A newly unsealed court filing puts a number on a problem search teams have argued about for years: how much traffic an AI answer removes before a publisher ever gets the click. The figure comes from the plaintiffs, not from a judge, and that distinction matters as much as the number itself.
The document, titled the News Plaintiffs’ Combined Summary Judgment Brief, landed on the docket September 17, part of a Manhattan federal lawsuit The New York Times and fellow publishers brought against OpenAI and Microsoft over copyright. It asks Judge Sidney Stein to find that the two companies infringed the plaintiffs’ copyrights and that their fair-use defenses do not hold up. Both companies reject that characterization of the evidence, and none of it has been adopted as a factual finding by the court.
Inside that brief, according to Search Engine Watch’s reporting on the filing, the plaintiffs cite Microsoft’s own representative data comparing Bing Chat to Bing Web Search. That comparison, as presented by the plaintiffs, shows click-through-rate reductions of 87 percent to 93 percent for Times domains, 83 percent to 91 percent for Daily News plaintiff domains, and 51 percent to 94 percent for Ziff Davis domains. These are the plaintiffs’ characterization of Microsoft’s numbers, filed to support a legal argument, not an audited or independently verified benchmark.
Nick Turley, who runs ChatGPT for OpenAI, is quoted in the same filing calling AI products “largely substitutive” for publisher visits and predicting they would only substitute more as the technology improved, per Search Engine Watch’s account of the Reuters-reported testimony. A separate internal Microsoft document cited by the plaintiffs quotes Brent Hecht, the company’s director of applied science, describing large-scale AI scraping as “the largest theft of labor in human history.” Both are characterizations pulled from internal material for a legal argument, not admissions of liability.
What separates this filing from the usual argument about whether AI answers are “transformative” is the range itself. A spread from 51 percent to 94 percent inside a single publisher group signals that substitution risk is uneven by content type, not a flat tax on every query. A search team cannot treat this as one number to plan against. The more useful exercise, once the underlying methodology becomes public through discovery, is mapping which query categories sit at the low end of that range and which sit at the high end, because that split is likely to track how directly a query can be answered without visiting the source.
Google’s separate AI Contribution Pilot, a Search Console-based program testing publisher payments tied to content that meaningfully feeds AI Overviews, AI Mode and Gemini, addresses compensation for the same dynamic this filing describes: an answer product satisfying a query before a click happens. It does not restore the click itself, and Google has not disclosed how individual payments are calculated. The filing and the pilot describe the same mechanism from opposite sides of the transaction: one puts a contested price on what was lost, the other tests a price for what remains.
None of this changes ranking mechanics for publishers today. It does change the evidence available to argue for licensing terms or product changes, since a range this wide, even unverified, is more concrete than a general claim about AI substitution. Search teams working with legal or business-development counterparts on licensing conversations now have plaintiff-side numbers to test against their own Search Console and referral data before the next negotiation.
Search Engine Watch (Radu Tyrsina) reported this September 18, 2026, based on the News Plaintiffs’ Combined Summary Judgment Brief filed September 17, 2026.
