What are Google's AI Overviews?
Google's generated answer box at the top of results, built from several linked searches and shown with cited sources.
AI Overviews are the AI-generated answer box Google shows at the top of its results, above the organic links, with a few sources cited next to the text. Google announced them at its I/O event on May 14, 2024, in the United States first, as the heir of the experiment then called Search Generative Experience; they reached Italy on March 26, 2025, and by May 2025 Google's official blog declared them live in over two hundred countries and more than forty languages. They are not a summary of the first page of results: Search Central documentation describes a query fan-out technique, where the system issues several related searches across different subtopics and builds the answer on those retrievals. The same page attributes that technique to AI Mode as well, the conversational surface Google keeps distinct and aims at comparison questions, with models and links of its own. It is the same index as always, queried differently and handed to a generative model that writes the text.
See also: Google Business Profile · The free profile deciding how a business appears on Search and Maps: hours, reviews and location, before the website.
How sources are chosen
The mechanism is retrieval plus generation: Google retrieves pages from its own index and uses them as the material the model generates on, a case of grounding against a proprietary corpus. On source selection the official position is blunt and should be read for what it says: in the "AI features and your website" page of Search Central, Google states that to appear as a supporting link a page only needs to be indexed and eligible to show in Search with a snippet, and that there are no additional requirements nor special optimizations. It adds that the models identify supporting pages while the response is being generated, so as to show a wider and more varied set of links than a classic search. Two practical consequences follow: the cited source does not necessarily match the first organic result, and crawler access remains the prerequisite, though no longer an on-off switch. With content signals, the initiative Cloudflare has been driving, a site can state separate preferences for separate uses of the same text: be indexed for search while denying the use of its pages as material for the generated answer. These are statements of intent, not technical blocks, and whoever receives them may ignore them; what changes is that the choice is no longer "in or out of Google", but "in the index, out of the box". The topic is covered in AI crawler access control.
An enterprise example
An industrial components manufacturer sees organic sessions drop on its technical documentation pages while rankings stay flat. Log analysis shows the affected queries are informational ones of the "difference between protocol A and protocol B" kind, exactly those that now trigger the generated box: the searcher reads the answer there and goes further only when the detail matters. Transactional pages, price lists and configurators, do not move. The correct reading is not that the site lost value, but that part of the informational demand has moved inside the results page and needs instruments other than visit counts to be measured.
Search Console, however, is no help in isolating it. Google's own documentation says that sites appearing in AI features, AI Overviews and AI Mode included, are part of the overall traffic in the Performance report under the "Web" search type, with no separate entry: impressions and clicks from the box are already mixed in with those from classic results and no filter separates them. Anyone who genuinely wants to quantify the shift goes back to server logs, where crawler hits and referrer sessions can still be told apart one by one. Logs are the reason the analysis in the example above was possible at all.
Why it matters for decision makers
The traffic effect is where the most partisan numbers circulate, so it is worth sticking to a single independent measurement with a published methodology: the Pew Research Center tracked the browsing of more than 900 US adults in March 2025, across 68,879 distinct Google searches, and recorded a click through to a site in 8% of the visits with the box against 15% of the visits without it. That holds for that sample and that month, it is not a universal law, but the direction is enough for a budget decision: anyone funding content on a forecast of sessions alone is measuring a channel that is narrowing.
It should be said that Google disputes this reading, and the two claims are better held side by side than resolved in favor of whichever is convenient. On August 6, 2025, Liz Reid, who heads Search, wrote on the company's official blog that total organic click volume from Google Search to websites has been relatively stable year over year, and that average click quality has actually increased. That is not necessarily a contradiction, because the two measurements are not of the same quantity: Pew observes the probability of a click when the box is present, across people tracked one by one, while Google reports an aggregate total across all of Search, where new queries can offset the ones that have stopped producing visits. What matters to a decision maker is that neither answers the question they actually have, which is what is happening to their own pages: that one is answered only on their own data. The question to bring into the meeting is not how to get cited, which is the business of GEO, but which pages still hold the traffic promise they were approved on.
Related terms
- Google Business Profile · The free profile deciding how a business appears on Search and Maps: hours, reviews and location, before the website.
- Share of model · Share of model is the share of mentions and citations a brand earns, compared with its competitors, in the answers of generative engines such as ChatGPT or Gemini.
- Grounding · Anchoring a model's answers to verifiable sources instead of its memory: RAG is the main technique used to achieve it.
- AI crawler access control · Deciding which AI crawlers may read your site, distinguishing training crawls from real-time answer retrieval.
- Content freshness · How much a content's date matters for AI citation or Google ranking: it weighs differently in each case.
A term that hits close to home? Write to me.
CONTACT ME