Glossary
What is grounding in AI search?
Grounding is connecting an AI model's answer to specific sources, such as search results, documents or databases, so its claims are based on retrieved information rather than only on training data. Grounded answers usually show citations. Google uses the term for Gemini answers grounded in Google Search.
Grounded vs ungrounded
| Ungrounded | Grounded | |
|---|---|---|
| Based on | Training data | Retrieved sources |
| Current facts | Up to a cutoff | Yes |
| Citations | No | Usually |
| Hallucination risk | Higher | Lower |
Grounding with search
Google offers "grounding with Google Search" for Gemini, and its consumer AI features are grounded in its index. Other assistants ground with Bing or their own indexes. Either way, the pages that rank for the underlying searches are the material the answer is built from.
What it means for you
If you want an AI answer to describe your product correctly, make sure the pages that rank for your category describe it correctly: your own site, directories, review sites and comparisons. A wrong description on a highly ranked third-party page can end up in grounded answers.
Example
A grounded answer about a SaaS's pricing quotes a two-year-old review site listing because it ranks above the SaaS's own pricing page. The founder updates the listing and adds a clear pricing summary to the top of her pricing page. Within weeks, answers cite the correct price.
FAQ
Does grounding stop AI hallucinations?
It reduces them. The model can still misread or mix sources.
What is Google-Extended's role in grounding?
Google-Extended controls whether your content is used for grounding Gemini apps and training, separate from Search.
Can I see which sources grounded an answer?
Usually, through the citations shown.
Sources
- Google Cloud: Grounding overview (Vertex AI)
- Google Search Central: Google's common crawlers (Google-Extended)