Grounding
groundedness / grounded answers / grounded in your data
Grounding is tying an AI answer to retrieved sources rather than letting the model answer from memory. Google uses the word as a synonym for retrieval augmented generation. Microsoft treats groundedness as something you measure on a scale, which is the more useful way to think about it.
Google introduces the word as a second name for something else. Retrieval augmented generation is a technique, also known as grounding, used to improve the quality, accuracy, and freshness of AI responses by relying on its core Search ranking systems to retrieve relevant, up-to-date web pages from the Search index. So when a supplier says an answer is grounded, the claim is about where the material came from rather than about whether it is right.
The rest of Google's description is the part worth holding on to. Its systems then review the specific information from those retrieved pages to generate a more reliable and helpful response, showing prominent, clickable links to relevant web pages that support the information in the response. More reliable, not reliable. The links are evidence that something was retrieved, which is a weaker claim than it looks at a glance.
Vendor tooling makes the weaker claim explicit by measuring it. In Microsoft's evaluation set, groundedness measures how well the generated response aligns with the given context without fabricating content, and the result is a pass or fail against a threshold on a one to five scale. A separate evaluator measures how relevant the retrieved context chunks are to addressing a query, judged by another model.
Read those two together and the shape of the thing appears. Grounding can fail at either step. The system can retrieve the wrong material and summarize it faithfully, or retrieve the right material and drift away from it. They are different failures with different fixes, and both are invisible in an answer that carries citations.
In practice
If somebody is selling you a grounded assistant, the useful question is not whether it is grounded but what the groundedness threshold is and who chose it. A score with a threshold is a decision somebody made about how much drift is acceptable, and that decision belongs to you rather than to the vendor who set the default.
Not the same as
- Accuracy
- Grounded means the answer is tied to retrieved material. If the retrieved material is wrong or out of date, a perfectly grounded answer is wrong too.
- Citing sources
- Links show that retrieval happened. They do not show that the sentence above them follows from what was retrieved, which is the thing the groundedness score is trying to measure.
Why it matters to you
Grounded is becoming the word that ends the conversation about whether an AI answer can be trusted, and it does not settle that at all. Knowing that it names a process, and that the alignment between answer and source is a separate measured quantity, gives you a question with a numeric answer instead of a reassurance.
What to ask or check
- 01What is the groundedness threshold, and who set it?
- 02Is retrieval quality measured separately from how faithfully the answer follows it?
- 03When the score fails, what happens to the answer the customer sees?
What people get wrong
That a grounded answer is a correct one. Grounding describes where the material came from, and the alignment between answer and material is scored separately against a threshold somebody chose.
Red flags
- Grounded used as a synonym for accurate.
- Citations offered as proof rather than as evidence of retrieval.
- No answer to what happens when the groundedness score fails.
Who owns it
The vendor builds the retrieval and sets the default threshold. Deciding how much drift is acceptable in answers given in your name is not a vendor decision.
Where you will see it
In the pitch for any assistant that answers from your documents, and in the links under an AI answer about your industry.