Gemini

Gemini as an answer surface, defined

Gemini is Google's family of generative AI models and the consumer app built on them; treated as a search surface, it is a place where a person asks a question and receives a written answer that may, but need not, carry links to web pages. Those links come from grounding — at answer time the model issues queries against the Google Search index, retrieves current pages, and attaches them to its response as related sources, which is what separates an answer reflecting the live web from one composed entirely out of model weights.

Three distinct things travel under the name, and conflating them produces most of the bad advice written about it. The Gemini app is the consumer product at its own address. Gemini in Chrome is the same assistant surfaced inside the browser, extended through 2026 to browse and act on a user's behalf. Grounding with Google Search in the Gemini API and in Vertex AI, Google's platform for building and running machine-learning applications, is the developer product — and the only place Google exposes the searches a Gemini model actually ran.

The distinction that matters most to a publisher is simpler. Gemini is not part of Google Search. It is a separate product with separate documentation, a separate publisher control, and no Search Console reporting of any kind. Everything established about AI Overviews or AI Mode is established about something else. Google said on 11 August 2026 that more than a billion people use the Gemini app every month, up from 750 million reported in February 2026 — first-party figures, unaudited, with no definition of a monthly user published and no subscriber split given.

Grounding is a tool call, and the model decides whether to make it

The clearest first-party evidence about how Gemini uses the web sits in a billing paragraph. Google's developer documentation, last updated on 26 August 2026, states that under Gemini 3 a project is billed for each search query the model decides to execute, and that if the model runs several searches to answer one prompt, that counts as several billable uses of the tool.

Two facts fall out of that sentence, and neither appears in Google's marketing. Searching is a decision the model makes rather than a step it always takes, so an answer may rest on no retrieval at all — which is one reason a brand can be described accurately by Gemini and never see a citation. And a single answer can rest on several searches nobody typed, the same shape as the query fan-out documented for AI Mode, established here independently and for a different product because Google had a charge to explain.

A fact published to justify an invoice is worth more than a fact published to sell something, because the company has a commercial reason to state it exactly. That is also the limit of it: this is documented for the API. Extending it to the consumer app is a reasonable inference and not a documented fact, and this page keeps the two apart because almost nothing else written on the subject does.

The API enumerates what the app conceals

A grounded API response carries three things a site owner would give a great deal to see. google_search_call contains the queries the model executed. google_search_result carries search suggestions for rendering in an interface. And annotations holds url_citation entries with start and end indexes, mapping a span of generated text to the source URL behind it.

So Google can enumerate a model's actual sub-queries and map each sentence of an answer to the page that supports it. It surfaces all of that to developers billed per query, and to no one else. Nothing comparable exists in the Gemini app, in Search Console, or in any free Google tool. The capability exists; access to it is a product decision.

Meanwhile the consumer app cites inconsistently by design. Google's help page on related sources says they may include public websites, files a user has uploaded, or documents and emails from Google Workspace, and states plainly that not all responses include related links or sources. No criterion is given for when they appear. The claim that Gemini always cites its sources is contradicted by Google's own documentation, and it spread by transfer: the developer product is sold partly on its promise to provide citations, grounded API responses do carry them, and the guarantee gets carried across to a consumer product that never made it.

The silence in Google's publisher documentation is the finding

Google's search documentation set does not cover Gemini at all. Its page on AI features and your website is scoped by its own title and text to AI Overviews and AI Mode; it mentions neither the Gemini app nor grounding. There is no Gemini eligibility rule — not even the indexed-and-snippet-eligible line that governs the two Search surfaces. There is no stated selection criterion, no measurement surface, and no publisher guidance of any kind.

That absence is the most important thing on this page, because it means every guide to being cited in Gemini is inference without exception. Two prerequisites are real and documented: be indexed by Google, and do not block Google-Extended. Past those two, nothing is documented and nothing has been independently replicated. GEO — generative engine optimization, the practice of trying to influence whether AI systems cite you — is nonetheless sold for this surface as confidently as for surfaces where an eligibility rule at least exists.

