How AI Overview Works: A Technical Walkthrough of Google’s AI-Generated Answers

Google AI Overview (AIO) is the AI-written summary at the top of many Google results pages. It writes one answer from a small number of web sources and links to them. For most businesses it is the busiest AI search feature there is, and the most valuable place to be named. Few people know how it actually works: what makes it appear, how it picks sources, and how it writes the answer. This guide walks through it in four stages, updated with what we saw in our own checks up to November 2026.

The four stages are: the trigger (which questions get an AI Overview and which don’t), source selection (how Google picks the pages it reads), content extraction (what it pulls from each page) and answer generation (how the Gemini model writes the final answer). Knowing the stages gives your optimisation work something concrete to aim at, instead of a black box.

What have our own checks shown, up to November 2026?

Our checks show that AI Overview rewards being clearly understood across several sources, and that a citation can disappear as fast as it arrives. Here is what we saw first-hand, with dates.

• Fast in, fast out (August to September 2026). Qwang Coffee, Alva’s own store, registered its domain on 23 August 2026. It was indexed in 1 hour, on page one in 24 hours, and cited in a Google AI Overview in 48 hours. It was listed first on 3 of 5 commercial questions in Singapore.
• More sources, more naming (September 2026). For a branding agency client, we checked 40 Google AI Overview results. A business that appeared on one source was named 27% of the time. A business on three or more sources was named 87% of the time. Source selection is where that gap opens up.
• AI Mode answers differ by market (4 November 2026). We asked “best GEO agency” in English and in the local language from Vietnam and from Thailand in Google AI Mode. In Vietnam only one agency appeared in both answers. In Thailand, none did. The same question in a different language can pull a completely different set of sources.
• Our own product (early 2026). AeroChat, Alva’s own product, reached its first Google AI Overview in under three weeks. Impressions grew 343%, clicks 127%, and branded search share reached 74%, period over period.

None of this predicts results for any other site. It is consistent with the four stages below: the trigger and the sources move, so the work has to keep up.

What is AI Overview and where does it sit in the results page?

Google AI Overview is the AI-written summary that appears at the top of the results page for many searches, above the normal list of links. It usually gives an answer in a few paragraphs, with links to the sources it drew from, and an option to expand it. It launched in May 2024 (first as Search Generative Experience, then renamed AI Overview), spread across markets through 2024 and 2025, and by November 2026 is a stable part of a large share of Google searches in most major markets.

The layout puts the AI answer first, the source links inside or beside it, and the normal ten blue links below. For many informational searches the AI Overview is the first thing a person reads, and on mobile it can fill the whole screen. The behaviour change is real: many people get their answer from the overview and never click through, even when the source is linked.

The model behind AI Overview comes from Google’s Gemini family. It is not a standalone system. It sits on top of the same index Google has built for two decades, and most of the pages it reads come from that index.

Stage 1: Which questions trigger an AI Overview?

Not every question produces an AI Overview. Google decides per search whether to show one, based on the type of question, the intent, the quality of available sources and how confident the model is that it can give a useful answer. The reported trigger rate has moved a lot. Industry studies reported it on 60 to 80% of searches in some categories at launch, tuned down through 2024 and 2025 to around 15 to 30% across most niches by mid-2025, and it has kept changing since.

Question types that tend to trigger AI Overview: informational (how does X work, what is Y, why does Z happen), comparative (X vs Y, best X for Y), how-to (how to do X, steps for Y) and definitional (what does X mean). Types that trigger it less often: clear buying searches (buy X, X pricing, X near me), highly local searches, navigational searches (someone looking for a brand they already know), and sensitive topics (specific medical or legal advice, politics) where Google tends to hold the AI answer back.

The practical point: first find out which of your target questions produce an AI Overview and which don’t. A question with no AI Overview doesn’t reward AI Overview work, so put the effort into the ones that do. Checking this question by question, and repeating it regularly, is now routine measurement, not a one-time check. Our guide to AIO tracking covers how.

Stage 2: How does AI Overview pick its sources?

Once Google decides to show an overview, it picks a small number of pages to write the answer from. It draws from two pools: the top 10 to 20 normal results for the search (so pages that already rank well are candidates), and a separate retrieval step that can pull a page from outside the top results if it answers the question better.

Within that pool, several signals decide which pages get used. Ease of quoting: pages where the answer is cleanly laid out (direct answers up front, FAQ sections, schema markup) are favoured. Source authority: sites Google treats as a primary source on the topic, such as established publishers, the brand itself and recognised experts, carry more weight. Clear structure: clean headings, proper markup, factual writing and named sources. Freshness: where recency matters, recent dates and current references count for more.

