Gemini SEO: get cited when the Gemini app searches Google

01 / 07

From prompt to source chip: how Gemini grounds an answer

Google defines grounding as “providing content from the Google Search index to the model at prompt time to improve factuality and relevancy,” and names the Gemini app as one place it happens (Google’s common crawlers). So there is no Gemini index to submit to and no Gemini crawler to court. The candidates are pages Googlebot has already crawled and indexed.

Google doesn’t publish how the consumer app decides when to search. Its developer documentation for the same capability describes the loop, and it matches what users see in the app:

1. 01

   Gemini decides whether a search would help

   The model “analyzes the prompt and determines if a Google Search can improve the answer.” Fresh, factual and comparison questions usually trigger it; drafting, brainstorming and small talk often don’t, which is why Google warns that “not all responses include related links or sources.”

   Your lever: Target questions that need current facts: prices, specs, comparisons, how-tos with versions and dates.

2. 02

   It writes its own searches

   If needed, the model “automatically generates one or multiple search queries and executes them.” They split the prompt into sub-questions, which is why Ahrefs found only 12% of links cited by ChatGPT, Gemini and Copilot ranked in Google’s top 10 for the original prompt (2025 data).

   Your lever: Cover the sub-questions around a topic on pages that each answer one well, rather than one thin page per keyword variant.

3. 03

   Google's index supplies the pages

   Results come from the Search index, with one filter Search itself doesn’t apply: pages whose robots.txt disallows Google-Extended can’t be used for grounding in Gemini Apps.

   Your lever: Be indexed (check URL Inspection) and leave Google-Extended allowed on anything you want cited.

4. 04

   The answer is written and its sources attached

   Sources sit “in-line throughout the response” or behind a Sources button at the bottom, which opens a side panel of links. If Gemini quotes a large amount of text from a page, that page is always listed. With Workspace connected, the list can include the user’s own Docs and emails too.

   Your lever: Write passages specific enough to quote: a number, a named method, a dated fact. Paraphrased common knowledge rarely earns a chip.

02 / 07

Google-Extended: the one switch that's Gemini's alone

Google-Extended isn’t a crawler. Google says it “doesn’t have a separate HTTP request user agent string”: Googlebot does the crawling, and the token only tells Google what it may do with what it fetched. Disallow it and your content can’t be used to train future Gemini models or for “grounding … in Gemini Apps and Grounding with Google Search on Vertex AI.” It “does not impact a site’s inclusion in Google Search nor is it used as a ranking signal.”

That bundling is the decision every Gemini strategy starts with. A publisher who blocks the token to keep content out of training also leaves the Gemini app’s citations, and there’s no setting that keeps one and drops the other.

• GooglebotSearch indexAllow

  Crawls and renders the pages in the Search index that Gemini grounds on. Blocking it removes you from Search, AI Overviews, AI Mode and Gemini alike.

  Honors robots.txt

• Google-ExtendedTraining + groundingYour call

  A robots.txt token you’ll never see in logs. Governs Gemini training and grounding in the Gemini app and on Vertex AI; no effect on Search or rankings.

  Honors robots.txt

Setting it: all, nothing, or by folder

Saying nothing is a choice too: with no Google-Extended group, Gemini may ground on, and train on, anything Googlebot can crawl. To leave Gemini entirely while staying in Search:

# Out of Gemini-app citations AND Gemini training.
# Search, AI Overviews and AI Mode are unaffected.
User-agent: Google-Extended
Disallow: /

The token takes path rules like any other group (Google’s own example allows one archive page and disallows the rest), so you can keep public guides citable and withhold the content you sell:

User-agent: Google-Extended
Allow: /guides/
Disallow: /members/
Disallow: /reports/

What each Google control switches off

The two AI opt-outs are mirror images. The Search Console control, rolled out to every site on 31 August 2026, removes pages from AI Overviews, AI Mode and Discover’s generative features and doesn’t mention Gemini. Google-Extended removes them from Gemini and leaves AI Overviews alone. Google’s help page sends you to Google-Extended only to limit the training of the models behind Search’s AI features, not their answers.

03 / 07

Gemini, AI Overviews and AI Mode: one index, three products

All three are Google products that ground answers in the same index, which is why advice for one gets pasted onto the others. They differ in where they live, which controls reach them and where their clicks land. Google has “brought together AI Overviews and AI Mode into one seamless Search experience”; the Gemini app is a separate product at gemini.google.com, in the mobile apps and in Chrome. For Search’s AI features, read the AI Overviews & AI Mode guide; this one stays with the app.

04 / 07

Fewer links, more names: what Gemini actually cites

Gemini is the most sparing linker among the big assistants. In Attrifast’s study of 1,200 buyer-intent prompts, each run three times on gemini.google.com (April–May 2026, logged out, mostly Gemini 2.5 Pro), Gemini cited a median of 2.4 unique domains per answer, against 3.1 for ChatGPT search, 3.6 for Claude and 6.4 for Perplexity. An agency panel of 11,028 answers per engine (July–October 2026, from tracked projects mostly in French) found Gemini attached any source to 68.4% of answers, against 96.3% for ChatGPT.