The evidence usually offered is a single agency study run between November 2024 and March 2025 across 1,000 Gemini queries, yielding roughly 4,004 cited URLs from 1,399 domains. It is the traceable origin of the widely quoted claim that FAQ markup lifts citations by 112%, a figure now attributed to AI Overviews about as often as to Gemini. Its authors state that their findings are correlational rather than causal. It has no control group, it is single-vendor, and it predates Gemini 2.5, Gemini 3 and Gemini 3.5 entirely. An eighteen-month-old correlational study of a model generation that no longer runs cannot support a claim about how Gemini selects sources in August 2026. Google's own generative-AI optimization guide of 15 May 2026, whose myth-busting section tells publishers that structured data is not required for generative AI search, is scoped to Search rather than to Gemini — which, if anything, leaves the Gemini case weaker still.

Google-Extended: the right lever here, the wrong lever everywhere else

Google's crawler documentation, last updated on 14 July 2026, scopes Google-Extended to whether crawled content may be used for training the Gemini models behind Gemini Apps and the Vertex AI API, and for grounding, which it defines as providing content from the Google Search index to the model at prompt time. It adds that the token does not affect a site's inclusion in Google Search and is not used as a ranking signal there.

Practitioner writing states both consequences backwards. Blocking Google-Extended does affect Gemini app grounding; it is the correct and only lever for this surface. Blocking it does not affect AI Overviews or AI Mode, which are governed by the search generative AI control in Search Console, live since 17 June 2026 — a control whose help page mentions Gemini nowhere and points readers back to Google-Extended for model training.

The net effect of that confusion is two mirrored errors. Publishers who wanted out of AI Overviews blocked Google-Extended, stayed in AI Overviews, and left Gemini instead. Publishers who genuinely wanted out of Gemini reached for the newer and more prominent Search Console control and remained in Gemini throughout. Both groups documented the work as completed.

Note also what Google does not commit to. It does not state whether an already-trained model retains material crawled before a block, and it publishes no timing for when grounding stops — in contrast to the Search Console control, where content is said to drop out within a day or two. This is a robots.txt token and a paragraph of documentation, which is the whole of the arrangement between Google and publishers on this surface; elsewhere in the field the terms have been set through litigation instead, as when News Corp sued Perplexity in October 2024.

Nothing measures this, and saying so is the honest answer

Gemini appears in no Google reporting surface built for site owners. The Generative AI performance report in Search Console covers AI Overviews and AI Mode; a companion report covers Discover. Gemini is in neither, and no plan for one has been announced.

What remains is server logs and analytics: sessions arriving with a Gemini referrer. That undercounts anything opened in a way that strips the referrer, and it cannot see a citation nobody clicked — which, given that Google says not every response carries links, may be most of them. Statcounter's measurement of global AI chatbot referrals in March 2026, reported on 3 April 2026, put ChatGPT at 78.16%, Gemini at 8.65%, Perplexity at 7.07%, Microsoft Copilot at 3.19% and Claude at 2.91%, with Gemini having risen from 2.31% a year earlier.

Set that against a billion monthly users and the tension is instructive rather than contradictory. A product at that scale contributing that share of measured referral traffic implies very few outbound clicks per user, which is consistent with Google's statement that not all responses include sources. Nobody has measured the link rate directly, and until somebody does, a site owner reasoning from Gemini referral counts is reasoning from the small visible remainder of an unmeasured whole.

There is also no repeated-sampling study of Gemini. Every sibling surface tested that way has proved volatile — AI Mode returned only 9.2% URL consistency across three same-day parses of the same keywords — so a single Gemini pull should be treated as a single draw, and a screenshot of a brand appearing in one answer as evidence of one answer. What would move this page's verdict is a study of Gemini citations with a published methodology, an adequate sample, repeated sampling and a holdout, or a Gemini dimension in Search Console. Neither exists as of 29 August 2026, and the grounding documentation remains the only place Google describes the mechanism at all.

Frequently asked questions