Google's new AI Mode is a preview of the future of search

This is a column about AI. My boyfriend works at Anthropic. See my full ethics disclosure here.

Just under a year ago, at its annual developer conference, Google signaled that a dramatic change was coming to its search results. In the near future, the company said, you would "let Google do the Googling for you": trusting the search engine to search the web on your behalf, summarizing its findings, and sparing you the need to visit many websites yourself.

It was an appealing proposition to Google, which would soon begin peppering the product it calls AI overviews with advertisements; and for the search engine's users, many of whom have already begun to replace their traditional Google searches with queries in ChatGPT, Perplexity, and other AI products.

To the millions of businesses that rely on Google to send them traffic, though, the bargain appeared much worse. AI overviews offered answers good enough that many users never bothered to click a link; some businesses have suffered accordingly. On Tuesday I wrote about the story of Chegg, the online education company, which has collapsed over the past two years as first ChatGPT and then Google took services the company offers students for $14.95 a month and began to offer them for free.

For the past year, we have arguably been in the first phase of Google search's AI evolution, with summaries generated by large language models sitting atop traditional search results. The question since then has been what will happen to the web once we enter phase two — the moment that AI-generated results become the primary way that Google answers most queries, with links to the sites from which Google derived that information relegated to the sidelines.

This week, phase two arrived — as part of an experiment available to subscribers to the company's Google One AI Premium service. On Wednesday, the company introduced "AI Mode," a preview of what Google search might come to look like over time.

"This new search mode expands what AI Overviews can do with more advanced reasoning, thinking and multimodal capabilities so you can get help with even your toughest questions," said Robby Stein, vice president of product for search, in a blog post. "You can ask anything on your mind and get a helpful AI-powered response with the ability to go further with follow-up questions and helpful web links."

Built on a customized version of the company's Gemini 2.0 model, Stein said AI Mode is well suited to answer questions that might previously have taken Google multiple searches. Unlike traditional Google search, you can ask follow-up questions of your queries; and unlike some chatbots, it's plugged into Google's search index and can offer real-time and local business information.

In an interview, Stein told me that AI Mode queries a broader set of websites than traditional search, and searches multiple related topics simultaneously to deliver more fleshed-out responses. Google has also tried to introduce some humility into AI Mode; it will acknowledge when it is uncertain about something, Stein said, or decline to offer an AI-generated summary when it has low confidence in the answer.

Happily — and to a greater extent than I had been bracing myself for — AI Mode also includes prominent links to the web. In a sample query looking to compare sleep tracking features of various smart devices, AI Mode offers a carousel of prominent links to websites with information about the subject directly beneath the first paragraph of the AI-generated answer; another column of links appears to the right of the result. Each paragraph also ends with a small link icon that directs you to the source of the information.

As presented in a demo, then, AI Mode stops short of doing all the Googling for you. The web is still there, and in some cases is presented quite prominently. I was glad to see it.

On Thursday afternoon, I got access to AI Mode, and spent some time hitting it with queries. In many ways, Google seems to have learned the hard lessons of its past few AI launches. Ask AI Mode whether you should put glue on your pizza and it will tell you no. ("Glue is not meant for consumption and can be harmful if ingested.") It will also dissuade you from eating rocks. ("Rocks are not digestible by the human body.")

One key difference between AI Mode as demo'd and as I experienced it is that the current version has fewer links to the web than the demo one. (Stein had warned me that this is the case; more carousels of links are coming soon, he said.) When I used AI Mode to search for popular spring break destinations, and houseplants that improve air quality, I got that right-hand column of links — but nothing in the body of the answer.

A GIF in the announcement post shows AI Mode creating a detailed table comparing smart rings, smart watches, and tracking mats; the version I got access today only generated text.

In general, I found AI Mode much more willing to answer my queries when they involved questions about products, travel, or other things I could buy. Other times, particularly when I asked about politics, AI Mode would decline to generate a response and instead only show links to websites that might be relevant.

