The Machine That Never Waits

The feature is called information agents. Ask a question in AI Mode, then tell Google to keep watching it. The agent runs in the background, continuously, scanning blogs, news, social posts, and Google's live data on finance, sport, and shopping. When something changes, it doesn't send a link. It sends a synthesised update: what happened, and what it means for the thing you asked about. Elizabeth "Liz" Reid, Google's VP of Search, introduced it on stage as the start of "the era of Search agents".

Google is calling it the next Google Alerts. That comparison undersells what changed.

What actually shipped

Alerts was a keyword match against an RSS-style crawl. You set the terms, it emailed you when they appeared somewhere. It was dumb by design, and that was the point: it fetched, it didn't interpret.

Information agents reason. You describe a goal in plain language, tracking a rental in a specific neighbourhood under a certain price, watching a developing news story, following a product launch, and the agent decides what counts as relevant, synthesises across sources, and can take action. It's rolling out in every country and language where AI Mode is already available, starting with Google AI Pro and Ultra subscribers in the US this summer. Free tier access hasn't been confirmed.

It arrived alongside two other changes that matter more once you put them next to each other. The search box itself was rebuilt, Google's own framing calls it the biggest change to the box in over 25 years, running on Gemini 3.5 Flash and accepting multimodal input. And Personal Intelligence, which pulls in Gmail, Photos, YouTube history, and search history to shape AI responses, expanded to nearly 200 countries with no subscription required.

Three announcements, one shift. Search stopped being a place you visit. It became a process that already has a profile of you, and now has the standing instruction to go and check on things without being asked again.

This isn't a Google story

Here's the detail that got missed in most of the coverage: every major AI product moved the same direction in the same eight weeks.

OpenAI rolled out Scheduled Tasks on June 17, retiring its older Pulse feature and giving Plus, Pro, Business, and Enterprise subscribers the ability to set recurring web monitoring jobs that surface only when the model decides a change is worth mentioning. Perplexity did the same inside Computer, its agent product, describing Scheduled Tasks as background jobs that run monitoring, competitive tracking, and daily briefings on a schedule you set, without your laptop needing to be open.

Three companies, three separate product teams, the same instinct: stop waiting to be asked, start deciding what's worth surfacing. That's not a coincidence of timing. It's a sign the industry has settled on where the next layer of the product sits, and it isn't in the query box.

The part Alerts never did

Nobody worried about Google Alerts. It had no opinion. It matched a string or it didn't.

An information agent has to decide what's worth telling you, which means deciding what's worth ignoring. That judgment used to sit entirely with the person doing the checking. You'd search again, skim ten results, notice the one that contradicted what you expected. Some of what you found was noise. Some of it was the thing that made you reconsider the question. You couldn't always tell which was which in advance, and that was, inconveniently, the entire value of doing it yourself.

Hand that filtering to an agent and the filtering doesn't disappear. It just stops being visible. You get told what changed. You don't get shown what the agent decided didn't. OpenAI is explicit about this mechanic in its own product documentation: monitoring jobs notify users only when the model judges a change to be significant, not on every check.

This is the same mechanism this site named in The Answer Engine and the End of Productive Disagreement, just moved from answering to monitoring. One voice, one synthesis, delivered on a schedule you didn't set and can't audit.

Who's actually running the agent

An agent that monitors a rental search needs your budget, your commute tolerance, your family size. One tracking travel needs your dates, your finances, your itinerary. The more useful the agent gets, the more of your situation it has to hold, continuously, to keep deciding what's relevant to you specifically. That is the profile this site went through in detail in The machine that knows you better than you do, now with a standing reason to stay current.

That data doesn't evaporate once the notification is sent. It's the same company that already holds your email and your photos, now with a standing, always-on reason to keep more of your life on file, indefinitely, in order to keep the monitoring useful. None of this requires bad intent to matter. An advertising business has never needed to act maliciously for its incentives to quietly shape what gets surfaced and what gets left in the background.

This isn't a hypothetical risk with no track record behind it. Google's handling of exactly this kind of data has already been tested in public and found wanting. An Associated Press investigation in 2018 found the company continued tracking users' location through web and app activity even after they'd switched off Location History. The same year, the Wall Street Journal reported Google had let third-party developers read users' Gmail messages. An agent that needs to know your budget, your commute, your travel dates to be useful is asking for more of exactly the kind of data that history says doesn't always stay where it's put.

There's also a colder, more structural cost sitting underneath the personal one. If an agent scans a site on your behalf so you never have to visit it, that site loses the visit. Multiply that across millions of agents doing millions of standing searches, and the traffic model the entire open web still runs on, the one this site has already tracked collapsing in Google Zero, takes another cut.

The scale of that redirection is already visible, though it is worth being precise about what the available data does and does not cover. Previsible's third AI Traffic Study, analysing 6.77 million LLM-driven sessions over 19 months, found ChatGPT commands 92.4% of trackable standalone AI referral traffic across the full dataset, up from roughly 84% share when the same team last published in December 2025. The study is Previsible's own and its author discloses he is the company's CPO and co-founder, so read it as a vendor dataset rather than an independent audit.

Standalone AI referral traffic, share by platform

December 2025, the last period with a full platform breakdown

ChatGPT

84%

Perplexity

8.9%

Gemini

4.5%

Copilot

2.1%

Claude

0.6%

The clicks that make it back to a publisher's site are concentrating hard around one product that isn't a search engine at all. Information agents add another layer of AI standing between a site and the person who might have visited it directly. That's not a separate story from the one here. It's the same mechanism, arriving from a different door.

What actually changes

Not that Google is watching you. It already was. What changes is the default posture toward staying informed, from something you do to something you subscribe to. Once checking is a thing that happens for you, the muscle for going and looking, on purpose, past the first synthesised answer, gets less use. Not because anyone forced that. Because it's no longer required.

Google Alerts asked you to set a keyword and wait for a match. Information agents ask you to set an intention and stop thinking about it. That's a small difference in the settings screen and a large one in what it trains a person to expect from knowing things: that it arrives, rather than that it's gone looking for.

The honest question isn't whether this is sinister. It's who ends up deciding what you needed to notice, once you've stopped being the one who checks.