How does Gemini grounding with Google Search work, and how do sources surface?
Last updated October 14, 2026. Written by the Third Coast AI team.
Gemini grounding works by having the model read a prompt, decide whether a Google Search would improve the answer, generate one or more search queries of its own, run them against Google's index, and then write an answer in which specific passages are tied to the pages that supported them, so a company surfaces as a source in Gemini when its pages are in Google's index, answer one of the queries the model generated, and are clear enough to be tied to a sentence in the response. Google documents the mechanism for developers using the Gemini API, and the consumer Gemini app shows the result as a Sources button and in-line links, though not every answer carries sources, because the model does not search every time.
This page walks through the documented workflow, how sources appear in the Gemini app, what the Google-Extended control does and does not affect, and how Gemini differs from AI Overviews and AI Mode even though all three draw on the same index. The AI Mode vs AI Overviews page covers the other two.
Does Gemini search the web every time?
No, and that single fact explains most of what confuses people about Gemini citations. Google's grounding documentation for the Gemini API describes grounding as a feature that "connects the Gemini model to real-time web content," and lays out the workflow in steps: the model "analyzes the prompt and determines if a Google Search can improve the answer," then "automatically generates one or multiple search queries and executes them," then processes the results and writes a grounded response. The decision to search sits with the model, so the same buyer question can produce a grounded answer with sources on one run and an answer from training data with no sources on the next.
The Gemini Apps privacy hub puts it in a sentence: "some Gemini responses are grounded on Search results." Some, not all. When we run a client's prompt bank in Gemini we run it with grounding on and we run each prompt several times, because a prompt that triggers a search three times out of five is a different measurement problem from one that always does, and reporting a visibility rate without knowing how often the model searched at all would be misleading.
How do sources appear in the Gemini app?
As a Sources button and as in-line links, when they appear at all. The Gemini Apps help page on sources says that when sources are available "you can find the Sources button at the bottom of the response or in-line throughout the response," that selecting it opens a side panel with the relevant links, and that "not all responses include related links or sources," adding that if there is no Sources button "Gemini Apps didn't provide any links for that particular response." The same page notes that if Gemini "directly quotes a large amount of text from a webpage, you'll see a link to that webpage in the sources list," which is the one case where a citation is guaranteed rather than discretionary.
On the developer side the structure is more explicit, and it tells you what the app is doing underneath. Google's grounding documentation for the Gemini API says the response includes the search queries the model executed, a set of search suggestions whose display is governed by Google's terms of service, and annotations in which "each url_citation annotation links a text segment (defined by start_index and end_index) to a source URL." In other words, a citation in Gemini is a span of the answer tied to one page, and the page earns the citation by having supplied that span. A page that states a fact in one clear sentence is easier to tie to a span than a page that implies it over three paragraphs.
| Step (Google's documented workflow) | What happens | What it means for your page |
|---|---|---|
| Prompt analysis | Model "determines if a Google Search can improve the answer" | Some runs never search; measure across repeated runs |
| Query generation | Model "generates one or multiple search queries and executes them" | Your page must answer the model's rewrites, not the buyer's exact words |
| Search results processing | Results from Google's index are read and matched to the draft | Index presence and snippet eligibility are the gate; no separate crawler |
| Grounded response | Text segments tied to source URLs; Sources button and in-line links in the app | One clear sentence per fact is easier to cite than a diffuse page |
Which crawler feeds Gemini, and what does Google-Extended control?
Gemini grounding draws on the same index as Google Search, crawled by Googlebot, and Google-Extended is not a crawler at all. Google's common crawlers list says Google-Extended "doesn't have a separate HTTP request user agent string," that "crawling is done with existing Google user agent strings" and the token "is used in a control capacity," and that it manages whether crawled content "may be used for training future generations of Gemini models that power Gemini Apps and Vertex AI API for Gemini and for grounding." The same page says Google-Extended "does not impact a site's inclusion in Google Search nor is it used as a ranking signal in Google Search," and Google's AI features documentation separately places AI Overviews and AI Mode outside its scope.
So a company that blocks Google-Extended keeps its Search rankings and its AI Overview eligibility, and gives up being used for Gemini training and, by Google's wording, for grounding in Gemini Apps and Vertex AI. We read that as a real trade-off for a brand that wants to be cited in the Gemini app, and we would not recommend the block for a commercial site whose buyers use Gemini, though it is a legitimate choice for a publisher whose concern is training rather than visibility. Nothing about the decision affects Googlebot's ordinary crawl, which Vercel's December 2024 crawler study measured at 4.5 billion fetches in a month against 569 million for GPTBot, a reminder of how much more of the web Google sees than anyone else.
Gemini is the engine where clients most often tell us they are invisible, and half the time the real finding is that the model never searched. We run every prompt several times with grounding on and report two numbers: how often it searched, and how often we were cited when it did.
Third Coast AI
Is being cited in Gemini different from being cited in AI Overviews?
Same index, different selection. AI Overviews answer a query on a results page, with Google's documentation saying no additional requirements apply beyond being indexed and snippet-eligible, and Ahrefs measured 38% of their citations coming from the top ten in March 2026. Gemini answers a conversation, decides for itself whether to search, and writes its own queries, which in our runs look more like the hidden sub-queries AI Mode issues than like the typed phrase. SparkToro's January 2026 study of 2,961 runs found Google's AI returning the same brand list for a repeated question less than 1 time in 100, and Gemini's discretionary searching adds a second source of variance on top of that. Sistrix's May 2026 citation drift study measured AI Overviews replacing about 5% of their sources each week and AI Mode about 56%, and while Sistrix did not measure the Gemini app, our repeated runs put it nearer the AI Mode end. The practical answer is to treat Gemini as the most conversational of the Google surfaces: the pages that win there are the ones that answer the follow-up a careful buyer asks, in one sentence a model can tie to a span, with the proof beside it.
- Keep every important page indexed and snippet-eligible, since Gemini grounding has no crawler of its own and draws on Google's index.
- Do not block Google-Extended if you want to be used for grounding in the Gemini app; it does not affect Search or AI Overviews either way.
- State each fact in one clear sentence with a number or a named source, so a citation span can be tied to it.
- Write for the model's queries, which means covering the sub-questions behind a buyer's question in separate passages; the decision factors page explains how we find those.
- Measure with grounding on and across repeated runs, and record how often the model searched at all. The measurement guide covers the method.
Frequently asked questions
No. Google's help page says not all responses include related links or sources, and the API documentation says the model decides whether a Google Search would improve the answer. When it does not search, there is nothing to cite.
Google says Google-Extended controls whether crawled content may be used for training Gemini models and for grounding in Gemini Apps and Vertex AI, and that it does not affect inclusion or ranking in Google Search. Blocking it is therefore a trade of Gemini visibility for training opt-out.
No. Google's documentation says Google-Extended has no separate user agent and crawling is done with Google's existing agents, so allowing Googlebot is what keeps your pages available to Gemini grounding.
In the API, each citation ties a text segment of the answer to a source URL, and the Gemini app shows this as in-line links and a Sources button. A page earns a citation by supplying the sentence the model used.
They share Google's index but not the selection step. Gemini decides whether to search and writes its own queries inside a conversation, while AI Overviews summarize a results page. Google says its AI features may use different models and techniques, so links vary.