How Google AI Overviews and AI Mode Work: How Sources Are Selected and Cited

Google AI Overviews and AI Mode answer with a Gemini model grounded in Google Search. The model generates related queries through query fan-out, Google's core ranking systems retrieve pages from the Search index, and the response links to pages that support it. A page is eligible if it is indexed and allowed to show a snippet. Google says no special markup, files or chunking are required.

This is part four of our series on how search and AI visibility work. It deliberately sticks to what Google itself documents, because most advice about AI Overviews mixes Google's statements with guesses. Where we add interpretation, we say so. For the practical enterprise playbook, read how to appear in AI Overviews and AI Mode. For what happens to clicks, read how much traffic AI Overviews take. The primary sources are Google's AI features documentation and its guide to optimizing for generative AI features, last updated in July 2026.

The short version

1. Trigger Google's systems decide whether an AI response would add value beyond classic results for this query.
2. Fan out The model generates related queries across subtopics to gather more information.
3. Retrieve Core Search ranking systems return relevant, current pages from the Search index.
4. Generate The model writes a response grounded in those pages and identifies supporting pages as it goes.
5. Link Google shows links to the supporting pages alongside the response.
6. Report Impressions and clicks are counted in Search Console under the Web search type.

AI Overviews and AI Mode: what Google says each one is for

Google positions the two features differently. AI Overviews help people get the gist of a complicated topic quickly and act as a starting point for exploring links. Google's Search Help page says they appear when its systems determine generative AI can be especially helpful, such as when someone wants to understand information from a range of sources. AI Mode is for questions that need deeper exploration, reasoning or comparison, and it supports follow-up questions. A searcher can move from an AI Overview into AI Mode by choosing to see more and asking a follow-up, with the original context carried over.

At AI Mode's launch in May 2025, Google said both AI Mode and AI Overviews in the U.S. used a custom version of Gemini 2.5. Model versions change, so treat the model name as a detail and the grounding design as the constant. Google also tells users that AI responses may include mistakes, and that the features cannot be turned off, though a Web filter shows text links only.

How AI Overviews and AI Mode select and cite sources, step by step

Step 1: Deciding whether to show an AI response

Not every query gets an AI Overview. Google's documentation says they show only when its systems determine they are additive to classic Search results. Google does not publish the criteria. The practical reading is that AI responses cluster on questions where synthesis across several sources helps, which describes much B2B research: comparisons, "how does X work", "what should I consider when choosing Y".

Step 2: Query fan-out

Google's guide defines query fan-out as concurrent, related queries that the model generates to request more information and fetch additional results. Its own example: a question about fixing a weedy lawn can spawn related searches about herbicides and weed prevention. At AI Mode's launch Google described the same technique as breaking a question into subtopics and issuing many queries at once, and said Deep Search extends it to hundreds of searches for a fully cited report.

For a B2B query such as "how to choose a managed SOC provider for a mid-size bank", fan-out might reach questions about regulatory expectations for banks, pricing models, in-house versus outsourced comparisons and response time commitments. The sub-questions are our illustration. The mechanism is Google's.

Step 3: Retrieval through core ranking systems

Each query is answered from Google's index using its normal machinery. Google calls this grounding: relying on core Search ranking systems to retrieve relevant, up-to-date web pages from the Search index to improve the quality, accuracy and freshness of AI responses. That means the systems described in Google's ranking systems guide, including passage ranking, reliable information systems and spam detection, shape which pages the model gets to read. There is no separate AI index and nothing to submit. Everything in how SEO works applies here.

Step 4: Generating the response and finding supporting pages

The model writes a response from what it retrieved. Google's AI features documentation adds one detail that matters for citations: while a response is being generated, the models identify more supporting web pages, which is why these features can display a wider and more diverse set of helpful links than classic web search. Put simply, links are attached to support what the response says, and the pool of candidates is bigger than the ten results for the original query.

Step 5: Showing links

Google presents supporting links with the response, and it adjusts the layout over time. Our piece on further exploration links tracks the current link surfaces. What Google documents is the purpose: links let people verify information and explore further, and Google says clicks from AI Overview results tend to be higher quality, with people more likely to spend more time on the site they visit.

Step 6: Counting impressions and clicks

Google's Search Console metrics documentation specifies how these features are counted. In an AI Overview, clicking a link to an external page counts as a click, standard impression rules apply, and the AI Overview occupies a single position, with all its links assigned that position. In AI Mode, clicks and impressions follow the same rules, position follows standard results page methodology, and follow-up questions are new queries counted separately. All of it appears in the Performance report under the Web search type.

What makes a page eligible, and what Google says is not needed

Google's AI features documentation lists three conditions. A page must be indexed, must be eligible to be shown in Google Search with a snippet, and must meet the Search technical requirements: Googlebot is not blocked, the page returns HTTP 200, and it has indexable content. Google says there are no additional requirements and no special optimizations necessary.

Its July 2026 optimization guide goes further and names things site owners do not need to do:

• No new files or formats. You do not need machine-readable files, AI text files, markup or Markdown. Google says adding them will neither harm nor help visibility in Search.
• No chunking. There is no requirement to break content into tiny pieces for AI to understand it.
• No special writing style. You do not need to write in a specific way for generative AI search.
• No required schema. Structured data is not required, and there is no special schema.org markup to add.
• No manufactured mentions. Pursuing inauthentic mentions is not effective, because core systems focus on quality and other systems block spam.