Semrush and Growth Memo measured the other side of that. Across 3,981 domain appearances from 115 prompts in 14 countries, Gemini named brands in 83.7% of appearances and linked them in 21.4%. It still recommends companies; it just doesn’t always say where it read about them.

Which sources win depends on the question

The studies disagree about Gemini’s favourite sources, and the disagreement is useful: each sampled a different kind of question.

• Forums and video, on everyday questions

  Observed

  Across 3M+ US queries on all topics (Ahrefs Brand Radar, September 2026), Reddit held 28.5% of the citation share among Gemini’s top 50 domains, YouTube 14.6% and Wikipedia 8.8%.

• Institutions, on industry questions

  Observed

  On BrightEdge’s industry prompts (finance, healthcare, B2B tech and more), 26% of Gemini’s citations went to government, academic and institutional sources (about 13% to .gov alone) and 0.2% to user-generated content.

• Vendor pages and YouTube, on buying questions

  Observed

  On buyer-intent prompts, Attrifast found vendor sites took 14.1% of Gemini’s citations, Wikipedia 12.1%, YouTube 9.7% and Reddit 8.4%. YouTube was cited about 2.5 times as often as on the other three engines.

• A concentrated pool

  Observed

  Gemini’s top 100 domains took 58.7% of its citations in the same study. The few links it gives keep going to the same sites, so a mention on one of them travels further than another page of your own.

• Google's quality bar decides who's in the pool

  Our read

  Grounding draws on Google’s index, so the candidates are pages Google’s systems surface. Google’s advice for its AI features, unique non-commodity content with first-hand insight, is written for Search, but nothing suggests the app’s pool differs.

• Why it names more than it links

  Our read

  Unlinked mentions plausibly come from what the model already knows and from third-party pages that discuss you. Google documents neither, but both point the same way: review sites, comparisons and YouTube are where to be named.

05 / 07

Deep Research, Chrome and agents: when Gemini reads your page itself

Not every Gemini visit is a grounded chat answer. Other paths reach your site, and they identify themselves differently, which matters for robots.txt, your WAF and your logs. Two have user agents of their own:

• Google-AgentUser-triggeredAllow

  Used “by agents hosted on Google infrastructure to navigate the web and perform actions upon user request.” IPs are published in user-triggered-agents.json, and Google is testing Web Bot Auth with the agent.bot.goog identity.

  Ignores robots.txt· User-triggered, so it generally ignores robots.txt.

• Google-GeminiNotebookUser-triggeredAllow

  Fetches URLs a user adds as sources in Gemini Notebook, called NotebookLM until July 2026. The old Google-NotebookLM token was supported until August 2026.

  Ignores robots.txt· User-triggered, so it generally ignores robots.txt.

The rest arrive without a Gemini label:

• Deep Research can “browse up to hundreds of websites” for one report, on Gemini 3, from a research plan the user can edit. Google doesn’t document its fetcher, so the safest assumption is the same index-plus-Google-Extended rule as chat.
• Gemini in Chrome sits in a side panel beside whatever tab is open and answers questions about that page. It reads what’s already in the user’s browser, so robots.txt never comes into it.
• Auto browse clicks, fills in forms and shops inside the user’s own Chrome, pausing for confirmation before purchases or posts. It launched for AI Pro and Ultra subscribers in the US in January 2026 and began reaching Android in the US at the end of June. Google documents no separate identifier, so expect it to look like that user’s browser.
• Gemini Spark, Google’s always-on personal agent, browsed in a remote browser at launch and since July 2026 can drive the user’s desktop Chrome through auto browse.
• Developers who pass your URL with the API’s URL context tool get the page “from an internal index cache” first, with a live fetch only as a fallback.

# 1. The user agent contains the token
Google-Agent

# 2. The IP is in Google's published list
https://developers.google.com/static/crawling/ipranges/user-triggered-agents.json

# 3. Or: reverse DNS ends in google.com or googleusercontent.com,
#    and a forward lookup returns the same IP
host <ip-from-your-logs>

06 / 07

Products and places: the feeds behind Gemini's shopping and local answers

For shopping and local questions, Gemini can lean on Google’s own data stores as much as on web pages, and those are fed directly, not crawled. Here’s what helps, what doesn’t, and what nobody has confirmed:

Indexed by GoogleRequired

The baseline for every grounded citation. Confirm with URL Inspection in Search Console.

Google-Extended allowedRequired

For Gemini-app grounding specifically. Check every hostname and any CDN-managed rules.

Merchant Center feedHelps

Google says Merchant Center and Business Profile “can help your products and services to be visible in both AI responses and other Google Search results.” Keep titles, GTINs, prices, availability and images complete.