And so I got no AI answer to “how has the Trump administration acted to restrict free expression so far in 2025?” or "why do conservatives want to shut down DEI programs?” (In fairness to AI Mode, Google's Gemini app refuses to answer that question at all; AI Mode at least offered some websites about the subject.)

Like most AI products in 2025, it's not totally clear when you should use AI Mode over Google's many similar alternatives. When I asked it to help me come up with a menu for a party I'm planning at my house, it offered some good if basic suggestions. Given the same query, Gemini — which has a memory feature and knows I'm a tech reporter — offered similar options but decided to give everything a tech-inspired name. ("Binary Code Caprese Skewers," "Data Stream Hummus," "AI Algorithm Spiced Nuts," and so on.)

In this case I prefer the far less cringeworthy suggestions of AI Mode. But even I can no longer remember when or how Google may be personalizing answers to my queries.

For my final test of AI mode, I tried some of the queries that have become popular on Claude: treating the AI more like a smart friend or confidante, and asking it for advice. It offered solid advice, though it was less creative than Gemini and other chatbots I have tried.

For the moment, AI Mode is just an experiment. But AI Overviews started as an experiment, too, and now they are default for more than 1 billion users of Google search. The good news is that Google is still working to highlight real websites, even as its AI gradually obviates the need to visit them. That bad news is that just because you show someone lots of links does not mean they will click on them.

On the podcast this week: Kevin and I talk through the implications of AI mode. Then, the Times' David Yaffe-Bellany joins us to explain why the strategic crypto reserve is neither strategic nor a reserve. And finally, listeners share their experiments in vibe-coding.

Apple | Spotify | Stitcher | Amazon | Google | YouTube

Governing

• Elon Musk’s bid to stop OpenAI transformation into a for-profit will have an expedited trial, a judge said. (Malathi Nayak / Bloomberg)
  • But his request to pause the transformation was denied. (Hayden Field / CNBC)
• The Trump administration said it would overhaul a $42 billion federal grant program for expanding high-speed internet and take “a tech-neutral approach,” including easing some rules that could benefit Starlink. (Cecilia Kang / New York Times)
• Trump’s USCIS may soon ask people applying for green cards, US citizenship, and asylum or refugee status to provide their social media accounts for review. (Gaby Del Valle / The Verge)
• Secretary of State Marco Rubio is reportedly launching an effort to use AI to revoke student visas of foreign nationals who appear to support Hamas or other designated terror groups. (Marc Caputo / Axios)
• A look inside the White House's new influencer-focused media strategy. (Drew Harwell and Sarah Ellison / Washington Post)
• A closer look at the proposed Take It Down Act, a bill that would criminalize nonconsensual intimate imagery, including AI-generated ones, but which could be weaponized by Trump, this author argues. (Adi Robertson / The Verge)
• Google has reportedly asked Trump’s Justice Department to reverse course on a push to break up the company over its illegal search monopoly, citing national security concerns. (Josh Sisco and Davey Alba / Bloomberg)
• Five Democratic senators asked the Justice Department to investigate whether Musk used his influence with Trump to pressure advertisers into returning to X. (Suzanne Vranica / Wall Street Journal)
• The House Judiciary Committee subpoenaed Alphabet and CEO Sundar Pichai in its investigation into big tech companies’ relationships with the Biden administration. (Emily Birnbaum / Bloomberg)
• ByteDance still hasn’t negotiated with prospective TikTok buyers, sources say, despite a looming sale deadline less than a month away. (Dan Primack / Axios)
• Former Google CEO Eric Schmidt, Scale AI CEO Alexandr Wang and Center for AI Safety Director Dan Hendrycks said the US should not pursue a Manhattan Project-style push into developing AGI, citing concerns of retaliation from China. (Maxwell Zeff / TechCrunch)
• A conversation with Ben Buchanan, former special adviser for AI in the Biden White House, on why he thinks AGI is coming and how the US government could prepare for it. (Ezra Klein / New York Times)
• Utah’s legislature passed the App Store Accountability Act, which requires app stores to verify users’ ages and require parental consent for minors to download apps. It’s the first state to do so; Meta is thrilled. (Lauren Feiner / The Verge)
• Microsoft’s $13 billion investment into OpenAI doesn’t qualify for a full investigation under merger rules, the UK’s competition authority said. (Upmanyu Trivedi / Bloomberg)
• Reddit introduced a Rules Check tool that tells users when their post could potentially violate a subreddit’s rules. This is great — more platforms should offer tools like this one. (Mariella Moon / Engadget)