What Google asks for instead is valuable, non-commodity content with a distinct viewpoint and expertise, plus the basics in its AI features documentation: allow crawling in robots.txt and in your CDN or hosting setup, make content findable through internal links, provide a good page experience, put important content in text, support it with good images and video, make sure structured data matches visible text, and keep business information current.

Controls: how to limit or opt out

AI Overviews and AI Mode are Search features, so they use Search controls. Google's robots meta tag documentation says nosnippet prevents content from being used as a direct input for AI Overviews and AI Mode, and max-snippet limits how much content may be used that way. The data-nosnippet attribute marks parts of a page not to be used as a snippet.

Whole page, no snippet and no direct input to AI Overviews or AI Mode:

<meta name="robots" content="nosnippet">

Whole page, previews limited to 120 characters:

<meta name="robots" content="max-snippet:120">

One section only:

<div data-nosnippet>Customer-specific pricing terms...</div>

Two points trip teams up. First, these controls also affect regular snippets, so opting out has a cost in classic results. Second, Google-Extended is not an AI Overviews switch. Google says it governs Gemini training and grounding in Gemini Apps and Vertex AI, and does not affect inclusion or ranking in Google Search.

How to measure performance in Google's AI features

• Search Console Performance report. Filter by page and query under the Web search type. Expect AI Overview impressions to raise impression counts for pages that rarely earned clicks before, with a shared position.
• Generative AI performance report. Google's optimization guide recommends the generative AI performance report in Search Console, introduced in June 2026, for tracking visibility in these features. Our guide to reading it covers interpretation.
• GA4. Google's default channel group has an AI Assistant channel, but it explicitly excludes AI Overviews and AI Mode. Clicks from these features arrive through Google alongside other search traffic.
• Manual checks. Record which domains appear for your priority queries each month. Google's guide warns that no third-party tool has access to its internal ranking or AI systems, so tools that report AI Overview presence are sampling what they observe.

What you control: the AI Overviews checklist

• Priority pages pass URL Inspection: indexed, crawl allowed, indexing allowed, and your intended canonical selected.
• No sitewide nosnippet, short max-snippet or stray data-nosnippet added by a plugin, template or legal review.
• Googlebot is allowed in robots.txt and not challenged by your CDN or firewall.
• Pages answer the sub-questions behind your head terms, not only the head terms.
• Important facts live in HTML text, not only in images, video or gated PDFs.
• Content contains first-hand, non-commodity detail: named examples, your own data, clear conditions.
• Structured data, where used, matches what the page visibly says.
• Internal links connect pillar pages to the detail pages fan-out is likely to reach.
• Search Console exports and the generative AI performance report are reviewed monthly.
• Google-Extended decisions are made separately and documented, since they do not govern these features.

Illustrative example: a managed security provider missing from pricing answers

Illustrative example. The company, queries and findings below are invented to show how to apply Google's documentation. They are not a client result.

A managed security services provider ranks on page one for "managed SOC services" but never appears in the AI Overview for "how much does a managed SOC cost for a mid-size company", which cites two competitors and an industry blog. Its team works through Google's documented conditions:

1. Indexed? URL Inspection shows the pricing page is on Google with the intended canonical. Pass.
2. Snippet-eligible? The page source contains max-snippet:0, added by the legal team a year ago to keep pricing language out of search previews. That also removes the page as a direct input for AI Overviews. Fail. The team agrees to remove it and wrap only contract terms in data-nosnippet.
3. Covers the fan-out? The likely sub-questions are pricing models, what drives cost, and in-house versus outsourced comparisons. The page says only "pricing tailored to your needs". The team adds a section on pricing models (per endpoint, per user, tiered), the five factors that change cost, and a comparison table of in-house versus managed staffing requirements. Publishing actual prices stays optional.
4. Non-commodity? The team adds anonymized examples from its own deal desk and names the author, the company's head of commercial operations.
5. Measure. The team tracks impressions and clicks for the page in Search Console, noting that an AI Overview citation shares one position, and records monthly which domains the AI Overview cites.

Nothing in the fix involved AI-specific markup or files. It was a preview control someone forgot about, and a page that did not answer the questions Google's fan-out was asking.

Glossary

• AI Mode: Google Search's conversational mode for deeper questions and follow-ups, grounded in Search.
• AI Overview: a generated summary at the top of some Google results, with links to supporting pages.
• data-nosnippet: an HTML attribute that keeps marked text out of snippets.
• Deep Search: an AI Mode capability Google described as running hundreds of searches for a cited report.
• Google-Extended: a robots.txt token for Gemini training and grounding in Gemini Apps and Vertex AI; not a Search control.
• Grounding: Google's term for retrieving pages from its Search index to base AI responses on them.
• max-snippet: a robots rule that caps preview length and limits direct input to AI features.
• nosnippet: a robots rule that blocks snippets and direct use of content in AI Overviews and AI Mode.
• Query fan-out: the model generating concurrent related queries to gather more results.

Next in the series

Google has one crawler for Search. Other engines have several each, with different jobs. How AI crawlers work covers GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot and the rest, and how to control them. To check your own pages against this list, request a free audit or read the B2B AEO and GEO playbook.