UCP checkoutHelps

With the Universal Commerce Protocol, eligible listings can show a Buy button on Gemini and AI Mode, paid through Google Pay while you stay merchant of record. Select merchants in the US, Canada and Australia; only listings with native_commerce show it.

Google Business ProfileHelps

Gemini’s Maps grounding draws on “places, reviews, photos, addresses, opening hours” across 250M+ places. Keep hours, categories and attributes current, and reply to reviews.

Structured dataNo effect

Google says no special schema.org markup is needed for its AI features. Keep it valid for rich results; it isn’t a Gemini switch.

llms.txtNo effect

Google Search “doesn’t use them,” and nothing documents the Gemini app reading one.

Snippet controlsUnconfirmed

nosnippet and max-snippet are documented for Search features. Whether the Gemini app honours them isn’t stated.

07 / 07

Finding Gemini in GA4, your logs and a prompt panel

Search Console won’t help here: its Generative AI performance report covers Google Search and Discover, and the Gemini app is neither. Gemini is measured where its clicks land.

1. 1GA4’s AI Assistant channel. Since 13 May 2026, GA4 sets medium ai-assistant and campaign (ai-assistant) when the referrer matches a recognised assistant, and Google names Gemini among them. For sessions before that date, filter on session source instead.
2. 2The referrer. Clicks from the web app arrive as gemini.google.com / referral (bard.google.com in older data). Google files AI Overviews and AI Mode clicks under Organic Search instead.
3. 3UTM tags. Since late September 2026, practitioners have seen utm_source=gemini on some outbound links. Google hasn’t documented it, so treat it as a bonus signal, and never add Gemini UTMs to your own URLs.
4. 4Direct traffic. Clicks from the mobile apps and the Android assistant often lose the referrer; one practitioner put mobile pass-through at around 9%, unverified. Watch for Direct sessions landing deep on the pages Gemini cites.
5. 5Logs. Count verified Google-Agent and Google-GeminiNotebook hits. Googlebot hits tell you nothing Gemini-specific.

# Session source / medium: the Gemini web app
^(gemini\.google\.com|bard\.google\.com) / referral$

# Session source: also catches UTM-tagged links
^(gemini\.google\.com|bard\.google\.com|gemini)$

A prompt panel for what analytics can't see

With links on barely a fifth of brand appearances, most of your Gemini visibility never becomes a click. A small panel you can repeat shows it:

1. 1Write 20–30 prompts a buyer would actually type: comparisons (“Tallowpine vs Asana for Workspace”), recommendations (“best PM app for a five-person team on Google Workspace”) and how-tos.
2. 2Run each in a fresh signed-out session, then signed in. With Workspace connected, sources can include the user’s own Docs and emails, which changes the answer.
3. 3Record three things separately: whether you’re named, whether you’re linked (a chip or a Sources entry), and which domains were cited instead.
4. 4Run the panel at least three times over a week. Attrifast ran every prompt three times, 48 hours apart, because one run proves little.
5. 5Send the same prompts through the API’s google_search tool to see the searches Gemini wrote, and cover the sub-questions you’re missing.

The Gemini run-sheet

1. 01Confirm Google has indexed the pages you want citedHigh impactURL Inspection on your top pages. There’s no other way into Gemini’s grounding pool.
2. 02Decide on Google-Extended on purposeHigh impactAllowed keeps you citable in the Gemini app; disallowed drops grounding and training together. Use path rules if only some content should be withheld.
3. 03Read robots.txt on every hostnameHigh impactLook for a Google-Extended group a CDN toggle or plugin added. Check the apex, www and every subdomain.
4. 04Map the sub-questions behind your buyers' promptsHigh impactRun them through the API’s google_search tool to see the searches Gemini writes, then answer each one well on a page of its own.
5. 05Make key passages quotableMediumSpecific numbers, named methods, dated facts and short definitions near the top of each section.
6. 06Earn mentions on the pages Gemini already citesHigh impactReview roundups, comparison posts, YouTube reviews and forum threads in your category. Earn them; never fake them.
7. 07Publish YouTube videos that answer real questionsMediumAccurate titles, chapters and descriptions. YouTube turns up far more in Gemini’s citations than in other assistants’.
8. 08Keep Merchant Center and Business Profile completeMediumProducts: a full feed, plus UCP checkout if you’re eligible. Places: hours, categories, attributes and review replies.
9. 09Keep pages usable by agentsMediumNative buttons and links, labelled inputs, no overlays, stable layouts. Verify Google-Agent before your WAF blocks it.
10. 10Set up Gemini measurement in GA4High impactThe AI Assistant channel, a gemini.google.com source filter and a check for utm_source=gemini.
11. 11Run a monthly Gemini prompt panelMediumSigned out and signed in, three runs per prompt. Record named, linked, and who was cited instead.