The short answerGemini is Google's AI assistant, built on DeepMind models. Google says it drafts several responses, pulls in sources such as Google Search, runs a safety check and shows the highest-ranked draft. In the Gemini API, the model decides whether to search. Gemini 3.5 Flash, which Google made the AI Mode default in May 2026, has a January 2025 knowledge cutoff.
How Gemini answers a question, in five steps
| Step | What happens | What it means for your brand |
|---|---|---|
| 1. It reads your prompt | Gemini uses "the post-trained LLM, the context in the prompt and the interaction with the user" to start a response. | Your brand name in a prompt is only context. Without a source, Gemini answers about you from training data. |
| 2. It decides whether to search | In the Gemini API, the model "determines if a Google Search can improve the answer." Google doesn't document the app's trigger. | If the model thinks it already knows, it may not look you up. Gemini 3.5 Flash, which Google made the AI Mode default in May 2026, has a knowledge cutoff of January 2025. |
| 3. It writes and runs queries | Gemini writes "the query Gemini uses to invoke these external tools." In the Gemini API, the model "automatically generates one or multiple search queries and executes them." | Your page has to match the sub-questions Gemini writes as well as the words the buyer typed. |
| 4. It drafts, checks and ranks | Gemini drafts several responses, runs a safety check on each, then ranks the rest "based on their quality." | You can't touch the ranking step. You can shape which sources it draws on. |
| 5. It shows the answer | Sources appear under a Sources button or inline. "Not all responses include related links or sources." | Gemini can name you without linking you. |
The same engine family sits behind several Google products, and each one sources answers its own way, as do ChatGPT and Perplexity (all as of November 2026).
| Surface | Model (as of Nov 2026) | When it searches | How it shows sources | What our tracking saw (26 Oct to 8 Nov 2026) |
|---|---|---|---|---|
| Gemini app | Flash-family models; Gemini 3.6 Flash went to all app users, Gemini 4 Argon to Google AI Pro and Ultra | Google hasn't documented it for the app. The API leaves it to the model. | Sources button, inline links, or none | Not tracked |
| AI Overviews | Gemini 3 (default) | Google hasn't documented it. The Overview appears only when Google judges it "additive to classic Search"; may use query fan-out | Supporting links | Web results in 96.3% of answers; median 6 pages cited |
| AI Mode | Gemini 3.5 Flash (default from May 2026) | May use query fan-out across subtopics | Supporting links | Not tracked |
| ChatGPT | See our ChatGPT explainer | Searched in 97.2% of answers we tracked | Sources list; cited 13.4% of listed pages | Median 15 pages listed, 2 cited |
| Perplexity | See our Perplexity explainer | Searched in every answer we tracked | Citations; cited 29.6% of listed pages | Median 10 pages listed, 3 cited |
Our figures come from LoudFace's tracking in Peec of 96 B2B SaaS agency-buying questions, 1,344 answers per engine (method at the end). Peec does not track the Gemini app, so every first-party number here about Google describes AI Overviews.
What is Gemini?
Three things share the name: an app, a family of models, and the engine inside Google Search.
The app. Google calls it "an interface to a multimodal LLM (handling text, audio, images and more)." Google "initially launched Gemini (then called Bard) as an experiment in March 2023."
The models. Google DeepMind builds them. The May 2026 Gemini 3.5 launch post was bylined by DeepMind's leadership, including Jeff Dean.
Gemini in Search. Google Search has run custom versions. In March 2025 AI Mode launched on "a custom version of Gemini 2.0." In May 2025 Google brought "a custom version of Gemini 2.5" into Search "for both AI Mode and AI Overviews in the U.S." When a buyer reads an AI Overviews answer, Gemini wrote it, just not the Gemini they chat with in the app. The app, AI Overviews and AI Mode can give one buyer three answers with three source lists.
Is Gemini generative AI? Yes. It writes new text for every prompt. Google says LLMs "are neither information databases nor deterministic information retrieval systems," so the same question asked twice can get two different answers.
How does Gemini build an answer?
Google's Gemini overview is Google's own step-by-step account of the pipeline. The page was last updated on July 25, 2024 and still mentions "Gemini Advanced," a plan name Google has since replaced.
Training comes first. The models "are pre-trained on a variety of data from publicly available sources," filtered for quality, then fine-tuned on "carefully selected examples of excellent answers" and on human ratings of competing responses.
For each prompt, Gemini drafts. It uses the model, the prompt and the conversation "to draft several versions of a response."
It pulls in outside sources. The drafts can rely on "external sources such as Google Search, and/or one of its several extensions, and recently uploaded files." Google calls this "retrieval augmentation."
