Grounding
Grounding is the practice of making an AI model base its answer on specific, trusted sources supplied at the time of the question, such as search results or company documents, rather than only on what it learned in training.
An ungrounded model answers from memory: patterns absorbed during training, frozen at its knowledge cutoff and prone to hallucination on details. A grounded model is handed the relevant material first and told to answer from it, often with citations, so its answers can be checked and kept current.
Grounding is the goal; retrieval-augmented generation is the most common way to achieve it. Grounding is also different from fine-tuning, which changes the model itself. A grounded model can answer about yesterday's price change without being retrained. It also lets the answer cite where each fact came from.
How it works
When a question comes in, the system first finds sources likely to contain the answer. That may be a web search, a search of the company's help articles, or a lookup in a database such as the customer's order. The best passages are placed in the prompt together with an instruction to answer only from them and to say so if they do not contain the answer.
The model then writes its reply using those passages. Well-built systems attach citations to each claim so a reader can open the source. Some go further and check, after the answer is written, that every claim is supported by the retrieved text.
Grounding reduces errors but does not eliminate them. The model can misread a passage, combine two sources wrongly or fall back on memory when the sources are thin. The quality of the answer can never exceed the quality of the sources: an out-of-date policy document gives a confidently grounded, wrong answer. How the sources are found matters as much as how the answer is written. If the search step returns the wrong passages, even a careful model will answer from the wrong material, so testing a grounded system means checking what was retrieved, not only what was said.
Example
An online garden shop adds an AI assistant to its help page. Each time a customer asks a question, the system searches the shop's own delivery, returns and product pages, passes the top passages to the model and requires a link to the page used. A customer asks whether live plants can be returned. The assistant answers that they can be exchanged within seven days if they arrive damaged, and links to the returns page.
The owner tests grounding by asking questions the documents do not cover, such as whether the shop delivers to an island it has never shipped to. A well-grounded assistant says it cannot find that information and offers the contact form. An early version guessed yes, so the instructions were tightened and a fallback added.
Why it matters for your business
If you put AI in front of customers, grounding is what keeps it to your real prices, policies and stock rather than plausible guesses that you may be held to. It also explains how AI search tools decide what to say about you: they ground answers in pages they can find and trust. Clear, current, well-structured pages make you a better source for both your own tools and other people's. When choosing a tool, ask whether it shows the sources behind each answer and what it does when it finds none.
Common mistakes
- Grounding an assistant in documents that are outdated or contradict each other.
- Not telling the model what to do when the sources do not contain the answer.
- Assuming citations prove an answer is correct without opening the cited page.
- Feeding the model whole manuals instead of the few passages relevant to each question.
Questions people ask
What is the difference between grounding and RAG?
Grounding is the aim of tying an AI answer to specific, trusted sources. RAG, retrieval-augmented generation, is a common technique for achieving it: retrieve relevant passages, then generate an answer from them. Grounding can also use other sources, such as live database lookups or a document the user uploads.
Does grounding stop AI hallucinations?
It reduces them significantly, especially for facts that are in the sources, but does not remove them. The model can still misread or misquote a passage, or invent details when sources are missing. Citations and clear fallback instructions make the remaining errors easier to catch.
How can I tell if an AI answer is grounded?
Look for citations or links to specific sources, then open them and check that they say what the answer claims. An answer with no sources, or with sources that do not contain the stated facts, is drawing on the model's training memory rather than being grounded.
