How Google AI Mode picks the pages behind its answers

An AI Overview is a box sitting on top of the ten links. AI Mode is not a box. It is the whole surface, a thread you keep talking to, and the links it shows are the ones its models pulled while the answer was being written.

If you want the short definition, the glossary entry on AI Mode has it. This page is about the part that decides whether your URL is in the running.

One query in, many searches out

Google publishes more about this surface than most engines publish about theirs. Its Search Central guidance says both AI Overviews and AI Mode may use a "query fan-out" technique, issuing multiple related searches across subtopics and data sources to build a response (Google Search Central, AI features and your website, last updated 2025-12-10).

That changes the arithmetic of being found. A classic search matched one query against one results page. A fan-out breaks your buyer's question into subtopics you never targeted, then goes looking for pages that answer each piece.

Google frames the upside for site owners plainly in the same document: while responses are generated, its models identify more supporting web pages, which lets it show a wider and more diverse set of links than a classic web search. More slots, decided by a process you cannot see.

How it reads the web and decides what to link

The retrieval-then-write shape is the same one behind every engine profiled here, described in how AI engines decide what to recommend. What is specific to AI Mode is that it runs on Google's index and Google's crawling, not a separate pipeline you opt into.

Google states the technical bar directly. To be eligible 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, and there are no additional technical requirements (Search Central, last updated 2025-12-10). Google also notes AI Mode and AI Overviews may use different models and techniques, so the links they show will vary.

The surface itself keeps moving. Google said in May 2026 that AI Mode had passed one billion monthly users, that Gemini 3.5 Flash became its default model globally, and that a follow-up question from an AI Overview now flows into an AI Mode conversation across desktop and mobile (blog.google, "A new era for AI Search", 2026-05-19). Treat anything specific about this surface as true on the date it was checked, including this page, checked 2026-07-30.

What a page needs to be usable inside it

Eligibility is a floor, not a strategy. The readable part is the part you own.

• Crawlable by Googlebot, allowed in robots.txt and not blocked by a CDN or host. Google lists this first among its best practices for AI features.
• Important content available in textual form. A point that exists only inside an image or a script is a point the model cannot lift.
• A direct answer high on the page, phrased so a machine can quote one clean sentence rather than reconstruct one from four paragraphs.
• Structured data that matches the visible text. Google states the match explicitly; schema that contradicts the page is worse than no schema.
• Internal links, so a fan-out that lands on one page of yours can reach the rest.

Old way: rank the page for the query. New way: assume the question gets split into five you did not write for, and make each page answer its own piece unambiguously.

The controls people get wrong

There is a common mix-up worth naming, because it costs traffic silently.

Google-Extended is the control for limiting AI training and grounding in some of Google's other systems. It is not the switch for AI features in Search. Google states that because AI is built into Search, robots.txt directives for Googlebot are the control for how sites are crawled for Search, and that to limit what is shown you use nosnippet, data-nosnippet, max-snippet, or noindex (Search Central, last updated 2025-12-10).

So a site that blocked Google-Extended thinking it opted out of AI answers did not opt out of AI Mode. And a site running an aggressive max-snippet to protect its content may have narrowed what AI features can show from it.

The damaging admission

Citedon does not query AI Mode. The scan runs your page's likely questions against four engines, ChatGPT, Perplexity, Gemini, and Claude, and reports which of them named you. AI Mode is not one of the four, so nothing we return is evidence about it.

Our robots.txt check also looks at named AI crawler tokens, including Google-Extended. For AI Mode specifically, the token that matters is Googlebot, and that is a Search Console question more than a Citedon one.

What we can measure for this surface is the readability layer: whether the page is reachable, whether the answer is in text a machine can extract, whether the structure and schema say the same thing the page says. That is the part Google's own guidance keeps pointing at, and the part that is yours.

The automated fix is WordPress-only. On Shopify, Wix, Webflow, or a headless stack, the scan still diagnoses the gaps, but you apply the changes yourself.

Why this one drifts faster than most

Two clocks run against you. Google is rebuilding this surface in public, with model swaps, new entry points, and agentic features announced in batches, so the behavior you observed last quarter is not a stable baseline. And your own site grows, with new pages shipping without the structure your fixed pages got.

There is a reporting wrinkle too. Google says sites appearing in AI features are folded into overall search traffic in Search Console, reported within the Web search type (Search Central, last updated 2025-12-10). You are not going to find a clean AI Mode column to watch. The measurable thing on your side stays the same: can a machine reach this page, and can it lift a clear answer from it.

How to check whether Google's AI surfaces can read your page

Take the page you would most want a conversational answer to lean on, and test the one thing that is testable.

Run a free scan on that URL to see how an engine parses it today and which structural pieces are missing. If you want to see the shape of the readout before you run one, there is a full example report, and the plain definition of the surface itself lives in the glossary entry on AI Mode.