Grounding license
By Gareth HoyleLast updated How we research
A grounding license permits an AI company to use content at the moment of answering, to support or cite responses, rather than to train a model.
A grounding license is a content license that lets an AI provider retrieve and show the licensor's content when answering users, usually through retrieval-augmented generation or a search index, without necessarily granting rights to train on it. Grounding keeps answers current and gives the model sources to cite.
What does a grounding license look like in practice?
News deals show the pattern. In January 2025 Google signed a deal with the Associated Press to deliver up-to-date news in its Gemini chatbot, as AP reported, and the report did not say whether training was included. Google also treats grounding as a separate use in its crawler controls: its Google-Extended token covers both Gemini training and providing Search content to the model at prompt time.
What terms matter in a grounding license?
- Whether training is excluded, permitted or licensed separately
- Attribution, linking and display limits on how much text appears in an answer
- Update frequency and the archive depth the buyer can query
- Recurring fees and what happens to cached content when the license ends
See the AI data deals tracker for published grounding and training deals.