AI Overviews: definition, and what they change for visibility

Ask Google a real question today and you often will not land on a list of links at all. A paragraph appears first, written for you, drawn from several pages, with a small panel of sources beside it. That block has a name, and it has quietly become one of the most contested pieces of real estate on the web. This page defines it, then shows you the machinery behind it. For the wider frame, our GEO agency pillar covers the discipline that grew up around these surfaces.

AI Overviews: the one-sentence definition

AI Overviews are AI-generated summaries that Google places at the top of the results page for some searches: instead of only listing links, Google writes a short answer drawn from several web pages and shows a panel of the sources it used.

Two words in that sentence carry the weight. Generated, because the text did not exist before you asked; a model wrote it. And sources, because the answer points back to a handful of pages, which is the only reason a business still has anything to play for. Google's own framing is that an Overview helps people get the gist of a complicated topic quickly and acts as a jumping-off point to explore links (Google Search Central, 2025). Read that as a promise and you will be disappointed. Read it as a new front door, one you either appear in or you do not, and you are thinking about it correctly.

How the surface is actually built

Google states that AI Overviews may use a query fan-out technique: the system issues multiple related searches across subtopics and data sources, then assembles a grounded answer from what comes back. The pages it genuinely used are surfaced as supporting links.

Fan-out is the detail most people miss, and it changes everything downstream. Your query is not answered once. It is broken into pieces, each piece is searched, and the answer is welded together from several retrievals (Google Search Central, 2025; Google, AI Mode in Search, 2025). That is why an Overview can cite five sources that have nothing to do with one another: each one won a different sub-question. Which is oddly good news. You do not have to be the best page on the topic. You have to be the cleanest available answer to one slice of it.

The eligibility rule is refreshingly blunt. To be shown as a supporting link, Google says a page must be indexed and eligible to appear in Search with a snippet. Block the snippet with nosnippet or max-snippet and you exclude yourself as surely as if you had blocked the crawler. I have watched a site do that by accident and spend a quarter wondering why it had gone quiet.

When an AI Overview appears

AI Overviews trigger far more often on long, question-shaped searches. In Pew Research Center's March 2025 browsing dataset, only 8% of one or two word searches produced one, against 53% for searches of ten words or more, and 60% for queries starting with a question word.

Those numbers draw the map of where the risk actually is (Pew Research Center, 2025). Your brand name, typed alone, is mostly safe. The long, spelled-out question a buyer types just before deciding is where the Overview lives. Google also says plainly that Overviews only show when its systems judge them additive to classic Search, and so often do not trigger. This is not a surface that has swallowed the SERP. It shows up precisely where the intent is richest.

And that is where our own numbers get uncomfortable. Of the 4836 French keywords Cicero Studio has analyzed, 34% draw fewer than 100 searches a month (Cicero Studio internal data). The long tail is not a rounding error, it is the bulk of the map, and it is made of exactly the queries that summon an AI Overview. Most teams optimise for the fat head and get summarised on the tail.

AI Overviews vs featured snippets

A featured snippet is an extract from a single page, and clicking it sends the user to that exact section. An AI Overview is model-written, synthesises several pages, and links out through a citation panel. One has a winner; the other has a shortlist.

People conflate the two constantly, and the confusion is expensive because the strategy diverges. Google's documentation is clear that clicking a featured snippet takes the user directly to the section of the page that appeared in it (Google Search Central, 2025). A snippet is a quote with an address. An Overview is a synthesis with footnotes.

What it changes for clicks and for your brand

Clicks fall when a summary is present. Pew Research Center tracked 900 US adults and found users clicked a traditional result in 8% of visits with an AI summary, against 15% without, while clicks on the summary's own source links happened in just 1% of visits.

Nearly half the clicks, gone (Pew Research Center, 2025). Google counters that clicks arriving from pages with AI Overviews tend to be higher quality, meaning users stay longer. Both things can be true at once, and probably are: fewer visitors, better ones. But notice what neither party disputes. The visit is no longer the unit of visibility. Being named is.

That reframing is what our audit work keeps running into. Across the 1207 SEO and GEO audits Cicero Studio has run (Cicero Studio internal data), the recurring problem is almost never a shortage of content. It is content the machines cannot cleanly quote or confidently trust. The business publishes; the machine reader cannot use what it publishes. That gap never shows up in a rankings report, which is exactly why it goes unnoticed for so long.

The limits, and the criticism

AI Overviews are unstable by design, occasionally wrong, and contested by publishers who attribute traffic losses to them. Treating them as a fixed, fair or permanent surface is the fastest way to build the wrong strategy on top of them.

Three criticisms are worth holding onto. The first is variability: the same question can produce a different answer, and a different source list, on a different day. Anything you measure here is a snapshot, not a rank. The second is accuracy: a synthesised answer can be confidently wrong, and the burden of that error lands partly on whoever it cites. The third is the publisher dispute, which Pew describes directly: publishers have attributed declining traffic to these summaries replacing traditional results. That argument is unresolved, and it runs alongside a regulatory push in Europe, where the AI framework leans on providers to be transparent about AI-generated content (European Commission, 2024). None of this makes the surface ignorable. It makes it something you observe honestly rather than something you promise results against.

Alexis Dollé

CEO & Founder of Cicero Studio

I test AI visibility the slow way, by hand, one real business question at a time, across hundreds of sites. What I keep finding is that the pages an AI Overview names are rarely the loudest ones. They are the clearest: a plain answer, stated early, backed by a source you can check. That is the standard we hold every page to at Cicero Studio.

LinkedIn →

How to think about being present

Since Google offers no special optimisation, presence comes from three plain conditions: be retrievable, be extractable, and be credible. The research on generative engines found that cited statistics and named sources raise visibility inside AI answers, while keyword stuffing does nothing.

That finding, from the original academic work on generative engine optimisation (Aggarwal et al., arXiv, 2024), is the least glamorous and most useful thing in this field. The machine rewards what a good editor always rewarded: answer the question early, say something specific, name your source. For the playbook rather than the principle, we wrote it up separately in getting cited by AI Overviews, and the cross-engine version lives in our method for appearing in ChatGPT and Google AI. The technical trap, blocked crawlers and content invisible to retrieval, is covered in AI crawlers and invisible websites. And if the vocabulary is still fuzzy, start with our GEO definition.

What this page does not cover

A definition is only as useful as its edges, so here is what I left out, and where each thing lives instead.

Going further

We publish our method in the open, because that is the only proof worth much in this business.