-93%”, and keeps ZD separate. Two ranges become one, and the group that owns each figure stops being visible.
That is worth sitting with before anyone blames a newsroom. The lawyers had every incentive to be precise, and they compressed their own numbers on their own page 1. Compression is what happens to numbers each time they move up a level of summary, including the level where they originate.
Two outlets, two different subsets
TechCrunch, by Rebecca Bellan: “Microsoft’s own data shows its Copilot ‘answer engine’ caused click-through rates for The New York Times’ domain to drop as much as 93% compared to traditional Bing search.” That is the ceiling of one of three ranges, accurate about that ceiling and silent about its floor and about the other two groups.
TheWrap, by A.J. Katz: “Click-through rates from Bing Chat were 87% to 93% lower for the Times’ websites and 83% to 91% lower for the Daily News plaintiffs’ sites when compared with traditional Bing searches, according to the motion.” Two of three ranges, both matching the brief exactly, and attributed to “the motion” rather than stated as established fact.
Neither story mentions ZD.
The same selection shows up on a second number. At page 9 (“Page 18 of 92”), citing SF844, the brief quotes an OpenAI Slack list of the “Top 100 domains by document count”, “which included 2,182,079 articles from the ‘chicagotribune.com’ domain and 2,064,805 articles from the ‘nytimes.com’ domain”. TechCrunch reports the nytimes.com figure, rounded to “more than 2 million”. The larger number belongs to the Chicago Tribune, a named plaintiff here through the DNP group.
The quote that lost its “perhaps”
The brief’s opening sentence, page 1 (“Page 9 of 92”):
This case is about, as Microsoft’s Director of Applied Science put it, ‘an astonishing theft of unprecedented proportions’; SF1437, perhaps the ‘largest theft of labor in human history.’ SF1652.
Two things in that sentence do not survive into one of the write-ups: the word “perhaps”, and the fact that the two phrases carry two different reference numbers.
TheWrap keeps both the phrases and the hedge: “describing the alleged copying as ‘an astonishing theft of unprecedented proportions’ and perhaps the ‘largest theft of labor in human history.’”
TechCrunch’s sentence reads: “In a January 2023 internal memo, Hecht called it ‘an astonishing theft of unprecedented proportions’ and ‘the largest theft of labor in human history.’” That is more specific than the brief in three ways at once. It drops “perhaps”, it places both phrases in one document, and it dates that document.
None of which makes the date wrong. TechCrunch may have sourcing we do not, and it says in its own body text that the exhibits “remain sealed” and that “The quotes below are presented without their original context.” What we can check is the document TechCrunch links to, and in it the strings “January 2023” and “January 2024” appear nowhere across all 92 pages. The brief’s own timing language carries no date either: at page 52 (“Page 61 of 92”) it attributes a different sentence, under the reference SF1672, to “Microsoft executive Dr. Brent Hecht, in a document he wrote soon after this lawsuit was filed”.
We are not resolving what the “perhaps” is doing. It may hedge the attribution, or it may hedge the characterisation while the attribution is solid. Both readings fit the visible text and we do not have the exhibit.
The reference numbers matter for the same reason. The brief cites the “doom loop” sentence at SF1672, “an astonishing theft of unprecedented proportions” at SF1437, and “largest theft of labor in human history” at SF1652. Three quotes, three references, and it introduces the “doom loop” line as coming from “A Microsoft document” rather than from a named person. Whether any of them share an exhibit is not something the visible text says, so we do not say it.
A document stamped REDACTED, reported as “unredacted”
TechCrunch’s headline says “new unredacted filings reveal”. TheWrap’s says “in Unsealed Docs”. Those are two different claims, and only one is clearly supported by the copy on the public docket.
First, the cover page. Our extraction renders the words directly under the brief’s title as the garbled string 5('$&7('. Substituting 5 to R, ( to E, ’ to D, $ to A, & to C and 7 to T decodes it self-consistently, every repeated character mapping to the same letter both times, and the result is REDACTED. That is an inference from a text extraction, not a reading of the rendered page, and we flag it as one.
Second, the body. Grepping the extraction for lines with unusually wide internal whitespace turns up at least six passages where something has been removed. Two of them: “generate outputs using up to characters of grounded content” at page 13 (“Page 22 of 92”), and “Microsoft could SF564, 748” at page 15 (“Page 24 of 92”). The gaps sit mid-sentence, exactly where a number or a clause belongs.
