Anthropic's citations documentation, OpenAI's file search guide and Google's Search grounding documentation were read on 23 October 2026. Field names and limits change, so check each vendor's current documentation before building.

Grounding· LLM operations

Grounding is the practice of tying a chatbot's answer to specific source material, such as your help articles, product data or a web search, so the model answers from that material instead of from its training memory alone.

By Chatbotscape Editorial· Methodology· Published 24 October 2026· Updated 24 October 2026

Grounding — Making a Chatbot Answer From Your Sources

Quick answer: A grounded chatbot is given the material it should answer from and is told to stay inside it. The usual method is retrieval-augmented generation: find the relevant passages, hand them to the model, ask for an answer based only on them. A citation is the visible receipt, a link or reference to the passage used. Grounding lowers the risk of invented answers, but it does not remove it.

Grounding, retrieval and citation

The three words describe different layers, and a chatbot can have one without the others.

Term	What it is	Visible to the customer?
Retrieval	Finding passages in your content that may answer the question	No, it happens behind the scenes
Grounding	Instructing and constraining the model to answer from those passages	Only through the quality of the answer
Citation	Showing which source or passage an answer used	Yes, as a link, footnote or quoted line

A bot can retrieve passages and still answer from memory if the prompt does not constrain it. A bot can show a source link that was attached to the conversation by the interface and not by the model at all. And a bot can be well grounded without showing any citation. When someone asks whether a chatbot is "grounded," the useful follow-up is: grounded in what, and how is that checked?

What counts as a source

Sources fall into two groups. Your own content, such as help-center articles, policy pages, product feeds or documents, is the common case for a knowledge base chatbot. Live web search is the other: Google's documentation describes adding a Google Search tool so that Gemini decides whether to search, runs the queries and returns an answer with citations. The first suits a company that wants answers limited to what it has published. The second suits open questions but gives you less control over what is cited.

How vendors return citations

Each major model provider exposes citations differently. These descriptions come from the vendors' documentation read on 23 October 2026.

Anthropic. You enable citations on each document you supply. The response then includes text blocks that point back to the supporting passages. Plain text and PDF documents are chunked into sentences by default, and each citation carries the cited text, the document index and title, and a location (character range, page range or content-block range). Anthropic states that cited_text does not count toward output tokens, and that the feature guarantees valid pointers to the supplied documents. Limits include citations being all-or-none across a request, text-only support, and a 400 error if combined with structured-output format settings.
OpenAI. In the file search tool, the message output includes an annotations array whose file-citation entries carry a file ID and filename. The guide notes that search results are not returned unless you ask for them with the file_search_call.results include option, and that duplicate entries can appear.
Google. Google's documentation describes url_citation annotations on the answer text, each with a URL, a title and start and end positions, so an interface can attach a source to the exact span of text. It also describes a search-suggestions snippet and points to its terms for display rules.

The field names differ, but the idea is the same: the answer text plus a list of pointers into the sources.

A citation is a pointer, not proof

This is our own reasoning rather than a vendor claim. A pointer can be valid and still not support the sentence it sits beside. Anthropic's guarantee, for example, is that the pointer refers to a real place in your documents, which is not the same as the passage proving the claim. A chatbot can cite the right page for the wrong sentence.

Common ways a cited answer goes wrong:

Fabricated links. When a model is asked to write URLs from memory rather than from supplied data, it can produce addresses that look plausible but do not exist. The safe pattern is to let your system, not the model, build the link from a source ID.
Wrong passage. The cited chunk is related but does not say what the answer says, often because retrieval returned a near match. Reranking and better chunking reduce this.
Stale source. The page is real and the quote is accurate, but the policy has changed since.
Quote taken out of context. The sentence is genuine and the meaning is reversed by what surrounds it.

Worked example (invented numbers)

A team audits 50 answers that showed a source link.

41 answers are fully supported by the cited passage.
5 are partly supported: the right page, but one detail is not in it.
4 are not supported at all.
Check: 41 + 5 + 4 = 50. Fully supported is 41 ÷ 50 = 82%, partly 10%, unsupported 8%.

If the chatbot shows sources on 600 answers a month, an 8% unsupported rate means about 600 × 0.08 = 48 answers where the visible source does not back the claim. Those 48 may be worse than answers with no source, because the link invites trust. This is why teams sample and score cited answers, for example with a golden dataset or an LLM-as-a-judge check reviewed by a person.

FAQ

What is grounding in an AI chatbot?

Tying the chatbot's answers to specific source material, usually by retrieving relevant passages and instructing the model to answer only from them.

Is grounding the same as RAG?

RAG is the most common way to ground a chatbot, but the terms are not identical. Grounding is the goal; retrieval-augmented generation is one method of reaching it.

Does grounding stop hallucinations?

It reduces them but does not remove them. The model can still misread a passage, combine passages wrongly or answer from memory if the prompt allows it.

Is a citation the same as grounding?

No. A citation shows a source; grounding is about how the answer was produced. A source link can be added to an answer that was not grounded.

Do all chatbot platforms show sources?

No. Some show them, some make it optional and some do not offer it. Check the vendor's documentation and test it on your own content.

Can the model write the source URL itself?

It can, but it should not be trusted to. Build links from stored source IDs so that every link points to a real page.

Retrieval-augmented generation — the usual pipeline behind a grounded answer.
AI hallucination — the failure grounding is meant to reduce.
Chatbot knowledge base — the content a grounded bot answers from.
Reranking — a step that helps pick the right passage.
LLM-as-a-judge — one way to score whether an answer matches its source.
Golden dataset — a fixed set of questions for repeatable checks.

Sources

Anthropic, Citations — platform.claude.com/docs/en/build-with-claude/citations, read 23 October 2026. Source of the document types, citation fields, the cited_text token note, the valid-pointer statement and the limitations.
OpenAI, File search — developers.openai.com/api/docs/guides/tools-file-search, read 23 October 2026. Source of the file_citation annotation fields and the file_search_call.results include option.
Google, Grounding with Google Search — ai.google.dev/gemini-api/docs/google-search, read 23 October 2026. Source of the url_citation annotations and the search-suggestions snippet.
Chatbotscape evaluation methodology. /methodology (continuously updated).