The number of sources actually linked in the final overview is small, usually 3 to 6, sometimes more for complex questions and fewer for simple definitions. That is a much tighter bottleneck than page one, where you are one of ten results. Being a candidate is necessary but not enough. Being one of the few actually used is the goal. We go deeper on this in how Google AI Overview selects sources.

You can read the selection pattern by looking at which sites are linked across the AI Overviews for your target questions. Tools such as Profound, Otterly, AthenaHQ and BrightEdge AI automate this and show each competing site’s share. For a quick first look at your own brand, the free AIO Visibility Checker shows whether Google’s AI Overview names you today.

Stage 3: What does AI Overview pull from each page?

Once the sources are chosen, the model pulls the passages it will use to build the answer. This is the stage where the structure of your page makes the most direct difference.

From what we and others have observed, extraction prefers: direct answers (one or two sentences answering the question at the start of a section, before the detail), FAQ sections (question and answer pairs, often marked up with FAQPage schema), numbered steps for how-to questions, clear markup (proper heading order, lists, tables where they fit) and facts with a named source (data with its source, claims with references, quotes with a name). It does poorly with long narrative paragraphs where the answer is buried. The model can find it, but it tends to pick cleaner sources first.

This is the stage you can most easily act on. The same article becomes much easier to quote with structural changes that don’t change what it says: open each section with the direct answer, add an FAQ at the end of long pieces, mark up steps properly, and break dense paragraphs into shorter ones. These changes usually help normal SEO too, so the work is rarely wasted on a single channel. Our guide to schema markup for AEO covers the markup side.

Stage 4: How does AI Overview write the answer and choose the links?

In the final stage the Gemini model writes the answer from the passages it pulled, with links back to the source pages. It is really writing: it paraphrases and combines across sources rather than pasting passages together word for word. The links point to the sources that fed the answer.

What appears in the final overview: an answer in plain language, in paragraphs and sometimes lists; links to the source pages; an option to expand it; sometimes follow-up questions; and on some searches a structured card (price ranges, comparison tables, step lists) built from the source data. Being named inside that answer is what gives a brand visibility, because the reader sees the name even if they never click.

Clicks from AI Overview are well below normal result clicks. Studies through 2024 and 2025 reported organic click-through drops of 30 to 60% on searches where an overview appears, varying by question type and format. So being in the answer produces brand visibility more reliably than it produces clicks. The two are different results and should be measured separately.

What makes a page more likely to be used, and how do you measure it?

Pulling the four stages together, the pages AI Overview keeps using share a few qualities: the answer sits in the first one or two sentences of the relevant section; FAQ sections match question-style searches cleanly; schema markup gives clear signals (Article, FAQPage, Organization, Product); the site is a recognised source on the topic, with a consistent description of what the business is and mentions on other sites; the content is fresh where the topic needs it; and the writing is factual, without padding.

Measuring AI Overview performance is now routine work. The numbers that matter: the trigger rate on your target questions (which ones produce an overview, and the trend), how often you are named in the answer, how often you are only linked as a source, how you are described when named, and what happens next (clicks from the overview and the enquiries or sales they bring). The guide to measuring AI search visibility sets out the method we use.

The realistic view in November 2026: AI Overview is not the only AI search engine, but for most businesses it is the one that matters most because of Google’s share of search. The structural choices it rewards are largely the same ones that help with ChatGPT, Gemini, Claude and Google AI Mode.

Conclusion

How AI Overview works, in short: Google decides whether a search gets an overview, picks candidate pages from the top results and a separate retrieval step, pulls the passages that answer the question, and has a Gemini model write an answer with links to the sources it used. The four stages give your work a concrete target instead of a black box.

The choices AI Overview rewards (direct answers, FAQ sections, schema markup, recognised authority, clean factual writing) also help across the other AI engines and in normal SEO. Understanding the stages is the starting point. Measuring the trigger rate and how often you are named, question by question and over time, is where it becomes ongoing work rather than a one-off rewrite.

What to do next

If you have just learned how AI Overview works, the next question is usually whether it names you, and what to do if it doesn’t.

• See whether Google’s AI Overview names you today with the free AIO Visibility Checker.
• If a competitor is named and you are not, read why AI recommends your competitor and what to do and my competitors are in AI Overview but I’m not.
• If you rank well but enquiries have dropped, see we rank on page one but enquiries dropped.
• Deciding who should do the work? Compare hiring a GEO agency, training your team or doing it yourself.
• Vetting an agency? Learn how GEO results are measured and the questions to ask an agency.

Frequently asked questions