Every draft gets a safety check. Each potential response is checked against "predetermined policy guidelines."
The survivors are ranked. "The remaining responses are then ranked based on their quality, with the highest-scoring version(s) presented back to the user." Outputs carry a SynthID watermark.
Google also lists where this goes wrong: "the query Gemini uses to invoke these external tools, how Gemini interprets the results returned by the tools, and the manner in which these returned results are used to generate the final response." Read that as a map of where your brand gets lost. A query that never matches your page. A misread of your page. A summary that drops you.
Don't ask Gemini how Gemini works when you test any of this. Google's help center says it "can misrepresent how it works," including how it cites sources.
When does Gemini search Google?
In Google's developer API, when the model decides a search will help. Google documents that for developers only. For the Gemini app, it's our inference.
The Grounding with Google Search docs describe the API flow. The model "determines if a Google Search can improve the answer." If it does, it writes one or more queries and runs them. Billing follows the same logic: developers pay "for each search query that the model decides to execute," with 5,000 free search requests per month shared across all Gemini 3 and newer models, then $14 per 1,000. Google's Vertex AI example shows one prompt about Albert Einstein turning into three queries.
Google once exposed this decision as a number. Its October 2024 "dynamic retrieval" setting scored each prompt from 0 to 1 and searched above a threshold, where "the default threshold value is 0.3." Current models make the call themselves.
For the app, Google's overview says only that Gemini "relies on external sources such as Google Search." No Google page we read states when the app searches. Our read: the app most likely works like the API, searching when the model judges its own knowledge isn't enough. Google hasn't confirmed it.
The search decision also sets how current an answer is. Google says "Gemini 3.5 Flash has a knowledge cutoff of January 2025." In May 2026 Google made that model the default for the Gemini app and AI Mode. For anything after January 2025, Gemini 3.5 Flash has to search to know about your company. If your pricing or product changed since, and the model decides it doesn't need to search, it describes the old version of you.
Where does Gemini get its information?
From training data, from Google Search at answer time, and from whatever the user connects.
Google Search. The live source Google names first. Google says Gemini "is grounded in Google's understanding of authoritative information." On subjective questions, like which agency or tool is "best," Gemini "should respond in a way that reflects a wide range of viewpoints." Your own page is one viewpoint. Gemini is meant to reflect more than one.
The user's own material. Links can point to "files you have uploaded, or documents or emails from Google Workspace if you've connected to Workspace." Gemini can also show when it used your past chats.
Deep Research. This agentic mode "can automatically browse up to hundreds of websites" and builds "a personalized multi-point research plan."
Reddit. In February 2024 Google said it "now has access to Reddit's Data API," with "enhanced signals" to "display, train on, and otherwise use" Reddit content. Google gave no price for the deal.
Which domains does Gemini cite most? Three third-party studies give different answers:
- Profound (vendor study, 15,155 brand setups tracked in May 2026): "Gemini's three most-cited domains, according to our data, are Reddit, YouTube, and Wikipedia."
- Semrush (2026 AI Visibility Index, 126 million US prompts, published June 26, 2026): Gemini draws "on a smaller pool of citations that includes Wikipedia, Reddit, and YouTube."
- BrightEdge (AI Catalyst, prompts across ten industries; sample size and dates not stated): "Approximately 26% of Gemini's citations come from government domains, academic institutions, and major industry institutional bodies combined. UGC and social content makes up only 0.2%."
Those findings don't line up. Different prompt sets and different ways of classifying a site are the likely reason. The safe conclusion is narrow: what Gemini cites depends on the question.
How does Gemini show its sources?
Sometimes, and not always in the same place.
The Gemini Apps help page says "you can find the Sources button at the bottom of the response or in-line throughout the response." It also says: "If you don't have the Sources button below a response, Gemini Apps didn't provide any links for that particular response."
Google promises a link when a quote runs long: "If Gemini Apps directly quotes a large amount of text from a webpage, you'll see a link to that webpage in the sources list." Short of that, a mention can come without a link.
How many sources does an answer carry? The two big studies differ:
- Profound says "Gemini cites 6.6 sources per answer, against AI Mode's 15.2, and names about 4.4 brands" (May 2026, 15,155 brand setups).
- Semrush says "Gemini cites an average of 3 sources per response," against 15 for ChatGPT (126 million US prompts, June 2026). It adds that on Gemini, "the overlap between mentioned brands and cited domains can be as low as 30%."
On Gemini, being named and being cited are separate events. Your brand can sit in the answer while a competitor's site, or a forum thread, gets the link.
