How AI Overviews are rewriting the first page of Google

For twenty years the first page of Google was a list. Today, on some share of queries, a generated answer sits above that list, summarising several sources and linking them.

What Google has published about this is unusually specific. What Google has deliberately left unsaid is equally specific. This article separates the two, then shows the checks we actually run when a client asks how to get in.

What Google has actually published

AI Overviews launched in the US in May 2024. In late October 2024 Google announced the rollout to more than 100 countries and put the figure at over one billion users a month. The current availability list includes Germany, Austria, Switzerland and Italy, with both German and Italian among the supported languages.

Three sentences from Google's own documentation matter more than any commentary around them.

First: there are no additional requirements to appear in AI Overviews or AI Mode, and no special optimisations are necessary. Second: to be eligible as a supporting link, a page must be indexed and eligible to appear in Google Search with a snippet. Third, and the one that gets quoted least often, AI Overviews are shown only when Google's systems judge them additive to classic Search, and therefore often do not trigger at all.

That third sentence should change the brief you write, because the question is not whether AI Overviews exist but which of your queries produce one.

What changes for your visibility

The most reliable public measurement comes from the Pew Research Center, published 22 July 2025. It covers 68,879 Google searches made by 900 US adults during March 2025:

• 18 per cent of the searches produced an AI summary.
• On visits to results pages with an AI summary, users clicked a traditional search result in 8 per cent of visits. Without a summary, 15 per cent.
• Users clicked a link inside the summary itself in 1 per cent of visits.
• 26 per cent of sessions ended on a page with an AI summary, against 16 per cent without.
• One and two-word searches produced a summary 8 per cent of the time. Searches of ten words or more, 53 per cent. Searches starting with a question word, 60 per cent.

That last group is the practical one. If you want to know which of your rankings are exposed, sort your queries by word count and isolate the ones beginning with how, what, why or which. Short brand queries and transactional queries largely stay out of it.

Google writes in its own documentation that clicks from results pages with AI Overviews are "higher quality", because users are more likely to spend longer on the site. It has published no figures behind that. We treat the sentence as a vendor statement rather than a finding.

Every Pew figure above comes from a US sample and should not be generalised to German- or Italian-language search results without local data.

How Google builds the answer

Google describes two mechanisms, and understanding them removes about half the folklore in circulation.

The first is retrieval augmented generation, also called grounding. Google's core ranking systems retrieve relevant, current pages from the Search index; the model then reads those pages and generates a response from them, with clickable links to the supporting pages. There is no separate AI index. It is the same index.

The second is query fan-out: the model issues a set of concurrent related searches to gather more material. Google's own example takes "how to fix a lawn that's full of weeds" and generates side queries such as "best herbicides for lawns", "remove weeds without chemicals" and "how to prevent weeds in lawn".

That is where the published record stops. Google has not disclosed how a specific passage inside a retrieved page gets selected as a cited source. There is no documented threshold and no confirmed weighting for individual formatting choices. Anyone telling you which structure "the models prefer" is describing an observation, not a rule. We say so out loud in client meetings, because the alternative is selling confidence we do not have.

What Google tells you not to bother with

Google's guide to optimising for generative AI features lists the practices that do not help. The list is worth reading in full, because most of it is currently being sold as a service:

• llms.txt and other special files. No machine-readable files, AI text files or dedicated markup are needed.
• Chunking content. No requirement to break text into tiny pieces. Google's systems handle multiple topics on one page, and there is no ideal page length.
• Rewriting content for AI systems. The systems understand synonyms and general meaning, so chasing every long-tail variant is unnecessary.
• Chasing inauthentic mentions. Google writes that seeking mentions across the web is less useful than it appears, because the spam systems apply to generative features too.
• Over-focusing on structured data. Not required for AI features, and no special schema.org markup exists for them. It stays worthwhile for rich results, which is a different goal.

There is a sharper warning attached to fan-out. Creating a separate page for every possible query variation, in order to influence rankings or generative responses, violates Google's scaled content abuse spam policy. A proposal for "one landing page per derived question" is a risk, not a strategy.

Getting cited: our order of checks

When a client asks how to get into AI Overviews, we work through seven points in this order. The order matters: point one is what makes everything after it effective.

1. Verify snippet eligibility instead of assuming it. Without snippet eligibility the page is not admissible as a supporting link. We regularly find a nosnippet directive or a max-snippet value set years ago by a plugin or a legal review and never revisited. Check it with the URL Inspection tool in Search Console, against the HTML Googlebot actually received, not the source view in your browser.

2. One question per page. We write the question verbatim as the H1 or the first H2 and answer it in the two sentences directly below. Not in the third paragraph, not after the introduction.

3. The answer has to stand alone. Practical test: copy the paragraph, paste it into an empty document, read it aloud. If it makes no sense without the paragraph before it, rewrite it. The same passage feeds featured snippets, so the work pays twice.

4. Related questions in the same document. Fan-out tempts everyone into building a thin page per sub-question. That is exactly what the spam policy rules out. We build the sub-questions as H3 sections inside the same URL.

5. Named author, visible dates. Full name in the byline, a real publication date, a real update date. We do it because readers need it and because Google maintains its own documentation on byline dates. We do not claim it produces citations.

6. Settle the entity. Identical company name across every page, Organization markup, an up-to-date Business Profile. Google's guide explicitly names Business Profile and Merchant Center as relevant to visibility in generative responses. The wider topic sits in our article on how to get mentioned in AI answers.

7. Images or video where they carry the explanation. Also from the official guide, and the only point that draws regular pushback. One real screenshot of a Search Console view beats three stock photos.

For bilingual businesses in South Tyrol we add an eighth step that nobody else needs: run the whole check separately on the German and Italian versions, with different question sets. Translating the questions is not enough, because the question a client in Bolzano asks in German is not the question they ask in Italian. On where the three optimisation surfaces diverge, see the difference between SEO, AEO and GEO.

AI Mode and what the numbers mean

AI Mode is the extended surface. It splits the question into subtopics, searches them in parallel and supports follow-ups. It is available in Germany, Austria, Switzerland and Italy, with German and Italian among the supported languages.

Two details surprise most clients. A follow-up question in AI Mode counts as a new query, so one session can generate several rows in Search Console. And the strongest model available in AI Mode, Gemini 3 Pro, is limited to English, with a country list that includes Switzerland and San Marino but not Germany, Austria or Italy.

On measurement, the starting point is that AI Overviews and AI Mode get no dedicated report. They roll into overall Search Console traffic inside the "Web" search type. Three quirks make the numbers treacherous:

An AI Overview occupies exactly one position. Every link inside it gets the same position number. An average position of 1.0 can therefore mean you were the seventh link inside an overview.

Impressions require actual visibility. For an AI Overview, an impression counts only once the link has been scrolled or expanded into view. If your source sits behind a "Show more" control, your page can be used without generating an impression.

Search Labs experiments are excluded entirely. Google does not feed that data into Search Console.

In practice: watch the gap between impressions and clicks separately for question-shaped queries and for long queries above roughly ten words, because that is where the Pew data shows summaries appearing most often. A falling click-through rate against stable impressions on exactly those segments is the clearest signal Search Console can give you.

What it will never tell you is whether an AI Overview appeared for a given query at all, and whether your business was named in it without anyone clicking. That needs a logged prompt set, repeated on a fixed schedule. The method is in our article on how to measure AI visibility with a prompt set.