Google AI Mode and AI Overviews explained
Technical scope Major labs & flagship models, google, vendor-analysis
Google now answers many searches before you click anything. At the top of a growing share of results pages, a Gemini-generated summary – an AI Overview – stitches together an answer from across the web, and for a subset of queries Google offers AI Mode, a full conversational search surface that behaves less like a list of links and more like a chatbot with a citation habit. According to Google's own announcements, AI Overviews rolled out broadly to US users in May 2024 and have since expanded to more countries and languages; AI Mode followed as a dedicated experience, graduating from the experimental Search Generative Experience (Google SGE) that Google had tested through Search Labs. Both are built on the Gemini model family, and both change the economics of the open web in ways publishers are still measuring.
This piece explains what each feature is, how the user experience differs from traditional search, what the early traffic data shows for publishers, where the summaries still fail, and how the answer paradigm compares with Perplexity and ChatGPT Search. Treat every specific figure here as time-sensitive: Google iterates on these products constantly, and you should verify current behaviour against Google's official Search blog before making decisions on it.
AI Mode and AI Overviews in one paragraph
AI Overviews are Gemini-generated summaries that appear above the traditional blue links for queries where Google's systems judge a synthesised answer useful. They cite sources inline, link out to a handful of pages, and aim to resolve the query on the results page itself. AI Mode is the more ambitious cousin: a dedicated, conversational search experience – reachable from a tab in Google Search – where you ask a question, get a structured generative answer with links, then ask follow-ups that keep context. AI Mode grew directly out of Google SGE, the Search Generative Experience that Google ran as an opt-in Labs experiment before folding its lessons into the shipping products. Where an AI Overview is a single enriched answer block, AI Mode is a whole session of them.
The user experience
For a factual or how-to query, the difference is immediate. Traditional Google search returns a ranked list and asks you to do the synthesis: open two or three results, reconcile them, form a view. An AI Overview does that synthesis for you and presents it as prose, with expandable citations and a cluster of source links off to the side. For "how long to boil an egg for a soft yolk" or "difference between USB-C and Thunderbolt", the answer arrives without a click.
AI Mode pushes further. It handles multi-part and comparative questions that a single search box historically mangled – "plan a three-day vegetarian itinerary in Lisbon with a budget hotel near the tram" – by decomposing the request, running multiple underlying queries, and assembling a composite answer. Google describes this internally as a query fan-out: the system breaks a complex prompt into sub-questions, retrieves for each, and recombines. Follow-up questions retain context, so the interaction resembles a conversation rather than a series of fresh searches. If you have used ChatGPT's voice and conversational modes, the rhythm will feel familiar, though AI Mode's retrieval is anchored to Google's live index.
The trade-off is legibility. A ranked list makes its sourcing obvious – you see the domains, the titles, the snippets. A generative answer compresses that provenance into inline citations that most users never expand. The answer feels authoritative precisely because it has hidden its working.
The publisher impact
This is where the stakes are highest, and where the data matters most. The open web's bargain with Google was simple: Google indexed your content and sent you readers. AI Overviews strain that bargain, because an answer rendered on the results page can satisfy the user without any click at all – the zero-click search, now escalated from a feature of featured snippets to a default for a large slice of informational queries.
The evidence is accumulating rather than settled. Industry analyses and publisher-side measurement, including reporting in Press Gazette on news-publisher traffic, have documented meaningful declines in click-through from search for queries where an AI Overview appears, with informational and definitional content hit hardest. Several studies circulated through 2024 and 2025 reported that the presence of an AI Overview correlated with lower organic click-through rates on the queries it covered, though methodologies and sample sizes vary and the numbers move as Google tunes placement. The honest summary is this: when Google answers the question itself, the pages that would have earned the click lose it, and the pages Google cites gain far less traffic than a top-ranked link historically delivered.
Not all content is equally exposed. Transactional queries, pages that require interaction (booking, purchasing, account access), and content where the user wants the primary source rather than a summary are more resilient. Pure reference and explainer content – the kind that answers a discrete question completely – sits directly in the blast radius. For publishers whose business was built on exactly that category, the shift is existential rather than incremental, and it compounds the broader pressures covered in our analysis of AI's economic and labour-market impact.