Does Gemini use the same sources as AI Overviews and AI Mode?
Not reliably. Google says even its two Search surfaces differ. Its guidance for site owners states: "AI Mode and AI Overviews may use different models and techniques, so the set of responses and links they show will vary." The Gemini app adds a third set.
The models differ by surface. In January 2026 Google made "Gemini 3 the new default model for AI Overviews globally." In May 2026 Gemini 3.5 Flash became "the new default model in AI Mode for everyone globally." In November 2025 Google announced that Search would route its most challenging AI Mode and AI Overviews questions to the Gemini 3 frontier model while "continuing to use faster models for simpler tasks," rolling out to Google AI Pro and Ultra subscribers in the U.S.
Retrieval differs too. AI Mode and AI Overviews both "may use a query fan-out technique," which Google defines as "issuing multiple related searches across subtopics and data sources." In AI Mode, Deep Search "can issue hundreds of searches." Our explainer on fan-out queries covers what that does to tracked prompts.
Outside studies measure the gap:
- BrightEdge found that "Gemini's top-100 citation overlap with AI Mode is only 27%, and with AI Overviews only 34%. Gemini actually has more in common with ChatGPT (39% overlap)." AI Mode and AI Overviews overlap at "roughly 59%."
- Profound: "Only about a third of the companies named in one showed up in the others."
What we saw in Google AI Overviews
These numbers describe AI Overviews, because Peec can't see the Gemini app. Across 1,344 AI Overviews answers to 96 B2B SaaS agency-buying questions, 26 October to 8 November 2026:
- It almost always used the web. 96.3% of answers carried web results, against 100.0% for Perplexity and 97.2% for ChatGPT.
- It cited nearly everything it listed. A median of 7 pages listed and 6 cited per searched answer. Overall, 95.8% of listed pages were cited, against 29.6% for Perplexity and 13.4% for ChatGPT.
- It wrote mid-length answers. A median of 297 words, between Perplexity's 55 and ChatGPT's 677, naming a median of 3 tracked companies.
- It drew on a wide pool. 985 distinct domains across 8,700 cited pages. 52 answers cited nothing.
Where those 8,700 citations went, by source type:
| Source type | Citations in AI Overviews | Share |
|---|---|---|
| Company and agency websites | 6,306 | 72.5% |
| YouTube | 924 | 10.6% |
| Reddit and forums | 471 | 5.4% |
| 366 | 4.2% | |
| Unclassified sites | 202 | 2.3% |
| Publishers and editorial sites | 174 | 2.0% |
| Other user-generated sites | 124 | 1.4% |
| Other reference sites | 76 | 0.9% |
| Review sites | 37 | 0.4% |
| Institutional sites | 20 | 0.2% |
| Wikipedia | 0 | 0.0% |
YouTube took 10.6% of AI Overviews citations and Reddit and forums 5.4%, while Perplexity and ChatGPT cited YouTube, Reddit, forums and Wikipedia zero times in the same window. Google-owned domains took 10.7% of AI Overviews citations, almost all of it YouTube, and 40.0% of AI Overviews answers cited at least one.
And Wikipedia got nothing. Profound and Semrush both list Wikipedia among Gemini's top domains. Our data covers a different surface and a narrow set of agency-buying questions, so both readings can hold. Plan for both.
The engines also barely shared sources. When AI Overviews and ChatGPT both cited something for the same question on the same day, only 4.6% of AI Overviews' cited domains were also cited by ChatGPT, and 74.9% of those pairs shared no domain. The figure for Perplexity was 6.5%, with 69.0% of pairs sharing nothing.
AI Overviews also moved day to day. Its cited domains for the same question on consecutive days scored an average overlap of 0.243 (0 means no shared domain, 1 means the same set), against 0.389 for Perplexity and 0.162 for ChatGPT. We looked at the same effect in does ChatGPT give the same answer to everyone.
Which models power Gemini in 2026?
As of November 2026, per Google's release notes:
- Gemini 4 Argon, released on October 13, 2026, is "rolling out globally for Google AI Pro & Ultra subscribers in the Gemini app." Google calls it "our newest frontier model."
- Gemini 3.6 Flash, released on July 21, 2026, is "an upgraded version of our Flash model," brought "to all Gemini app users globally."
- Gemini 3.5 Flash began rolling out "to everyone globally" on May 19, 2026.
