What does grounding mean in AI?

Basing an answer on content retrieved at answer time - web search results, a connected tool or a feed - instead of only on what the model learned in training. Grounded answers usually show their sources.

What is an ungrounded answer?

An answer built only from the model's training data. It can be out of date and usually shows no sources.

How can my store appear in grounded answers?

Let the AI search crawlers reach your pages, serve the content in the HTML, and keep structured data and feeds current. Grounded answers can only cite what the assistant can fetch.

Does Google-Extended affect grounding?

Yes, for Gemini apps and Vertex AI: it controls whether Google may use your content for Gemini training and grounding. It does not affect Google Search or AI Overviews.

Should I measure grounded and ungrounded answers separately?

Yes. They measure different things - what can be found now versus what the model remembers - so mixing them hides both.

Canonical definition

Grounding in AI Answers – Definition

An AI answer is grounded when the assistant builds it from content it fetched for that question – web search results, a connected tool or a feed – rather than only from what it learned in training. Grounded answers usually cite their sources. For a store, grounding decides whether current prices and pages can appear at all.

In one sentence

Grounded answers read the web now; ungrounded answers remember it – and a store needs different things for each.

Grounded and ungrounded answers

	Grounded	Ungrounded
Based on	Content fetched at answer time: web search, a connected tool, a feed	What the model learned in training
Freshness	As current as the source	As old as the training data
Sources shown	Usually, as citations or links	Usually not
What decides if your store is in it	Search crawler access, current and readable pages, feeds	Whether your content was in the training data, and how often you are mentioned elsewhere

Most assistants mix both. The same question can get a grounded answer in a search mode and an ungrounded one in a plain chat.

What it means for a store

Grounded answers depend on access. If OAI-SearchBot, Claude-SearchBot or PerplexityBot cannot reach your pages, grounded answers cannot cite them. See GPTBot vs OAI-SearchBot.
Ungrounded answers depend on reputation. Training crawlers (GPTBot, ClaudeBot) and mentions on other sites shape what a model “knows” about your brand. Blocking training is a valid choice, but it can make the model know less about you.
Google-Extended is about grounding, too. It controls whether Google may use your content for Gemini training and for grounding answers in Gemini apps and Vertex AI. It does not affect Search or AI Overviews.

Grounding and measurement

When you measure AI share of voice, keep grounded and ungrounded runs apart. They answer different questions: grounded runs show whether your current pages are found; ungrounded runs show what the model remembers. The gap between the two is often the most useful number. angeo/module-aeo-brand-visibility can query providers with search grounding on or off for this reason.

Grounded does not mean correct

A grounded answer can still misread a page or pick a weak source. Grounding tells you where an answer came from, not that it is right.

Questions

What does grounding mean in AI?: Basing an answer on content retrieved at answer time – web search results, a connected tool or a feed – instead of only on what the model learned in training. Grounded answers usually show their sources.
What is an ungrounded answer?: An answer built only from the model’s training data. It can be out of date and usually shows no sources.
How can my store appear in grounded answers?: Let the AI search crawlers reach your pages, serve the content in the HTML, and keep structured data and feeds current. Grounded answers can only cite what the assistant can fetch.
Does Google-Extended affect grounding?: Yes, for Gemini apps and Vertex AI: it controls whether Google may use your content for Gemini training and grounding. It does not affect Google Search or AI Overviews.
Should I measure grounded and ungrounded answers separately?: Yes. They measure different things – what can be found now versus what the model remembers – so mixing them hides both.

Sources

Checked 24 September 2026 against Google’s Gemini API grounding and crawler documentation and OpenAI’s crawler documentation. Disclosure: we build the brand-visibility module referenced above.