Google Legal AI Push Pits Gemini Against OpenAI and Claude

The short version

Google has entered the contest to sell generative AI to lawyers. Its main product is Gemini Enterprise for Legal, an agentic platform now in preview at several of the world's largest law firms. OpenAI and Anthropic are going after the same customers. The more revealing part of the story, though, is how Google describes the product. It is selling lawyers answers that are grounded in authoritative, often licensed sources and that keep a citation trail back to each document. That is the standard publishers say Google's consumer search product is moving away from, as AI Overviews keep more readers on the results page and send fewer clicks to the sites that supply the material.

What Google launched

Google Cloud announced Gemini Enterprise for Legal in late August. Most coverage dates it to August 252930, and at least one write-up gives August 2622. It is a legal version of the broader Gemini Enterprise platform, aimed at law firms and corporate legal departments29. The launch partners are Cleary Gottlieb, Freshfields, Weil, and Williams & Connolly2130. Google released a financial-services version the same day and called the two the first of a planned series of packaged industry products30.

The design is not a lawyer chatbot. Google built it from four parts: reusable legal "skills" written by domain experts, connectors to the systems firms already run, agents that carry tasks through those systems, and an open partner ecosystem, all managed from one governance layer29. Listed workflows include regulatory horizon scanning, data-subject access requests, contract review and playbook benchmarking, redaction for motions to seal, and NDA drafting22. The connector list covers Microsoft 365 and Google Workspace, iManage, NetDocuments, Docusign, Everlaw, RelativityOne, Thomson Reuters HighQ, the Free Law Project's CourtListener, and AI vendors Harvey and Legora28.

Since launch, Google has been showing firms what it can do. At a September 16 event it demonstrated contract summaries, grounded research in NotebookLM, and custom "Cloud Legal AI Helpers" for deal work. The naming overlap with the legal platform suggests the product line is still being tidied up24. On October 8, Cooley said it is building a Gemini Enterprise agent that proposes redactions of confidential material in court filings, with attorneys making the final calls25.

Three labs, one market

Google is not first. Anthropic has been releasing tools for lawyers built on Claude since January3. Engadget dates the product now called Claude Legal Solutions to about three months before Google's preview23. OpenAI followed on September 17 with Astra for Law. It is a legal version of its GPT-6 Astra model, paired with an index of US case law, statutes and regulations, and integrated with Relativity, Clio, Intapp and Thomson Reuters3. OpenAI says it has tested legal applications with Sullivan & Cromwell, Latham & Watkins and Wachtell Lipton, among others3. Engadget counts three of the four major labs as having launched or pitched legal products this year, with xAI promoting Grok for legal work too2321.

The three products have converged on the same design: a frontier model wrapped around authoritative legal data and existing firm software. All three also partner with Harvey and Legora, the specialist vendors that sit between the labs and the firms328. So the labs are competing to be the infrastructure layer, and the legal-tech startups are customers of all of them.

Google's edge is reach. Firms already using its cloud and productivity tools can add legal AI without changing environments24. Its weak spots are disclosure. The launch materials published no legal-specific pricing, no general-availability date, and no accuracy or error rates29.

Grounding is the sales pitch

The pitch is aimed at legal AI's best-known failure. Researcher Damien Charlotin's database lists close to 2,000 cases worldwide where AI-fabricated material reached a court, involving more than 800 lawyers23. Google Cloud CEO Thomas Kurian said at launch that every step of these agentic workflows must be "accurate, factual, and grounded in legal authority"30. Google says research outputs are anchored in primary legal authority rather than training data alone, that connectors inherit existing ethical walls and permissions, and that customer data is never used to train base models30.

Independent analysts are less sure. One review argues that retrieval moves the risk rather than removing it: a system can miss the controlling case, rank outdated law too highly, or cite a real decision while misstating its holding21. The same review notes that no independent public audit has measured error rates for either Gemini Enterprise for Legal or Anthropic's 2026 legal package21. Google's general counsel, Halimah DeLaine Prado, has called AI a complement to lawyers rather than a replacement28. That fits the position all the vendors take, which leaves responsibility for each citation with the lawyer who signs the filing.

The contrast with search

This is where the legal launch meets the search story. In its enterprise pitch, Google treats grounding and traceable citations as non-negotiable. In consumer search, the main venue for AI answers is AI Overviews, which Google says reaches more than 2.5 billion monthly users11. One compilation of independent research reports that 11% of Overview claims are not supported by the sources they cite11.

The traffic data is consistent and getting worse. Chartbeat's 2026 Publisher Playbook found Google Search referrals to news sites fell 40.2% between July 2025 and July 2026, and Search's share of publisher pageviews dropped from 9% to 5% over two years12. A randomized field experiment found 39.8% fewer outbound organic clicks on searches that showed an Overview. Only 7.1% of the remaining clicks went to links inside the AI answer, and users reported no improvement in satisfaction17. Pew's often-cited figure is that users click a traditional result 8% of the time when an AI summary appears, compared with 15% when it doesn't18. AI chatbots are not making up the difference: Chartbeat puts AI referrals at about 0.01% of publisher traffic12.

The Overviews are also showing up on more searches. DemandSphere tracked Overviews on 26% of branded queries at the start of September and 90% by September 27, with no announcement from Google16. Since late August, some Overviews have expanded automatically into full AI Mode answers, pushing organic links further down the page18. For SEO teams, the mechanics have changed. No submission process gets a page into an Overview, the cited sources rarely change, and the advantage goes to being cited rather than clicked: cited brands reportedly get about 120% more organic clicks per impression16.

Google's response, and where the coverage differs

Google has made concessions. In 2026 it added inline links, hover previews, and "Subscribed" labels to AI answers1418, and an AI Overviews opt-out reached every site on August 3118. More significantly, it is now paying some publishers. An invite-only pilot pays roughly 100 sites when their content contributes to AI answers1519. That breaks with Google's long-held argument that search traffic was payment enough15.

On whether the payments matter, the coverage splits. Business-oriented write-ups call it a pivotal shift toward fairer terms19. Reporting based on The Information's sources is more critical. Many small and midsize participants receive less than 0.1% of their ad revenue, larger publishers are staying out to push for better terms, and Google decides on its own what the content is worth15. The second account seems more credible, because it rests on actual payout figures rather than on how the program is framed.

The courts have not given publishers much leverage. On September 30, Judge Amit Mehta dismissed the Chegg and Penske Media suits over AI Overviews. He held that an expectation of traffic "is not an agreement" and rejected the tying and standing theories16. Europe has been more willing to act: a Munich court found Google answerable for what its Overviews say about two publishers16.

The takeaway

Read together, these developments show a split strategy. Lawyers face sanctions over bad citations, so Google sells them grounded, auditable, licensed retrieval through partners like Thomson Reuters, and charges enterprise prices for it. The open web supplies the raw material for AI Overviews and gets fewer clicks, a small and opaque payment pilot, and little help from US courts. The legal AI race is a real contest between Google, OpenAI and Anthropic, and Google's distribution gives it a strong position. But Google's own legal product shows how much it values verified sources when a customer pays for them, and publishers are likely to keep pointing to that contrast.