The large language models (LLMs) war intensifies because AI tools aren't just getting smarter, but also getting themselves in more places.
Since OpenAI launched ChatGPT in late 2022, the moment ignited a fierce competition among tech giants and startups to dominate generative AI. It disrupted the status quo, exposing how unprepared many established players were for consumer-facing large language models. Companies scrambled to respond, and Google, long-regarded as the undisputed giant, was fearing for its life.
But along the way, Google has steadily caught up and, in many ways, surged ahead through relentless iteration on its Gemini family of models.
What began as a defensive move has evolved into a comprehensive push, leveraging Google's unmatched scale in data, compute (via custom TPUs), and distribution through Android, Search, and other products.
And now, the company said that it has upgraded AI Overviews, which is its AI-generated summaries that appear at the top of many search results, to run on the powerful Gemini 3.
With it, AI Overviews can now make more accurate, nuanced responses globally, handling complex queries with superior reasoning. This upgrade makes AI Overviews even more reliable and helpful, addressing earlier issues like occasional factual errors.
(1/2) Today we’re starting to test a new way to seamlessly go deeper in AI Mode directly from the Search results page on mobile, globally.
This brings us closer to our vision for Search: just ask whatever’s on your mind – no matter how long or complex – and find exactly what you… pic.twitter.com/mcCS7oT2FI— Robby Stein (@rmstein) December 1, 2025
Building on that foundation, Google has integrated AI Mode directly into AI Overviews for mobile users.
Now, when an AI Overview appears, users can tap "See More" to enter a conversational interface, asking follow-up questions seamlessly without starting a new search. This creates a fluid, chat-like exploration of topics, keeping people engaged within Google's ecosystem rather than bouncing to external sites. The shift supports both quick answers and deeper dives, aligning with how people increasingly seek information.
This approach underscores what sets Google apart in the ongoing AI race.
While competitors like OpenAI excel in standalone conversational experiences, Google combines frontier intelligence with real-time web knowledge, vast personal context from its apps, and seamless access across devices. The result is an AI that's not just smart but contextually aware and embedded where people already live their digital lives.
As the LLM landscape matures, Google's integrated ecosystem positions Search to evolve from a link provider into a truly personal, proactive intelligence companion.
However, like any major shift in how billions access knowledge, it comes with notable catches, limitations, and potential downsides that users, publishers, and privacy-conscious individuals should weigh carefully.
One of the most immediate and widely discussed drawbacks is the dramatic impact on web publishers and the traditional search ecosystem. AI Mode often delivers comprehensive, synthesized answers without displaying the classic list of organic links, or it buries them far below the fold. Early data shows AI Overviews already reduce click-through rates to websites by around 30-35%, and full AI Mode experiences amplify this effect significantly.
Sites that aren't directly cited in the AI's response can become effectively invisible, threatening traffic-dependent businesses, creators, and the open web. This shift prioritizes Google's ecosystem retention over driving users outward, raising concerns about reduced discoverability for smaller publishers and a more closed information landscape.
(2/2) This means you’ll continue to get an AI Overview as a helpful starting point, and now you can also ask conversational follow-up questions in AI Mode – right from where you are.
It’s one seamless experience: a quick snapshot when you need it, and deeper conversation when…— Robby Stein (@rmstein) December 1, 2025
Accuracy and reliability remain persistent challenges, even with Gemini 3's upgrades.
While the model excels in benchmarks for reasoning and knowledge when it has solid grounding, hallucinations continue to plague large language models, including Gemini variants.
Some evaluations of related Gemini 3 models (like Flash) show extremely high hallucination rates in uncertain scenarios, where the AI prefers to invent details rather than admit gaps in knowledge.
AI Overviews and conversational AI Mode can occasionally misinterpret complex queries, provide outdated facts, or blend sources inaccurately, especially on nuanced, fast-moving, or niche topics. Google has improved grounding with real-time web access, but errors still occur, and users must verify critical information rather than trusting summaries blindly.
Privacy stands out as a major point of contention, particularly with Personal Intelligence. By opting in to connect Gmail and Google Photos, users grant the AI access to deeply personal data, emails containing financial details, health info, travel plans, private conversations, and photo libraries full of family moments.
Although Google stresses that this is strictly opt-in, revocable at any time, doesn't involve mass training on personal data, and includes safeguards like on-device processing where possible, critics argue the convenience tempts over-sharing.
Misinterpretations could lead to awkward or off-target recommendations (e.g., assuming outdated interests from old photos), and the broader integration fuels long-standing worries about Google's data practices. Even with controls, granting such access normalizes deeper surveillance-like personalization, and any future policy shift could erode trust.
Usability hurdles also emerge in practice.
AI Mode's chat-like interface, while fluid for follow-ups, can feel less discoverable or navigable for casual users compared to traditional results. Some report friction in finding sources, refining queries, or escaping the AI bubble to explore original pages. The experimental nature of features like Personal Intelligence means occasional glitches, context misreads, or responses that "go off track," requiring manual corrections via thumbs-down feedback or follow-ups.
In essence, while Google's integrated approach: combining frontier models, real-time search, and personal context offers genuine advantages over standalone chatbots, the trade-offs include publisher traffic erosion, lingering factual risks, heightened privacy exposure for those who opt in, and a search experience that feels more curated (and controlled) than exploratory.
These elements highlight why many users remain cautious, sticking to classic results or cross-checking AI outputs, even as the technology advances rapidly.
Before this, Google introduced Personal Intelligence within AI Mode.
Available as a Labs experiment for Google AI Pro and Ultra subscribers in the U.S, users who opt in can connect their Gmail and Google Photos, and unlock tailored suggestions grounded in their own life.
For instance, planning a trip might pull in hotel bookings from email alongside family photos to recommend kid-friendly activities, or shopping advice could factor in past purchases and local weather.
Gemini 3 handles these without directly training on personal data en masse, maintaining strong privacy controls while delivering uniquely relevant insights, something standalone chatbots struggle to match without similar integration.