What Are AI Overviews? How Google's AI Answers Work

How does Google build an AI Overview?

The mechanism Google describes for both AI Overviews and AI Mode is query fan-out. Rather than answering from the top ten results for what you typed, the system decomposes your question into a set of related sub-queries, runs them in parallel, and retrieves the most relevant passages across all of those result sets. Those passages are then merged into a single attributed answer.

Google's own AI features documentation frames this as an advantage for publishers: the fan-out issues multiple related searches and can therefore surface a wider and more diverse set of helpful links than a single ranking would.

Three consequences follow, and they matter more than any tactic:

1. The unit of retrieval is the passage, not the page. A 3,000-word guide can be cited for one well-scoped paragraph buried in the middle. Sections that answer one question cleanly are individually retrievable; a section that only makes sense after reading the previous four is not.
2. You compete on sub-queries you never targeted. The fan-out generates questions adjacent to the original. Coverage of the surrounding question space matters more than exact-match phrasing of one head term.
3. Ranking first is not the qualification it used to be. Analyses through 2026 consistently report a falling overlap between the classic top ten and the set of pages cited in AI Overviews.

What does Google actually require?

Very little, and it is worth quoting because the market around this topic is loud with the opposite claim. Google's documentation says plainly:

There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.

And, on the recurring suggestion that you need new files or markup:

You don't need to create new machine readable files, AI text files, or markup to appear in these features.

That second sentence is the one to remember when someone sells you an AI-readiness package. The entry requirement is that the page is indexed and eligible for a normal snippet. Everything downstream of that is ordinary quality work: the page has to be worth quoting.

How much traffic do AI Overviews cost?

Be careful with numbers here, because most circulating figures come from vendor panels with undisclosed methodology. The most solid public dataset is from the Pew Research Center, which tracked real browsing behaviour rather than modelling it.

What Pew measured (March 2025)	With AI summary	Without
Clicked a traditional result	8%	15%
Clicked a link inside the summary	1% of visits	—
Ended the browsing session on that page	26%	16%

The study covered 900 consenting US adults and 68,879 Google queries, of which 12,593 triggered an AI summary — about 18% of searches at that time. Coverage has grown substantially since; estimates for 2026 commonly land somewhere between a third and half of queries, though these vary by keyword set and none are Google-published, so treat any single figure as an indication rather than a fact.

The honest summary: on queries where an AI Overview appears, expect roughly half the click-through you used to get. What that does to your total traffic depends entirely on how many of your queries trigger one, which is measurable in your own Search Console impression-versus-click trend and nowhere else.

What controls do you have?

The controls are the old robots preview rules. They are blunt, and the trade-off is real.

Control	What it does	The catch
nosnippet	No preview text may be used from the page	Removes your normal search snippet too
data-nosnippet	Excludes a specific HTML element	Element-level, so usable for surgical exclusions
max-snippet:[n]	Caps preview text at n characters	The same cap applies to ordinary snippets
noindex	Removes the page from Search entirely	Removes it from everything
Google-Extended	Opts out of training and grounding in Gemini apps	Does not affect Google Search or AI Overviews

The important line in that table is the last one. Google-Extended is frequently misdescribed as an AI Overviews opt-out; it is not. It governs training and grounding for Gemini apps, and blocking it costs you nothing in Search rankings — which is exactly why it is a reasonable default if you want to limit training use while staying visible.

There is no control that removes you from AI Overviews while preserving your normal snippet. If someone tells you otherwise, ask them to name the directive.

What should you actually change?

Almost nothing structural, and that is the point. The work that helps a passage get retrieved is the same work that has always helped a reader:

• Answer the question in the first two sentences of the section that promises it. Passage retrieval rewards self-contained answers; a paragraph that begins "as we saw above" is not extractable.
• Use headings that are the question, not a clever label. "How much traffic do AI Overviews cost?" retrieves; "The elephant in the room" does not.
• Include the specifics — numbers, dates, named sources. The GEO study from Princeton and IIT Delhi, presented at KDD 2024, tested nine content strategies across roughly 10,000 queries and found that adding citations, quotations, and statistics raised visibility in generative answers by up to 40%. That is one of very few controlled results in this space, and it points at substance rather than markup.
• Keep the page fast and stable. Retrieval requires successful crawling and rendering; see Core Web Vitals explained for the field metrics that matter to real users either way.
• Be citable as a source. Named authors, dated updates, a corrections policy, real evidence. The same signals that make a website look trustworthy to a person are what make it worth quoting.

What not to do

Two failure modes are worth naming, because both are actively harmful.

Don't publish volume to chase AI visibility. Google's spam policies define scaled content abuse as generating many pages "for the primary purpose of manipulating search rankings and not helping users," and name generative AI used to produce pages "without adding value for users" as a specific example. The policy is neutral about production method: AI-assisted is fine, AI-produced-at-volume-with-nothing-to-say is not.

Don't buy the file-and-markup story. Special AI text files, "AI-optimised schema," and similar products are sold as prerequisites for AI visibility. Google's documentation says no new files or markup are needed, and the field data on the best-known of those proposals is damning — see does llms.txt do anything for the numbers.

The uncomfortable conclusion is that there is no lever here. AI search rewards pages that contain a genuinely good answer to a specific question, from a source a system is willing to name. That has always been the job.

Sources: Google Search Central — AI features and your website · Google Search Central — Spam policies · Pew Research Center — Google users are less likely to click links when an AI summary appears · Aggarwal et al., GEO: Generative Engine Optimization, KDD 2024