Google launches AI Overviews in Search

Background

For most of the web's life, a search engine answered a question with a list of links, and the reader did the rest: pick a page, skim it, decide whether to trust it. Generative AI suggested a different arrangement — let a model read the results and write the answer. Google's first public step in that direction was the Search Generative Experience, unveiled at its I/O developer conference in May 2023^[2]. AI Overviews grew out of that experiment^[2].

The launch

On 14 May 2024, at Google I/O, Google said it had begun rolling out AI Overviews to everyone in the United States, with more countries to follow^[1]. The feature was powered by a new Gemini model customised for Google Search, and the announcement was made in a post by Elizabeth Reid, VP of Search^[1]. Google also set out an expectation for scale: it said it expected to bring AI Overviews to over a billion people by the end of 2024^[1].

What a user sees is a short generated passage placed above the ordinary list of results. TechCrunch described AI Overviews as a snapshot of information at the top of the results page^[3]. The design assumes that many queries can be answered by synthesising material that is already published, and that a reader would often rather have the summary first and the sources second.

When it went wrong

The rollout was quickly followed by trouble. Erroneous answers appeared, including advice to apply glue to pizza, and the problems led Google to temporarily restrict the tool in May 2024^[2].

The failure mode is worth understanding, because it follows from how summarising systems work. A model compresses sentences taken from web pages into one fluent paragraph. If the underlying page contains a joke, sarcasm, an outdated claim, or a statement quoted in order to be refuted, the compression can carry it forward as if it were fact. On the results page the paragraph arrives as continuous prose, stripped of the cues — the site's reputation, the surrounding context, the date of publication — that a reader would normally use to judge a link. A list of links lets a reader disagree with a source; a summary asks the reader to disagree with the search engine.

Expansion

On 28 October 2024, Google expanded AI Overviews to more than 100 countries and territories^[3]. The expansion added languages including English, Hindi, Indonesian, Japanese, Portuguese and Spanish^[3].

Why it matters

The significance of AI Overviews is less about any single answer than about what now sits in the first layer of the results page. A conventional results page handed the work of selection and judgement to the reader. An overview performs part of that work in advance: it shortens the path from question to statement, and moves some of the responsibility for being right onto the summarising model. The pizza episode shows what that shift costs when it fails — the more a generated answer reads like a settled conclusion, the more damage a wrong one does.

Seen from the launch to the expansion, AI Overviews also mark a change in what a search product is. The engine no longer only indexes and ranks pages; it writes the first piece of text a user reads. How that arrangement settles alongside links, and alongside the sites whose material is being summarised, remained an open question after the rollout.