Getting Cited in Gemini and Google AI Mode: Three Surfaces, Three Different Controls

Gemini is named in 18 other posts on this blog at the time of writing and explained in none of them — because "showing up in Gemini" isn't one target, it's three separate surfaces. AI Overviews and AI Mode, both inside Google Search, draw on the same index and the same Googlebot crawl. Gemini Apps, a separate product that sits outside Search entirely, run on their own training and citation logic. Treating all three as one "Gemini visibility" goal leads to the wrong tactics: a control that matters for one (allowing Googlebot in robots.txt) does nothing for another (Google-Extended). This post separates the three surfaces, maps each control to the surface it actually governs, and explains which of them GEON's own Gemini scan measures.

Three names, three products

AI Overviews is the summary box that appears above regular Google Search results. We've covered its structure in depth elsewhere and won't repeat it here: Google AI Overviews Reward Structure, Not Rank →.

AI Mode is a separate experience, a distinct tab inside Search. Google announced it on March 5, 2025 as a limited Labs experiment — a mode that "expands what AI Overviews can do with more advanced reasoning, thinking and multimodal capabilities." At launch it ran on a customized version of Gemini 2.0. Per Google's help page (as accessed in July 2026), AI Mode is now available in dozens of countries and languages and uses the same "query fan-out" technique: splitting a question into subtopics, searching each one in parallel, then combining the results.

Gemini Apps (gemini.google.com and the mobile apps, formerly Bard) aren't part of Search at all — they're a standalone conversational product. Unlike Search, they don't fire on every query automatically; the user talks to Gemini directly, and the model decides for itself whether to check the web.

AI Overviews and AI Mode: one index, one control

Google's own documentation deliberately collapses this pair: "to be eligible to be shown as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown in Google Search with a snippet... there are no additional technical requirements." Everything you do for AI Overviews eligibility — indexability, snippet eligibility, allowing Googlebot in robots.txt — automatically carries over to AI Mode. There's no separate opt-in, no separate meta tag, no separate crawler. The same document states both features may use the same "query fan-out" technique.

The same page describes the control mechanism: "AI is built into Search and integral to how Search functions, which is why robots.txt directives for Googlebot is the control for site owners... To limit AI training and grounding in some of Google's other systems, read more about Google-Extended." Google itself draws the line here between "Search's AI features" and "Google's other systems" — Gemini Apps and Vertex AI.

This matches what we wrote about robots.txt exactly: Google-Extended is a robots.txt product token, not a crawler, and it doesn't change Googlebot's crawling behavior. Google's own crawler list confirms it: "Google-Extended doesn't have a separate HTTP request user agent string. Crawling is done with existing Google user agent strings; the robots.txt user-agent token is used in a control capacity." Two separate posts on this blog verify the same underlying fact — no discrepancy.

The takeaway: if your goal is showing up in AI Overviews or AI Mode, Google-Extended is irrelevant to you. The only requirement is standard Search index eligibility.

Gemini Apps: outside Search, a different lever entirely

Gemini Apps operate on a completely different axis. Per the same crawler documentation, Google-Extended is a standalone product token that web publishers can use to manage whether content Google crawls from their sites may be used for training future generations of Gemini models that power Gemini Apps, and for Vertex AI's grounding on public web data. In other words, the token covers not just training future models but whether today's systems can ground themselves in your site in real time — a control layer that has no equivalent on the AI Overviews/AI Mode side at all.

One more distinction matters here, because it's commonly conflated: Google-Extended is a decision you make as a publisher about your website's content. Whether a Gemini user's own chat gets used to train models is a completely separate, account-level setting — "Keep Activity" — documented in the Gemini Apps Privacy Hub (last updated July 15, 2026). As a GEO team, the lever you control is the publisher-side Google-Extended decision; a user's chat data is a separate privacy matter outside your site's control.

Practical consequence: disallowing Google-Extended in your robots.txt changes nothing in Google Search — not your ranking, not your AI Overviews visibility, not your AI Mode visibility. What changes is whether your content trains future Gemini models and whether today's Gemini can ground itself in your site in real time.

Why citation behavior differs

In AI Overviews and AI Mode, citation is part of generation itself: the query fans out into sub-questions, each is retrieved separately. Google describes this as a retrieve-then-generate model — find the source first, then write on top of it — but that doesn't explicitly guarantee the source shown is the exact passage that fed the answer, only implies it.

Gemini Apps work differently. Per Google's help page, Gemini shows a "Sources" button below some responses; those links can include general web pages, files you've uploaded, or connected Workspace documents, all in one list. The same help center also documents a separate "double-check response" feature: it runs an independent Google Search after the response has already been generated and compares it against the statements — the source it surfaces isn't necessarily the one the model used to generate the answer in the first place. That's close to the inverse of AI Overviews' retrieve-then-generate order.

There's a third layer at the API level: Google's "Grounding with Google Search" tool connects the model to real-time web content. Per Google's own description, the model first decides whether a search would help, then generates and runs its own search queries, then synthesizes the results into a response with inline citations (annotations) — all current models, including Gemini 2.5 Flash, support this tool. That's a retrieve-then-generate shape close to AI Overviews' fan-out, but it happens inside a developer's API call, not the Search UI — it isn't automatically the same as the consumer Gemini app's default behavior or Search's AI Overviews/AI Mode.

What GEON's Gemini scan actually measures

Separating these three surfaces also clarifies GEON's own measurement. GEON's Gemini scan runs google/gemini-2.5-flash through OpenRouter with the platform's web plugin — the provider's own native search mechanism — so, unlike DeepSeek, it runs grounded (see functions/src/types/scan.ts, PLATFORM_MODELS). In practice, that's close to the "Grounding with Google Search" tool described above: the model searches when it judges it necessary, then synthesizes.

But the limit needs to be stated plainly: this is an API call — not Google Search's AI Overviews/AI Mode box, and not the consumer Gemini app. A brand getting a mention in a GEON scan doesn't mean that brand appears in AI Overviews or made it into the Gemini app's "Sources" list; the three are separate surfaces requiring separate measurement. What GEON measures is how Gemini's grounded API behavior positions your brand. Your AI Overviews/AI Mode visibility inside Search needs to be tracked separately, through Search Console's Performance report, where Google reports both under the standard "Web" search type.

Three steps for Monday morning

Keeping the three surfaces separate isn't about adding complexity — it's about applying the right control to the right surface.

1. For AI Overviews and AI Mode, meet the standard Search technical requirements: the page must be indexable, snippet-eligible, and Googlebot must not be blocked in robots.txt. There's no separate opt-in — both come with the same condition.
2. For Gemini Apps, make a deliberate Google-Extended call in your robots.txt; any of the three strategies from our robots.txt post — citation-friendly, selective, or full opt-out — applies here too.
3. Read GEON's measurement correctly: the Gemini scan measures grounded API behavior, not Search's AI boxes. Don't fold the two into one number.

If you want to track all three surfaces per platform in one place, GEON is built for exactly that.