Where AI Overviews fail
The feature's early public failures became a genre of their own. Shortly after the US rollout, AI Overviews surfaced answers advising users to add non-toxic glue to pizza sauce to stop cheese sliding off, and to eat a small rock per day for minerals – the former traced to a sarcastic Reddit comment, the latter to satirical content from The Onion. These were widely reported, and Google acknowledged them and rolled back or constrained the feature for certain queries.
The failures are instructive because they are structural, not incidental. A retrieval-augmented summariser is only as reliable as the sources it trusts and its ability to judge tone, authority and intent. When the web's answer to a nonsensical query is a joke, a literal-minded model repackages the joke as fact. When few authoritative sources exist, the model fills the gap with confident synthesis – the familiar hallucination problem, now presented with Google's imprimatur above the organic results. Google has added guardrails, reduced AI Overviews for queries where they underperform, and improved its handling of low-quality and satirical sources, but the underlying tension remains: a system optimised to always produce an answer will sometimes produce a wrong one, and the interface design makes that wrong answer look as trustworthy as a correct one.
SEO implications
The optimisation question has split the SEO industry. Coverage in Search Engine Land and elsewhere converges on a few durable observations, even as tactics churn.
First, inclusion in an AI Overview does not map cleanly onto classic ranking. Pages cited in Overviews are often, though not always, drawn from results already ranking on page one, but the model shows a preference for content that answers sub-questions directly, in clear, extractable passages. Content structured around specific questions – with unambiguous, self-contained answers near the relevant heading – is more liftable than content that buries its conclusion.
Second, the entity and authority signals Google has long rewarded matter more, not less. The model needs to trust a source enough to synthesise from it. Established topical authority, clear authorship, and corroboration across reputable sources raise the odds of being the page Google quotes rather than the page it ignores.
Third, the realistic goal has shifted from "capture the click" to "be the cited source and capture the fraction who click through for depth". That is a worse deal for most publishers, and no amount of optimisation fully restores the old traffic. The defensible strategy is to produce content a summary cannot replace: original reporting, proprietary data, genuine analysis, interactive tools, and depth that rewards the reader who does click. Thin explainer content that an AI Overview can fully absorb has the least future. Practitioners building audience-facing tools and workflows around this shift may find our coverage of AI agents and their developments useful context, since agentic retrieval is where search is heading next.
Versus Perplexity and ChatGPT Search
Google is not alone in answering rather than listing, but the paradigms differ in ways that matter.
Perplexity built its product from the premise that the answer is the interface. It foregrounds citations more aggressively than Google – every claim typically carries a visible numbered source – and its audience skews toward research-oriented users who want a sourced synthesis quickly. Its weakness is reach and index depth: it depends on third-party search infrastructure and web access, and it lacks Google's distribution.
ChatGPT Search, OpenAI's web-connected mode, grafts live retrieval onto a conversational model that millions already use for non-search tasks. Its strength is the continuity of the assistant – search is one capability inside a broader workflow – and its audience is people who already live in ChatGPT. Its weakness, historically, has been freshness and the breadth of a dedicated search index.
Google's advantage is structural and hard to overstate: it owns the index, the distribution, the default search box on billions of devices, and now a competitive model in Gemini. AI Mode and AI Overviews meet users where they already are, which is why their impact on the web dwarfs that of standalone competitors. The risk Google carries is the inverse of its advantage – it has an enormous existing business built on sending clicks outward, and every answer it keeps on the page cannibalises that business while antagonising the publishers whose content feeds it. Perplexity and OpenAI have no legacy click economy to protect; Google has nothing but.
What publishers should actually do
The strategic reality is uncomfortable but clear. Treat search as one channel among several rather than the foundation, and accelerate the move toward audiences you own – newsletters, apps, communities, memberships – where no intermediary sits between you and the reader. Invest in content that resists summarisation: original data, first-hand reporting, strong authorial voice, tools and experiences. Structure the content that remains explainer-shaped so that if Google is going to quote someone, it quotes you, and so that the minority who click arrive on a page worth their time.
None of this restores the traffic economics of 2015, and pretending otherwise helps no one. The useful posture is to assume the answer layer is permanent, to measure your own exposure query by query rather than trusting industry averages, and to verify current behaviour directly – Google's placement logic, citation patterns and the share of queries carrying an AI Overview all shift from month to month. For the model capabilities driving all of this, our explainer on the Gemini model family and our profile of Google DeepMind cover the research and engineering behind the search experience you now meet by default.