What is generative engine optimization (GEO)?
How AI assistants choose which sources to cite, and what you can change.
On this page
Short answer
Grounding is when an AI model bases its answer on sources it looks up at the moment of the question, such as search results, map data or a company’s own documents, instead of relying only on what it absorbed in training. A grounded answer usually shows links to those sources. Google’s Vertex AI documentation lists four kinds of sources to ground a model in, from Google Search and Maps to a company’s own data. For a business, it means the live pages on your site, and not only what the model once read, can end up in an answer.
How Google defines it
Google’s Vertex AI documentation describes grounding as the ability to connect model output to verifiable sources of information. It names the main benefit as fewer hallucinations, meaning fewer statements that are not factual. Google offers several kinds of sources for developers who build on its models:
- Google Search, for general knowledge from the web;
- Google Maps, for answers about places and locations;
- a company’s own data, through managed retrieval or an existing search index such as Elasticsearch;
- other search APIs and web indexes.
The Gemini API documentation on grounding with Google Search walks through the steps. The model first decides whether a search would improve the answer. If yes, it writes one or more queries and runs them. Then it reads the results, builds the answer and returns it with citations tied to the exact parts of the text they support.
How OpenAI describes search
OpenAI’s developer documentation for its web search tool tells a similar story in different words. The model can choose to search or not, depending on the prompt, and a developer can force a search. The answer comes back with citations that carry the URL, the page title and the place in the text where the source is used. A separate list records every URL the model consulted. OpenAI requires apps that show these results to make the citations visible and clickable.
Both vendors agree on the main point: in a grounded answer, the model writes from pages it just read. That is where your site comes in.
Grounded or not: how to tell
| What you see | What it likely means | What you can do |
|---|---|---|
| answer with links to pages | a search ran, the answer is grounded | check that the linked page states your facts correctly |
| answer with no links in ChatGPT | it may come from training alone | ask again with search on and compare |
| wrong fact with a link | the source page is wrong or outdated | fix your page, or ask the other site to correct theirs |
| wrong fact, no link, same with search on | an outdated page of yours is probably still online | search your site for the old fact and update it |
| Perplexity answer | Perplexity’s help center says every answer carries numbered citations | open each cited source |
Google’s Vertex AI documentation describes grounding as the ability to connect model output to verifiable sources of information.
What it means for a service business
Grounding turns your website into live input for AI answers. This has three practical effects.
- Old pages keep working against you
A price list from 2023 that is still indexed can be read and quoted today.
- Plain facts win
The model looks for a sentence that supports a claim. A page that states “implants from $X, two to three visits” gives it one.
Read more
A page that says “affordable, high-quality care” gives it nothing.
- Fixes show up faster
Training data changes only with a new model release. Grounded answers change once the source page is updated and read again.
We saw the third effect on our own site. With Russian-language queries in Almaty, Google AI Overviews, which build answers from Search results, linked to aoe.kz in 0 of 17 tracked queries on September 22, 2026. By October 6 it was 2 of 20, after three queries were added to the set. The first links showed up within a week of a page rewrite. ChatGPT, meanwhile, did not name us more often. Each assistant searches in its own way, so we track them one by one.
What to check on your site
- One live page per fact
Prices, address, hours, service list. Remove or redirect older copies.
- Facts in text
Not only in images, PDFs or widgets that load late.
- A date on pages that change
It helps a reader, and a model, judge which version is current.
- Search bots allowed
OpenAI says sites that opt out of OAI-SearchBot will not be shown in ChatGPT search answers.
Read more
Perplexity runs PerplexityBot for its search results. Check robots.txt for both.
- Google-Extended set on purpose
This robots.txt token lets you opt out of Gemini training and of grounding in Gemini Apps and Vertex AI.
Read more
It is separate from Search, so it is not the switch for AI Overviews.
- Profiles outside your site
Directories, review sites and your Business Profile are grounding sources too. Keep their facts in line with the site.
Common mistakes
- “Grounded” taken to mean “correct”. Grounding reduces hallucinations but does not remove them. A model can still misread a correct page.
- Confusing grounding with training. Blocking a training bot does not remove you from grounded answers, and allowing it does not add you.
- Blocking Google-Extended to leave AI Overviews. AI Overviews are part of Search. This token does not take you out of them.
- Fixing the site but not the sources about you. If a directory lists an old phone number, a grounded answer can still repeat it.
Grounding turns your website into live input for AI answers.
FAQ
Is grounding the same as RAG?
RAG stands for retrieval-augmented generation, the engineering pattern of fetching documents first and writing from them. Grounding is the wider term vendors use for tying an answer to sources. Google lists RAG over your own data as one way to ground a model.
Can we make ChatGPT ground its answer on our site?
You cannot choose the sources for someone else’s question. You can make your pages easy to find, allowed for search bots, and clear enough to quote. Our AI citations guide shows how we check whether that works.
How does grounding relate to query fan-out?
Fan-out decides what to search for. Grounding writes the answer from the results. Read more in our query fan-out guide.
Where can I learn the other terms?
The AI search glossary has short definitions. For what models learn about you without searching, see our LLM SEO guide.
