Gemini

Gemini is Google's family of multimodal models, sold through the Gemini API and used inside Google products, and — when grounding is switched on — capable of answering from live web sources and returning citations alongside the text. For a publisher, the second half of that sentence is the part that matters: an ungrounded Gemini answer cannot reach your page at all, and a grounded one can quote it and name it.

It is a family, not a model

Google's model directory lists a Gemini 3 series alongside a Gemini 2.5 series and separate audio, generative media, music generation, tool and agent, and specialised task models. Within the Gemini 3 stable line Google describes Gemini 3.8 Flash as engineered for long-horizon software engineering, autonomous agents and complex enterprise workflows, Gemini 3.7 Flash for complex coding, agentic workflows and reliable multi-step execution, and Gemini 3.5 Flash as providing baseline speed and foundational performance.

The practical consequence is that "Gemini said X" is an underspecified statement. Limits, modalities and behaviour attach to a model code, not to the brand, and Google publishes those per model page. Any claim about what Gemini can accept or produce should name the model code it was measured on.

Long context, and why it does not remove retrieval

Google's long-context documentation states that many Gemini models come with large context windows of 1 million or more tokens, and that Gemini was the first model capable of accepting 1 million tokens.

A window that large changes what fits in one call: an entire codebase, a long transcript, a full set of filings. It does not remove the selection step, because a corpus is unbounded and a window is not. Something still has to decide which documents enter it. That decision is the whole game for a publisher — see context window for what the limit is, and RAG for the architecture that does the choosing.

Grounding with Google Search

Grounding is the feature that connects a Gemini model to live web content. Google describes its purpose in three parts: increase factual accuracy by basing responses on real-world information, access real-time information to answer questions about recent events, and provide citations by showing the sources for the model's claims.

The response shape is what makes it auditable. A grounded call returns the search queries the model actually executed, an HTML snippet for rendering search suggestions, and inline url_citation annotations that tie segments of the generated text to specific source URLs. So the citation is not a courtesy the model may or may not extend — it is structured data the developer receives and chooses how to display.

Two things follow. First, being cited by a grounded Gemini answer requires being findable by the searches it ran, which are machine-chosen sub-queries and not the user's phrasing. Second, whether a reader ever sees your name depends on the application, because Google gives developers control over how sources are displayed in their interface.

Gemini in Search is a separate question

Google Search's AI Overviews and AI Mode are documented by Google Search Central, not by the Gemini API docs, and their eligibility rules are ordinary Search rules: a page must be indexed and eligible to be shown in Google Search with a snippet, fulfilling the Search technical requirements, and Google states there are no additional requirements nor other special optimizations necessary. Google also describes a query fan-out behind those features — multiple related searches issued across subtopics and data sources.

Google's AI-features documentation does not name the model behind those surfaces, so treating AI Overviews and the Gemini API as one system is a mistake that leads to optimising against the wrong controls.

Failure modes

Quoting a capability without a model code. Limits move between releases and differ across the family. A number attached to "Gemini" rather than to a model page goes stale invisibly.

Assuming the long window means everything gets read. It means everything can fit. What enters it is still a retrieval decision made before the model sees anything.

Expecting a visible citation. The API hands back annotations; the application decides whether to render them. Your source can be used in an interface that shows nothing.

Reaching for a Gemini-specific opt-out. For Search surfaces there isn't one. Google points at nosnippet, data-nosnippet, max-snippet and noindex, which also govern the classic snippet, so the control is a trade rather than a switch.

Writing for a phrasing the user typed. Under fan-out, the query that retrieves you is one the system composed. Sections that answer a single question completely are the ones that survive that indirection.