Grounding

A model on its own answers from what it learned in training, which stops at a cutoff and can be wrong. Google Cloud defines grounding as the ability to connect model output to verifiable sources of information, and says that tying output to a data source reduces the chances of inventing content and gives auditability, because the response carries links to its sources^[1]. For a website, that is the part of an AI answer it can reach: the sources are pages.

The word names a mechanism, and one vendor’s account shows its steps. Google’s Gemini API describes Grounding with Google Search as connecting the model to real-time web content, so it can answer beyond its knowledge cutoff and cite verifiable sources. When a search helps, the model generates one or more search queries and runs them, processes the results, and writes a response whose inline citations each link a segment of the text to a source URL^[2].

Google describes the same step inside its own AI features. It lists retrieval-augmented generation, “also known as grounding”, as a technique that relies on its core Search ranking systems to retrieve relevant, up-to-date pages from the Search index, and then shows prominent, clickable links to pages that support the response^[3]. Microsoft describes grounding as the system that connects AI to current, authoritative information, and says its Bing Webmaster Tools updates give early insight into how content contributes to AI-generated experiences, including citation and grounding signals^[4].

A robots.txt token reaches part of this, and only part. Google says its Google-Extended token lets a site manage whether content Google crawls may be used to train future Gemini models and for grounding in Gemini Apps and in Grounding with Google Search on Vertex AI, and that it does not affect a site’s inclusion in Google Search or serve as a ranking signal^[5]. The AI features inside Search are set elsewhere. The Search generative AI control in Search Console sets whether a site is included in AI Overviews, AI Mode and the generative AI features of Discover, including whether its content helps ground their responses^[8]. Robots.txt rules for Googlebot govern how a site is crawled for Search as a whole^[6], and Google-Extended is the token that limits training of the models those features use^[8]. In QueryRev’s study, Google-Extended is named and blocked on 411 of the 6,294 sites whose robots.txt could be read, 6.5%^[7].

Grounding is not a score a tool hands over. What a team can measure is whether an answer is tied to sources, and to which. For one question on one surface, record the answer, split it into its claims, and note for each claim whether a source is shown for it and whether that source says it.

Then count two things apart. The share of claims that rest on a source shows how grounded the answer is; the share of cited sources that are yours shows what the grounding does for your site. A well-grounded answer can cite sources that do not include your page.

Repeat the check on the same questions and surfaces on later dates. A grounded answer is built at the moment of the question from whatever is retrieved, so one reading is a sample and the pattern across readings is the finding.

Illustrative: an invented site, to show the method.

A home-energy site asks one surface a question: “how fast does a heat pump pay back”. The answer has four sentences and three linked sources, and the site’s guide is the second. Read sentence by sentence, the payback, fuel-price and installation sentences each lean on a source. The sentence that gives an installed price leans on none: it comes from the model’s own recall.

The site opens the page behind source 2 and finds that it gives no price range. It adds a dated price table with the source of its figures, keeps the question in its tracked set, and reads the next answers to see whether the price sentence now has a source, and whether it is the site’s.