Cited Answers API: Facts with Sources for AI Agents

Send a question, get a short factual answer and the sources behind it. Each answer comes from a Gemini model with Google Search grounding switched on, so the claims are checked against live search results and every source is returned with the exact piece of the answer it supports.

Built for agents that make many small calls: one row per question, a stable JSON shape, and an answer that arrives with its citations or is not charged.

What you get per question

field	meaning
question, index	the question and its position in your list
answer	the answer text
grounded	true only when Google Search backed the answer with at least one citation
status	answered, ungrounded, failed or skipped
error	why a row is not answered, otherwise null
citations[]	the sources, see below
search_count	how many searches the model chose to run; null when none ran or the API did not say
model, provider	the model asked and the provider that answered (always Google on a charged row)
finish_reason	stop, or length when the answer was cut at maxAnswerTokens
billed	the pricing event this row was charged as, or null

Each citation:

field	meaning
url	the page the source points at, resolved from Google's redirect. null if it could not be resolved
grounding_redirect	Google's original vertexaisearch.cloud.google.com/grounding-api-redirect/... link, unmodified
domain	host of url without www.
title	the source name as Google gives it (usually the site's domain, not the page title)
start_index, end_index	UTF-8 byte offsets into answer, as Google reports them, not character offsets
supported_text	the slice of answer those offsets cover

supported_text is Google's segment, passed through as is. Segments are sentence-like and can start mid-token (a segmenter that treats Node. as a sentence end starts the next segment at js:). One span can have several sources, and the same page is listed once per distinct span.

Input

• questions: up to 20 per run, up to 1000 characters each.
• model: default google/gemini-3.5-flash-lite. The list is Gemini 3+ Flash and Flash Lite models that support search grounding.
• language, country: optional hints added to the instructions. They steer the answer, they do not filter sources.
• maxAnswerTokens: 100 to 2000, default 600.

What is charged

• A small actor-start event per run ($0.001).
• One grounded-answer event per question ($0.075) whose answer came back from Google with at least one citation.
• Nothing for a question that failed, came back from a non-Google provider, came back empty, or came back with no citation. Those rows are still returned, flagged, so you can see why.
• A run never asks a question it cannot bill: the remaining maximum charge is checked before each call.

Things worth knowing

• The number of searches is the model's decision. A simple fact usually takes one search, a comparison or multi-part question two to four, and a nonsense query three. Searches are the main cost of an answer. OpenRouter has no setting that caps them on a Google model (tested: max_uses, max_tool_calls and stop_server_tools_when were all ignored), and asking for a single search in the prompt did not reduce them, so the price is set against a measured spread, not a cap. Over 21 mixed questions the model ran 1 to 4 searches (median 2), and one repeated multi-part question ran up to 6.
• A question is not always grounded. The model decides whether to search. A greeting or "reply with OK" runs no search and is returned as grounded: false, free. Arithmetic still triggered a search in testing.
• Citations are returned as Google gave them. grounding_redirect is never altered. url is read from the first redirect hop with one request; the destination page is never fetched.
• Google's grounding terms apply to the grounded results. Read them before building on the output.