What Is Grounding Source?

How Grounding Source Works

A modern AI answer is rarely written from memory alone. When a system uses retrieval-augmented generation, it first fetches a set of candidate documents, then writes an answer that stitches together passages from a handful of them. The act of tying a generated sentence back to one of those retrieved passages is called grounding, and the passage it lands on is the grounding source. Without one, a claim is just the model asserting something; with one, the claim carries a receipt a reader can check.

This is visible in production. In the Gemini API’s Grounding with Google Search, a grounded response ships with a groundingMetadata object. Inside it, groundingChunks lists the web sources the model used — each entry a uri and a title — and a groundingSupports array maps individual spans of the answer text (by start and end index) to the chunks that back them. That mapping is the machine-readable version of a grounding source: this sentence rests on that document. The same structure drives the citations you see beneath AI Overviews, Perplexity answers, and other answer engines.

Because grounding happens at synthesis time, being retrievable is necessary but not sufficient. A page enters the candidate pool only if an AI crawler can reach and read it. But whether it becomes the cited grounding source depends on extractability — whether a passage states its claim cleanly enough to be quoted without the rest of the page for context. That is the property that separates a page that ranks from a page that gets cited.

What Counts as a Grounding Source

Grounding sources fall into a few practical categories depending on where the retrieval happens:

• Live web documents — pages pulled from a search index at answer time, as in Google AI Mode or Gemini’s Google Search grounding. These are the grounding sources SEO and generative engine optimization can actually influence.
• Provided context documents — a file, transcript, or knowledge-base article handed directly to the model in the prompt, common in enterprise assistants. The model must ground its answer strictly in that supplied text.
• Vector-store passages — chunks retrieved from an internal database, where the grounding source is a single embedded passage rather than a full page.

In every case the grounding source is a passage, not a domain. A single long article can contribute several distinct grounding sources, and the same document can be a grounding source for one query and invisible for the next.

Example of Grounding Source

The clearest documented illustration of grounding sources at work is Google DeepMind’s FACTS Grounding benchmark, published on December 17, 2024. The benchmark exists to answer one narrow question: when a model is handed a source document and asked something about it, how faithfully does the answer stay anchored to that grounding source rather than drifting into invented detail?

The setup is specific and verifiable. FACTS Grounding contains 1,719 examples, split into an 860-example public set and an 859-example private, held-out set. Each example pairs a user request with a context document — the grounding source — of up to 32,000 tokens, roughly 20,000 words, spanning domains like medicine, law, finance, and technology. A response only passes if it is, in the benchmark’s words, fully grounded in information contained in the provided document, with no hallucinations.

The scoring is itself instructive about how grounding is judged in practice. FACTS runs a two-phase evaluation: first it checks whether a response is even eligible — that it genuinely addresses the request rather than dodging it — then it scores factual accuracy strictly against the grounding source. To reduce single-model bias, it aggregates the verdicts of three frontier LLM judges: Gemini 1.5 Pro, GPT-4o, and Claude 3.5 Sonnet, and the final number is the average of their scores across every example.

The lesson generalizes straight to content strategy. FACTS measures whether a model stays tethered to a grounding source it was given; retrieval-based AI search adds an earlier step, deciding which passage becomes that source in the first place. Both stages reward the same trait: a claim that is stated plainly and can be checked against the text supporting it. A page that supplies numbers with a named origin and self-contained phrasing is easier for a system to adopt as a grounding source and easier for it to keep anchored to once adopted — which is why front-loading a verifiable, quotable claim is the highest-leverage move for earning citations.

Grounding Source and Citation Readiness

Earning grounding sources is less about volume and more about supply: you are supplying the sentence a machine can quote without staking its own credibility on your ambiguity. That is the discipline of citation readiness — writing passages engineered to be lifted intact. When your text is the cleanest available evidence for a claim, and an AI crawler can reach it, you become the grounding source. Track how often that happens with citation share: the percentage of AI answers on your topics that name you as the source, rather than a competitor.

Frequently Asked Questions

The Bottom Line

A grounding source is the evidence trail behind an AI answer — the exact passage the system leaned on to justify a sentence, exposed to the reader as a citation. It shifts the unit of optimization from your whole page to the individual claim: the site that supplies the cleanest, most checkable statement of a fact is the one the machine anchors to, and the one it names.