“Unsealed” means newly public on the docket, and this filing is. “Unredacted” means nothing is blacked out, and this copy has blacked-out passages in it. In a story whose whole value is that hidden material is now visible, the difference is the story.
What Microsoft said, in one of the two
TheWrap is the only one of the two carrying an on-record Microsoft response. It published at 20:02:21 UTC and was modified at 21:47:22 UTC, so the statement landed after the story did. Per TheWrap, a Microsoft spokesperson said: “These comments reflect one employee’s individual perspective, are not a legal analysis, and do not represent the company’s views.”
TechCrunch reports that “OpenAI and Microsoft did not return requests for comment”. TheWrap reports that “OpenAI did not immediately respond to TheWrap’s request for comment”. If you are citing “Microsoft’s response” to this filing, you are citing TheWrap alone.
The second document in this case we have read, and 1557 is not 1587
On 8 September we published “Two Experts, One 8.2 Million Conversation Sample, and the Numbers 59,545 and 51”. Same case, different filing, and the docket numbers are close enough to be worth spelling out.
That post read Microsoft’s Rule 56.1 Statement of Undisputed Material Facts, docket entry 1557 in NYT v. Microsoft, filed 4 September by the defendant. This post reads Document 1587-1, the News Plaintiffs’ own combined brief, filed 17 September by the other side. Nine days apart, same summary judgment round, opposite parties. We have not fetched the full docket sheet, so we are not saying which document is a motion, an opposition or a cross-motion.
The gaps differ in kind too. The earlier piece was about two experts on the same side producing 59,545 and 51 from one 8.2-million-conversation sample, with the filing printing both and defining neither. This one is about one filing producing three ranges and two newsrooms each keeping a different subset. A measurement problem, then a compression problem.
What this cannot tell you
No court has found anything. This is a brief, the motion has not been decided, and every quotation above is the plaintiffs’ characterisation of documents produced in discovery, not a finding of fact.
We have the brief, not the exhibits. Every quote reaches us through the plaintiffs’ selection of it, in a document written to win a motion, and the underlying documents are sealed.
We read a text extraction, not the rendered pages. pdftotext reflows, and a layout artefact can look like a redaction. The six wide-whitespace passages and the cover-page decode agree with each other, which is why we report both, but a rendered read is the stronger check and we have not done it.
We checked four things, not everything: the click-through figures, the quoted phrases, the domain counts and the redaction status. Nothing here is a general verdict on either outlet’s accuracy, and others covered this filing too. We read these two because they cite the same document and published sixteen minutes apart, which is as close to a controlled comparison as press coverage gets.
What would change this reading
The exhibits would. If the sealed Hecht documents become public with dates on them, TechCrunch’s “January 2023” and “January 2024” either match or they do not, and that question closes either way.
What would not change is the selection. One filing gave three ranges, and its own introduction merged two of them on page 1. One newsroom kept the ceiling of one range, the other kept two ranges in full, and 51% to 94% is still on page 14 where anyone can read it.
Sources
- News Plaintiffs’ Combined Summary Judgment Brief, Document 1587-1, filed 17 September 2026, 92 pages. Every figure, quotation and page reference attributed to the brief here comes from that PDF.
- The docket in The New York Times Company v. Microsoft Corporation, CourtListener, for the case caption.
- Rebecca Bellan, TechCrunch, “Microsoft exec called AI scraping ‘the largest theft of labor in human history,’ new unredacted filings reveal”, 2026-09-17, 19:46:08 UTC.
- A.J. Katz, TheWrap, “OpenAI’s Head of ChatGPT Warned Publishers Faced an ‘Existential Threat’ in Unsealed Docs”, 2026-09-17, 20:02:21 UTC, modified 21:47:22 UTC.
- Our reading of the other filing: “Two Experts, One 8.2 Million Conversation Sample, and the Numbers 59,545 and 51”, 8 September 2026.
Every page above was fetched on 17 September 2026 and returned HTTP 200, and the brief was downloaded in full and read directly rather than through either write-up. The only arithmetic on this page is the sixteen-minute gap between the two publication timestamps, and both are printed above.