How Gemini Chooses What Sources to Include
Gemini answers from two places: its trained knowledge and live Google Search grounding. When it grounds, it runs real searches, pulls candidates from Google’s index, and binds each cited source to the exact claim it supports. Five steps, from question to cited answer.
Decide whether to search at all
Not every prompt triggers retrieval. Gemini first judges whether live web results would improve the answer or whether its own training knowledge is enough.
A retrieval classifier scores the prompt
The model scores how much the query would benefit from fresh information, from 0 to 1. If the score clears a retrieval threshold (about 0.7 by default in the API), grounding activates. Current events, prices, local facts, and specific claims score high; creative writing and general knowledge score low.
Below the threshold: no sources
If Gemini answers from model knowledge alone, no retrieval happens and no citations appear. Everything that follows in this diagram only applies to grounded answers.
Break the question into search queries
Gemini rarely runs the user’s prompt as a single search. It decomposes the question into narrower sub-queries, a process called query fan-out, and executes real Google searches for each.
Query fan-out
“Best CRM for a small law firm” might fan out into searches on CRM comparisons, legal-industry requirements, pricing, and reviews. Each sub-query retrieves its own complementary sources, so one answer can draw on pages that never compete for the same search result.
Real searches, visible in the API
These are genuine Google Search executions. The Gemini API even returns the exact queries the model ran alongside the answer, so developers can see how the question was decomposed.
Build the candidate pool from Google’s index
This is what separates Gemini from other AI assistants: its retrieval backend is Google Search itself. The candidate pool is shaped by the same ranking signals covered in the page-ranking diagram.
Search ranking is the input
Pages Google cannot crawl or ranks poorly rarely enter the pool. But grounding is not just the top 10: because of fan-out, a page ranking well for one narrow sub-query can be cited even if it is invisible for the main query.
Knowledge Graph entities
Gemini leans on Google’s entity infrastructure. Brands, people, products, and places it recognizes as entities get their associated domains weighted up on queries about them. A brand with a Knowledge Panel and consistent entity signals is easier to cite.
Beyond web pages
The pool is not limited to articles: YouTube videos, Maps and local data, and other Google-indexed surfaces can be retrieved when they fit the sub-query.
Quality filters carry over
Search’s quality systems (E-E-A-T signals, spam filtering, helpful-content scoring) shape the pool before Gemini ever reads a page, so low-quality or spammy pages are mostly filtered out upstream.
Select sources claim by claim
As Gemini writes the answer, it decides which retrieved sources actually support each statement. Citations are attached to specific spans of text, not to the answer as a whole.
Semantic fit to the claim
A source is chosen because a passage in it directly supports a sentence being generated. Pages that state facts cleanly and self-containedly are easier to attach a citation to than pages where the claim is spread across paragraphs.
Diversity and corroboration
The synthesis layer prefers a spread of independent sources over many pages from one site, and weighs agreement between sources when claims conflict. Fresh pages win for time-sensitive claims.
Section-level extraction
Gemini can cite a single section of a long page. Clear headings where each block answers one question give a page more citable surfaces than one long undifferentiated essay.
Bound to exact text spans
Under the hood, each citation records which characters of the answer it supports. The API exposes this as grounding chunks (source URLs and titles) plus supports mapping response segments to source indices.
Show the sources
The same selection logic surfaces differently depending on where Gemini is running.
Gemini app
Grounded answers show inline link chips next to supported claims and a sources panel listing the pages the answer drew from, with click-throughs to the originals.
AI Overviews and AI Mode
Inside Google Search, Gemini-powered summaries cite with inline linked text that expands into source attributions, sitting above the traditional results and tied tightly to the live SERP.
Gemini API
Developers receive the structured grounding data (sources, text-span supports, and the searches that were run) and must display a Google Search entry point alongside grounded answers.
Want your pages to be the ones that get found and cited?We build SEO, AI search optimization and content strategy into every site we design.
Get a Free Quote