What is grounding

Two things that often get mixed together belong apart. Retrieval-augmented generation is an architecture: it describes how you build a system that searches for documents before writing. Grounding is a property of the result: it describes how far what the answer says is held up by those documents. One is the blueprint of the machine, the other is the measure of what comes out the far end.

The distinction is not academic. A system built with retrieval can return a poorly grounded answer, because the model received the right fragments and still wrote a sentence that none of them supports. The opposite also happens, when a model without retrieval mentions from memory a source that exists and that does in fact say what it claims. Retrieving and supporting are different operations, and the first does not guarantee the second.

Grounded, in practice, means three conditions hold together. The cited source exists and is accessible. The source genuinely deals with the matter the sentence is about. And the source claims what the sentence attributes to it. Serious evaluation systems check all three separately, because each fails on its own and the last one breaks most often.

The mechanism has a retrieval part and a stitching part. Retrieval hands the model a set of fragments together with the user's question. The stitching is what separates a grounded answer from any old summary: the system preserves the correspondence between each stretch of generated text and the fragment it came from, and that correspondence travels with the answer instead of being lost during writing.

In Google's implementation that information arrives through the programming interface itself. Alongside the text it returns the queries the model launched on its own, the results it obtained, and annotations marking, with a start and end position inside the answer, which segment corresponds to which source address. With those marks the application can place the footnote at the exact point rather than piling links at the end of the text. The product offering this connection to the search index is called Grounding with Google Search, and its terms of use require the accompanying search suggestions to be displayed.

There is a variant that never touches the open web. A system can be grounded against a closed corpus, an internal document base or a product catalogue, using the same scheme of source marks. Grounding describes the link to the source, whether that source is public or private, which is why it turns up in conversational search engines as much as in internal company assistants.

What no marking scheme solves is the quality of what was retrieved. If the selected fragment says something wrong, the answer will be perfectly grounded in an error, with its impeccable link underneath.

Grounding is often sold as a guarantee of truth, and it is not. It lowers the invention rate because it forces the model to start from a specific text, but in exchange it introduces a new dependency: the answer inherits the errors, the omissions and the age of whatever was retrieved.

The available evidence is uncomfortable. A paper published in April 2025 evaluated seven popular models across some eight hundred medical questions and around fifty-eight thousand statement and source pairs. Between 50 and 90 per cent of the answers were not fully supported by the sources they cited, and in some cases the source contradicted the claim. Even with web search enabled, roughly a third of individual statements went unsupported. The field was medicine and the figure does not transfer wholesale to other topics, but it fixes the order of magnitude of the problem.

From that follows the practical decision, and it is twofold. Anyone publishing must assume their pages will be cited in support of sentences they never wrote, which turns the clarity of every claim into a matter of reputation. And anyone consuming AI answers inside a workflow needs a verification step of their own for anything with consequences, because the link says where the system claims it got the fact, not that the fact is correct.

• Put the verifiable claim and its figure in the same sentence or paragraph as the term they refer to, so the retrieved fragment stands up without the rest of the page.
• Date every perishable fact inside the visible text, because the model cannot tell a 2019 figure from one from this year if the page does not say so.
• Keep the page holding the fact crawlable and free of walls, since a fragment nobody can retrieve never grounds anything.
• Check regularly which specific sentence gets attributed to the brand in generated answers, rather than merely counting whether the brand appears.
• In your own systems, store the source marks the model returns and display them at the exact point in the text, not as a list of links at the end.
• Set out in writing which kinds of claims require human verification before publication, using risk criteria rather than case by case judgement.

• Treating the presence of links under an answer as proof of accuracy. They indicate declared origin, and that is the whole of what they tell you.
• Confusing grounding with retrieval-augmented generation. A system can have retrieval and still produce poorly grounded answers every day.
• Measuring only the percentage of answers carrying a citation, without checking whether the citation supports the specific sentence it is attached to.
• Splitting a fact across distant sections of the page, so that no retrievable fragment contains the complete claim.
• Requiring consent, a login or script execution to read the content you want to see cited.