Industry

• OpenAI is reportedly planning to charge up to $20,000 per month for specialized AI agents aimed at supporting PhD-level research. PhD-level research is typically ... not that expensive. (Stephanie Palazzolo and Cory Weinberg / The Information)
• Meta is rolling out two facial recognition tools in the UK, one to prevent scams using celebrity likeness, and another to help people get their compromised Facebook or Instagram accounts back. (Ingrid Lunden / TechCrunch)
• YouTube is rolling out its Premium Lite ad-free subscription tiers to all users in the US. (Jack Greenberg / YouTube Official Blog)
• YouTube has surpassed 125 million Music and Premium subscribers worldwide. (Murray Stassen / Music Business Worldwide)
• Apple made some Mac announcements:
  • The new MacBook Air now has an M4 processor, a new sky-blue color option, and a lower price. (Antonio G. Di Benedetto / The Verge)
  • New Mac Studios will have an M4 Max chip for the base model, and an M3 Ultra chip for the upgraded model. (Wes Davis / The Verge)
  • A look at the M3 Ultra chip, Apple’s most capable processor. (Andrew E. Freedman / Tom’s Hardware)
• The latest betas for iOS and iPadOS offer AI-generated summaries of reviews on the App Store. If they're anything like the text-message summaries, maybe don't rely on them too much. (Mahmoud Itani / Macworld)
• Amazon is testing an AI dubbing system for movies and television shows on Prime Video. (Ian Carlos Campbell / Engadget)
• Discord is reportedly in early talks with investment bankers to prepare for an IPO as soon as this year. (Lauren Hirsch and Mike Isaac / New York Times)
• Google co-founder Larry Page has reportedly formed a new startup, Dynatomics, that aims to use LLMs within product manufacturing. (Jessica E. Lessin / The Information)
• The Los Angeles Times began to offer AI-generated counterpoints underneath its opinion pieces, but a quick review of the product found a series of embarrassments. (Laura Hazard Owen / NiemanLab)
• The Association for Computing Machinery awarded Andrew Barto and Richard Sutton the Turing Award, for their work on developing reinforcement learning that has proved vital to AI chatbots. (Cade Metz / New York Times)
  • Barto and Sutton say they are concerned that AI companies are pushing out products without adequate safeguards and testing. (Cristina Criddle and Melissa Heikkilä / Financial Times)
• AI startup Sesame’s demo for its new conversational speech model sounds surprisingly realistic, testers say. (Benj Edwards / Ars Technica)
• A quarter of the current startup batch coming out of Y Combinator have 95 percent of their codebases or more generated by AI. (Ivan Mehta / TechCrunch)
• Digg is coming back, revived by co-founder Kevin Rose and Reddit co-founder Alexia Ohanian. (Mike Isaac / New York Times)
• Flipboard CEO Mike McCue is pitching his company’s new Surf browser as a decentralized alternative to social media algorithms. (John Markoff / New York Times)
• Trading volume for the Pump.fun memecoin tokens has plummeted from a peak of $3 billion in January to about $170 million now, a 94 percent decline. (Brandon Kae and Ivan Wu / The Block)

Those good posts

For more good posts every day, follow Casey’s Instagram stories.

(Link)

(Link)

(Link)

Talk to us

Send us tips, comments, questions, and posts: casey@platformer.news. Read our ethics policy here.