What is grounding in AI?

Search for "grounding" and you will get three unrelated things: the earth connection of an electrical installation, a psychology technique for calming anxiety and the one that matters here, the one from artificial intelligence. In AI, grounding is the moment an assistant stops answering from memory and goes to look something up. For a store it is the moment that counts, because it is when your product page can make it into an answer and get cited. Without grounding, the model answers with what it learned in training and has no specific source to link.

How it works

A language model keeps what it learned in its parameters. The paper that named retrieval-augmented generation (RAG), by Patrick Lewis and eleven other authors and presented at NeurIPS 2020, called that parametric memory, and added a non-parametric memory to it: a Wikipedia index queried with a trained retriever. The result was that models with retrieval generated language that was more specific, more varied and more factual than the same kind of model with parametric memory alone.⁠^{[1]arXiv (Lewis, Perez, Piktus, Petroni, Karpukhin, Goyal, Küttler, Lewis, Yih, Rocktäschel, Riedel and Kiela; NeurIPS 2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks}

Google defines grounding in its crawler documentation: giving the model content from Google Search's index at the moment of the question, to improve the accuracy and relevance of the answer.⁠^{[2]Google Search Central. Google's common crawlers} In its guide to Google's AI search features it uses it as a synonym for RAG: a technique that improves the quality, accuracy and freshness of answers by relying on its usual ranking systems to retrieve relevant, current pages from its index. Then its systems review the specific information on those pages to generate a more reliable answer.⁠^{[3]Google Search Central. Optimizing your website for generative AI features on Google Search}

The Gemini API shows the process step by step when Google Search is turned on as a tool:⁠^{[4]Google AI for Developers (Gemini API). Grounding with Google Search}

1. The app sends the user's question to the model with the search tool turned on.
2. The model analyzes the question and decides whether a Google search can improve the answer.
3. If so, it generates the queries and runs them itself.
4. It processes the results, synthesizes the information and writes the answer.
5. It returns a grounded answer, with the queries it ran and with citations that tie each passage of the text to the URL that backs it.

Google sums up what it is for in three benefits: more accuracy, because it reduces hallucinations by basing the answer on real information; access to recent information; and citations that show where the model's claims come from.⁠^{[4]Google AI for Developers (Gemini API). Grounding with Google Search} Reducing is not eliminating, and it is worth reading it that way.

For a store, three of those steps matter: that the model decides to search, which queries it writes and which pages it ends up citing. The first does not depend on you. The third does, in part: in Google Search, the pages that get cited come from its index, retrieved with the same systems that order the usual results.⁠^{[3]Google Search Central. Optimizing your website for generative AI features on Google Search}

Who decides whether your pages serve as a source

At Google there are two paths, and they are controlled separately. Google-Extended is the name Google lets you use, from robots.txt, to decide whether the content it crawls from your site can be used to train the next generations of Gemini models and for grounding in the Gemini apps and in the Grounding with Google Search service of Vertex AI. It does not affect your site's inclusion in Search and it is not a ranking signal.⁠^{[2]Google Search Central. Google's common crawlers} To show up as a supporting link in AI Overviews or in AI Mode, what counts are Search's rules: the page has to be indexed and able to be shown with a snippet.⁠^{[5]Google Search Central. AI features and your website}

At Bing, the word already appears in the tools for publishers. Since February 10, 2026, the AI Performance report in Bing Webmaster Tools shows grounding queries: the key phrases the AI used when it retrieved the content that ended up cited in its answers. Microsoft warns that the data is a sample of citation activity, not the total.⁠^{[6]Microsoft, Webmaster Tools blog. Introducing AI Performance in Bing Webmaster Tools Public Preview}

In a store

Fictitious example. A customer asks an assistant which lightweight travel tripod can hold a DSLR with a telephoto lens. The same question is resolved in two ways depending on whether the assistant searches or not:

The right-hand column is the one a store can work on. What the assistant cites from your product page will be what the page says: if the maximum load or the folded weight is not written on the page, the assistant cannot take it from there.

How to check it

1. Ask an assistant with search a purchase question from your category, without your brand, and note whether it cites sources and which. Repeat it another day: a single answer is a sample of one.
2. Open each cited page and look for the sentence. The citation says where the information came from; that grounding reduces hallucinations does not mean it eliminates them.⁠^{[4]Google AI for Developers (Gemini API). Grounding with Google Search}
3. Look in robots.txt at what you tell Google-Extended and decide knowing what it controls: training and grounding in Gemini and Vertex AI, not Search.⁠^{[2]Google Search Central. Google's common crawlers}
4. If you have your site in Bing Webmaster Tools, open AI Performance and read the grounding queries that led to your pages being cited.⁠^{[6]Microsoft, Webmaster Tools blog. Introducing AI Performance in Bing Webmaster Tools Public Preview}
5. Confirm in Search Console that your product pages are indexed and that nothing keeps them from showing a snippet, the condition for serving as a supporting link in Google's AI features.⁠^{[5]Google Search Central. AI features and your website}

Related concept: RAG

RAG and grounding are used as synonyms, and Google does so in its own guide.⁠^{[3]Google Search Central. Optimizing your website for generative AI features on Google Search} The useful difference is one of scope. RAG names an architecture: a retriever that looks up documents in an index and a generator that writes the answer with them, as in the 2020 paper.⁠^{[1]arXiv (Lewis, Perez, Piktus, Petroni, Karpukhin, Goyal, Küttler, Lewis, Yih, Rocktäschel, Riedel and Kiela; NeurIPS 2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks} Grounding names the goal, that the answer rests on an external source, and that source can be a search engine's index, as in the Gemini API or in Bing's answers.⁠^{[4]Google AI for Developers (Gemini API). Grounding with Google Search}⁠^{[6]Microsoft, Webmaster Tools blog. Introducing AI Performance in Bing Webmaster Tools Public Preview} Put another way, RAG is one way of doing grounding, and grounding with a search engine is the kind that can end up citing your store.

To go on, the GEO (Generative Engine Optimization) entry explains how to work so a generative engine cites your store, and the AI search reading path puts the rest of the terms in order.