Grounding

Grounding is a technique that ties an AI model's answer to sources retrieved at the moment of the prompt, such as Google Search or Bing results, instead of relying only on what the model learned in training. Gemini, ChatGPT search, Microsoft Copilot and Google AI Mode use it to answer with current facts and cite their sources.

Why grounding matters for your brand

Grounding decides which sources an AI answer can lean on, and only a grounded answer can cite a page at all. A model that answers from memory alone has nothing to link to.

Grounding also changes how often answers are right. On SimpleQA, a benchmark of short factual questions, OpenAI reported in March 2025 that GPT-4o answered 38% correctly on its own, while GPT-4o search preview, its search-enabled version, scored 90%. OpenAI says its web search tool runs on the same model as ChatGPT search.

For a brand, the consequence is direct. When an engine grounds its answer, what it says about you depends on the pages it retrieves at that moment, not only on what the model learned months earlier. Those pages can be yours, a competitor's or a review site's.

How grounding works

Google gives a plain definition: grounding means "providing content from the Google Search index to the model at prompt time to improve factuality and relevancy." The Gemini API documents the steps of Grounding with Google Search:

Analyze the prompt. The model decides whether a Google Search can improve the answer. If it judges that it can answer from memory, it does not search.
Search. It writes one or several search queries and runs them, the same idea as query fan-out.
Synthesize. It reads the results and writes the answer.
Attribute. It links parts of the answer to the pages that support them. These become the citations.

Consumer products work the same way. Google says AI Overviews and AI Mode rely on retrieval-augmented generation, "also known as grounding", using its core Search ranking systems to retrieve pages from its index. OpenAI says ChatGPT "may search the web automatically when your question would benefit from current information." Microsoft says Copilot is powered by Bing's search index.

Because retrieval goes through a search index, grounding inherits search rankings. A page that does not rank for the searches the model runs is unlikely to be retrieved, and a page that is not retrieved cannot be cited.

Example: what a grounded answer contains

Developers can see grounding at work. In the Gemini API's generateContent interface, a grounded response carries a groundingMetadata field: webSearchQueries lists the searches the model ran, groundingChunks lists the web sources with their URL and title, and groundingSupports links each segment of the answer to the sources that support it. Google's newer Interactions API returns the same information as search call steps and url_citation annotations.

Take a prompt a buyer might send:

Which accounting software do freelancers in Canada recommend?

The metadata could look like this (illustrative): two searches such as "best accounting software for freelancers Canada" and "FreshBooks vs QuickBooks Canada", then five sources, including two comparison articles, a forum thread and a vendor's pricing page. In groundingSupports, the sentence that recommends a tool points to a comparison article, not to the tool's own website.

That last detail is the useful one. The brand named in an answer and the page cited for it can be different, and the page cited is the one that earned the brand its place.

Grounding vs training data

Grounding and training data are the two places an AI answer can come from, and a brand reaches them in different ways.

	Training data	Grounding
When content is read	Before the model is released	At the moment of the prompt
Freshness	Stops at the model's knowledge cutoff	As fresh as the search index
Sources shown to the user	None	Cited pages
OpenAI crawler involved	GPTBot	OAI-SearchBot

How to get your pages used in grounded answers

Rank for the searches the model runs. Grounding retrieves through a search index, so rankings for the main query and its sub-queries come first.
Let the right crawlers in. OpenAI asks sites to allow OAI-SearchBot to be eligible for ChatGPT search. For AI Overviews and AI Mode, Google applies the same robots.txt rules for Googlebot and the same snippet controls as for Search.
Keep key facts in text. Google lists "making sure that important content is available in textual form" among the SEO basics that still apply to its AI features.
Find the sources that win instead of you. Hawkscan lists the pages cited by ChatGPT, Gemini, Microsoft Copilot and Google AI Mode, how many answers cite each one and whether it mentions your brand. A page that cites your competitors but not you is a grounding source worth winning.

Common questions

Is grounding the same as retrieval-augmented generation?

Almost. Google's own guide describes retrieval-augmented generation (RAG) as "a technique (also known as grounding)." In practice, RAG names the method, retrieving documents and then generating text, while grounding names the result: an answer tied to sources a user can check. Grounding can also use other data than web pages, such as Google Maps in the Gemini API.

Does blocking Google-Extended remove my site from AI Overviews?

No. Google-Extended only manages whether Google may use your content to train future Gemini models and to ground answers in Gemini Apps and in Grounding with Google Search on Vertex AI. Google states that it "does not impact a site's inclusion in Google Search." AI Overviews and AI Mode follow Googlebot's rules and snippet controls such as nosnippet.

Does every AI answer use grounding?

No. In the Gemini API, the model first decides whether a search would improve the answer, and ChatGPT searches automatically when a question "would benefit from current information." An answer that skips the search comes from training data alone and carries no citations, so it cannot send anyone to your site.