Source
Mehta (2025), LIT-006, the remedies opinion in United States v. Google LLC (D.D.C., ECF No. 1436, filed 2 September 2025), read in full in NOTE-004. Page references are the opinion’s own (“op. p.”).
What was actually shown
The wrong and its fruit. Liability rests only on exclusive distribution agreements, not on product design, Chrome or self-preferencing (op. pp. 9–11). The fruits the court identifies are freedom from threats, scale and revenue (op. pp. 82–97). Scale is carried forward from the liability opinion: Google receives 9× the queries of all rivals combined (19× on mobile); 13 months of NavBoost data is about 17.5 years of Bing’s; of 3.7M unique phrases over 7 days, 93% were seen only by Google and 4.8% only by Bing (op. pp. 89–95). Long-tail queries are about a third of volume and 90% of distinct queries.
The causal standard. “Significant causal connection” is not but-for causation. “The more drastic the remedy, the greater the causal connection required” (op. p. 73, following New York v. Microsoft). The court grades the liability-phase inference as stronger than in Microsoft, enough for behavioural remedies, not structural ones (op. pp. 74–82). So Chrome and Android divestiture, the payment ban and choice screens are rejected.
The tailoring. What is shared is the input scale produced: a one-time snapshot of six index fields per document, the data underlying Glue and RankEmbed “at least twice” under a cap and privacy treatment, and five-year search and text-ads syndication (op. pp. 136–186). What is not shared is what Google engineered from it or what scale earned: ranking signals, models, the Knowledge Graph, ads data. The court accepts the testimony of Google’s own expert, qualified on cross-examination, that sharing raw Glue and RankEmbed data will not let rivals clone Search (op. pp. 159–163; NOTE-004 C16).
What could have come out otherwise. The court could have found the causal inference strong enough for structural relief, or found scale not a fruit at all. It did neither.
What this does not say
- That the remedy works. No one has used the data yet. The cap, the privacy treatment, the syndication taper and who qualifies are delegated to the Technical Committee and the final judgment (NOTE-004, Limitations). Both privacy experts agreed raw logs can re-identify users, and the EU Digital Markets Act precedent excluded 99% of queries after filtering (op. pp. 161–164). Whether it does is QUESTION-352.
- That search data is a moat for LLMs. The court found plaintiffs “did not establish that Google’s scale advantage in Search translates into a quality advantage in GenAI search-assisted responses” (op. p. 156), and so ordered no Gemini training data shared. That is a failure of proof on this record, not a finding of no advantage. Citing the opinion for the moat would be citing it for the opposite of what it holds.
- That remedies declined were judged ineffective. Several were refused for want of evidence (popularity signals, synthetic queries, publisher opt-out, Vertex access never requested), not on the merits.
- That this is settled law. It is a district-court opinion, and the court says it is “prepared to revisit” a payment ban (op. p. 128).
- Anything about how LLMs work. The GenAI findings restate testimony at a lay level under the court’s own heading “Greatly Simplified”.
Connections
- THEORY-150 and THEORY-470, Doctorow’s account of platform decay, name competition and interoperability as checks, and the talk read in NOTE-041 counts this case among the antitrust revival. This opinion is the record’s one primary document on how such a check is designed, and it chose data sharing over structure.
- THEORY-152 (exit and voice) treats exit as a recuperation mechanism. The court’s default-bias findings (op. pp. 83–85; of users paid to switch to Bing for 14 days, 33% stayed) bear on how far exit operates in this market.