Your plan decides your model. Google's plan page for the United States lists:
| Plan | US price | Model access (Google's wording) |
|---|---|---|
| Free | $0/month with a Google Account | "Access to our Flash-Lite model" |
| Google AI Plus | $4.99/month | "Access to our Flash model" |
| Google AI Pro | $19.99/month | "Access to Gemini 4 Argon," plus the Pro and Flash models |
| Google AI Ultra | From $99.99/month (a $199.99 tier also exists) | "Highest access to Gemini 4 Argon" |
Google's pages don't agree on the free tier. The plan page says Flash-Lite, while the release notes say 3.6 Flash went to all app users. Our read: free users get Flash-family models, and Google hasn't said which one. The free tier does include Deep Research.
A founder on the free plan and a CMO on AI Pro can ask the same question and get answers from different models.
Can a website opt out of Gemini?
Yes, without leaving Google Search. The control is a robots.txt token called Google-Extended.
Google's crawler documentation says it manages whether your content "may be used for training future generations of Gemini models" and for grounding in Gemini Apps and Grounding with Google Search on Vertex AI. Three details matter:
- It doesn't touch Search. "Google-Extended does not impact a site's inclusion in Google Search nor is it used as a ranking signal in Google Search."
- It has no crawler of its own. It "doesn't have a separate HTTP request user agent string." Google's existing crawlers read it.
- It blocks grounding as well as training. Google's overview says "Gemini will also not use such content for grounding."
Google also says it "intends to support the AIPREF shared machine-readable training preference signal" in early 2027.
AI Overviews and AI Mode follow Search rules. A page must be "indexed and eligible to be shown in Google Search with a snippet," and "There are no additional requirements." You don't need "new machine readable files, AI text files, or markup," or any special schema. To limit what Search shows, Google points to robots.txt rules for Googlebot and to nosnippet, data-nosnippet, max-snippet or noindex.
So a B2B SaaS that blocks Google-Extended stays in AI Overviews but drops out of grounded answers in the Gemini app. If your buyers research in Gemini, that's an own goal. Check your robots.txt before you spend anything on content.
How does a B2B SaaS get named in Gemini?
Gemini names you from what it learned in training or what it retrieves. For Gemini 3.5 Flash, which Google made the AI Mode default in May 2026, training data stops at January 2025, so for anything current, retrieval decides.
Let Gemini ground on you. Remove any Google-Extended block. Keep your key pages indexed with a snippet, which is also all AI Overviews and AI Mode require.
Answer the sub-questions. Gemini writes its own search queries, and in the Gemini API the model can generate "one or multiple" for a prompt. AI Mode and AI Overviews may fan a question out across subtopics. A buyer asking for the best AI search agency for B2B SaaS may set off searches on pricing, proof and comparisons. Your pages need to rank for those pieces. Our AI Mode explainer and our guide to Google AI Overviews optimization go deeper on the Search side.
Get onto other people's pages. On "best" questions, Gemini is meant to reflect "a wide range of viewpoints." On the agency-buying questions we tracked, AI Overviews cited company and agency websites most, then YouTube, Reddit and forums, and LinkedIn. That is your working list: comparisons on other companies' sites, a YouTube video that answers the question, real threads on Reddit, and posts on LinkedIn.
Put the answer at the top. Gemini links a page when it quotes "a large amount of text" from it. A clear, sourced answer in the first lines is the text most likely to be lifted.
Track each surface on its own. BrightEdge put Gemini's citation overlap with AI Overviews at 34%. In our data, AI Overviews and ChatGPT shared only 4.6% of cited domains on the same question and day. Winning one surface tells you little about the next.
LoudFace runs this work for B2B SaaS teams through its generative engine optimization service. The free AI visibility audit checks where your brand stands across ChatGPT, Claude, Gemini and Perplexity against your top competitors.
How did we measure this?
Every first-party number here comes from our LoudFace project in Peec, pulled live on 9 November 2026. Peec asks each tracked question once a day per engine from a US location. The project tracks B2B SaaS agency-buying questions. We dropped the 4 that name LoudFace, leaving 96. Over 14 days, 26 October to 8 November 2026, that gave 1,344 answers per engine on ChatGPT, Perplexity and Google AI Overviews.
Peec does not track the Gemini app, so every first-party figure describes AI Overviews.
A listed page is an entry in an answer's sources list. A cited page is one the answer attaches a citation to. Source types follow Peec's classification. Day-to-day overlap is the average Jaccard score of cited domains for the same question on consecutive days, over 1,178 day pairs for AI Overviews, 1,248 for Perplexity and 564 for ChatGPT. Cross-engine overlap uses 877 AI Overviews and ChatGPT pairs and 1,292 AI Overviews and Perplexity pairs where both engines cited at least one domain.
The source mix describes agency-buying questions, not Google in general.
