Two publisher groups sue Brave over a 30-billion-page index resold to AI labs

If your organisation publishes to be found, the so what is blunt: content you paid to produce may now be grounding data inside someone else's assistant, with no click, no lead and no licence fee. That applies to UAE and GCC brands with a real content operation, not only to newspapers. The asymmetry is the story. Brave's own AI pages market the index for LLM training and RAG, while Google publishes Google-Extended in robots.txt so publishers can withhold content from AI training and grounding. One vendor documents a switch. The other documents that it has none.

Where most teams will get it wrong: they read this as a newspaper story, then add a Disallow line for a crawler that has no user agent to match. That line does nothing. Brave's webmaster page says it gives no robots.txt instructions and runs no opt-out, which leaves noindex as the only lever it points at, and noindex also removes you from Google. Blocking every AI crawler by reflex is wrong too, because it costs you discovery in the assistants that already send qualified traffic. Set a policy per content class, enforce it at the edge, and stop depending on public indexes for the answers that represent your brand.

Read the vendor documents rather than the coverage, because the numbers are the argument. Brave's own API documentation describes the LLM Context endpoint as pre-extracted page text built for AI agents, LLM grounding and RAG pipelines (https://api-dashboard.search.brave.com/documentation/services/llm-context): it launched on 6 February 2026, its default extraction pipeline changed on 31 July 2026, and one call returns up to 50 URLs with a token budget set from 1,024 to 32,768 tokens (8,192 is the default) and 512 to 8,192 tokens per URL. Brave's own Search API page (https://brave.com/search/api/) prices that at $5 per 1,000 requests on the Search plan at 50 queries per second, and $4 per 1,000 requests plus $5 per million input and output tokens on the Answers plan at 2 queries per second, and its own FAQ confirms that results may be stored and used to train models on a plan that includes storage rights. One request that ships up to 32,768 tokens of your pages to a model therefore costs half a cent. Per-click economics do not survive that comparison. Brave's own published runs put grounding at 85.3% accuracy and $18 per 1,000 questions on FreshQA, against $49 to $89 for Parallel Advanced, Exa and Tavily. Google documents the opposite posture: Google-Extended has been a robots.txt token since 2023, covering training of future Gemini models and grounding in Gemini Apps and Vertex AI (https://developers.google.com/crawling/docs/crawlers-fetchers/google-common-crawlers), and on 30 October 2026 Google said it will support the shared IETF AIPREF training token, user-agent: __usage_ai-training, in early 2027 (https://developers.google.com/search/blog/2026/10/ai-training-control).

How we would approach it. First, inventory: run a site: query on Brave Search, diff the result set against your gated, premium and paywalled routes, and keep the export as a baseline. Second, enforce at the edge rather than in robots.txt: block Brave's AI training and preview crawlers by user agent, and block the IP ranges Brave documents for its crawlers, on routes you never want resold, with X-Robots-Tag: noindex as a second layer. Third, retrieval you control: hold the corpus your brand should answer from in a self-hosted knowledge base such as Kryton, and put a gateway such as FastLLM Proxy in front of every model provider so routing, budgets and access control sit with you. An nginx sketch, using the crawler identity and ranges documented at https://brave.com/webmasters/:

# Block by crawler identity and by Brave's documented verification ranges
geo $brave_bot_net {
    default            0;
    34.80.255.72/29     1;
    34.138.255.72/29    1;
}
map $http_user_agent $brave_ai_bot {
    default                0;
    "~*BraveBot-Training"  1;
    "~*BraveBot-Preview"   1;
}

Routes you never want resold through the Search API location /premium/ { add_header X-Robots-Tag "noindex, nofollow" always; if ($brave_bot_net) { return 403; } if ($brave_ai_bot) { return 403; } }

Everything else: allow BraveBot for search, deny the AI crawlers

location / { if ($brave_ai_bot) { return 403; } }

The view from the UAE: there is opportunity alongside the risk. Arabic and regional publisher content is thin in the indexes that Western model builders license, so these APIs are a distribution channel worth choosing deliberately, page class by page class, rather than sitting inside by default. Put the question to your vendors in writing: is Brave Search in the subprocessor list, and is it used for training, grounding or both? Keep those answers next to your agent platform, so that when a licensing negotiation or a claim like this one lands you can show provenance instead of inferring it. Azrty is a UAE company, and our team in Dubai does this work: index audit, edge enforcement, a private retrieval layer and the governance around it.