The Web Is Turning Into a Trust Setting

AI search makes answers faster. But no-AI search, publisher controls, and source visibility are becoming the new tools for deciding what to trust

Image generated by the author

There is now a search page for people who want AI turned off. Not a manifesto. Not a browser extension made by someone angry on GitHub. Just a page: noai.duckduckgo.com.

DuckDuckGo describes it as a version of DuckDuckGo Search where all AI features are turned off and AI-generated images are filtered out by default. On its broader AI help page, the company says its AI features are optional and can be turned off or tuned down.

The important part is not DuckDuckGo itself. The important part is the label. A search engine now has to tell you when the search is not using AI.

That small button carries more meaning than it first appears to. It is not proof that AI search is bad. Sometimes a synthesized answer is exactly what you want, and nobody enjoys opening eleven tabs to find out whether a laptop charger is compatible, whether a drug has a common side effect, or why a line of code keeps failing at midnight.

The appeal is obvious but the problem is quieter. When the search starts writing the answer for you, the old work of looking begins to move out of sight.

The old search page was messy in a useful way

The old results page was never innocent. It had ads at the top, strange SEO pages in the middle, forum threads that were half-useful and half-insane, government PDFs from 2018, old blog posts with one perfect paragraph, recipe sites padded with family history, and enough affiliate sludge to make every “best” list feel like a trap.

It was not a library. It was a street market with rankings. But the mess gave you something to inspect.

You could see the date. You could squint at the domain. You could open three tabs and notice that two of them were quoting the third. You could decide that the source with the clean design had done less work than the ugly page written by someone who had clearly been living with the problem for years.

The old search page did not make you wise. It made the contest visible. That visibility mattered because search was not only a way to find information. It was a small ritual of judgment. The reader still had to walk through the shelves, even if the shelves were badly arranged.

AI search changes the shape of that walk. At its strongest, it does not simply add a paragraph above the links. It enters the web on your behalf, comes back with a cleaned-up answer, and asks you to accept that the route was good enough. The answer arrives polished. The path back to the source gets harder to find.

Google is rebuilding the search box

Google has been unusually clear about where it wants search to go. At I/O 2026, the company announced what it called “a new era for AI Search”, describing a new AI-powered Search box as the biggest upgrade to the search entry point in more than 25 years, with agents available “just by asking a question.”

The company also said AI Mode had surpassed one billion monthly users, with queries more than doubling every quarter since launch.

Those are Google’s own product and usage claims, not independent measurements. But they show the direction of the machine.

The new search box is designed for questions that do not fit neatly into keywords. Google says it can take text, images, files, videos, and Chrome tabs; continue a conversation from an AI Overview into AI Mode; and use agents operating in the background to monitor the web for updates related to a question.

This is not just a better autocomplete. It changes who does the information labor. Classic search gave you a map, however flawed. AI search offers to drive.

The trade is seductive because the driver is fast. It summarizes, compares, rephrases, and continues forward when the old results page would have made you start over.

But once the machine drives, the route becomes part of the trust problem. Not because it is always wrong. Because even when it is useful, what it consulted may be harder to trace than you expect.

The source may not be where you were trained to look

Imagine asking a question and getting an AI Overview. The answer looks calm. It has links. Nothing about the card feels suspicious. It does not look like a hallucination. It does not scream. It does not ask you to believe without evidence.

So you do what a careful reader does. You scroll down to check the sources. That is where the old habit starts to slip.

Google’s own Search Central documentation says AI Overviews and AI Mode may use “query fan-out”, issuing multiple related searches across subtopics and data sources to develop a response. Google argues this can surface a wider and more diverse set of helpful links than classic search.

That may be true. It also means the link layer in an AI answer is not simply the old first page wearing a summary on top. Two independent preprints make the shift concrete enough to matter, with the caveat that preprints are not settled consensus.

One study measured more than 55,000 trending queries across 19 topical categories over a 40-day window. It found that Google AI Overviews appeared for 13.7% of queries overall and for 64.7% of question-form queries.

More revealingly, it found that nearly 30% of AI Overview-cited domains did not appear in the co-displayed first-page results. The same study decomposed AI Overview responses into atomic claims and found that 11% were unsupported by the cited pages. Read the preprint.

A second preprint compared traditional Google Search, AI Overviews, and Gemini across 11,500 user queries. It found that AI Overviews appeared for 51.5% of representative real-user queries in its dataset, and that retrieved source sets were substantially different across the three systems. Read the preprint.

The lesson is not that every AI answer is unreliable. The lesson is that AI search creates a different evidence surface.

You may still get links, but those links no longer mean exactly what they meant on the old results page. They may come from a different route, a different search process, a different selection layer. The answer can be cleaner than the web it came from, while the path back to that web grows less legible.

A summary can be useful and still requires a kind of trust that the old results page did not.

Publishers get their own version of the button

The reader is not the only one looking for a control. Publishers are getting one too.

Google has introduced Search Generative AI performance reports in Search Console, designed to show site owners how their pages appear inside generative AI features such as AI Overviews and AI Mode. The reports include visibility signals such as impressions, pages, countries, devices, and dates.

That sounds technical until you notice what the new dashboard admits: Visibility inside AI search has become different enough from ordinary search visibility that it needs its own measurement system.

A publisher used to ask a simpler question: did search send people to my page? Now the question has multiplied.

A page might appear in an answer without generating a visit. It might be part of the answer layer while the domain itself never becomes the reader’s destination. Measuring any of this separately from ordinary search requires its own dashboard.

Then comes the harder question. Can I say no?

On June 3, 2026, the UK Competition and Markets Authority announced a conduct requirement for Google Search. The CMA said publishers would be able to opt out of having their content used to power AI features in Google Search, including AI Overviews.

It also said Google must ensure publisher content is properly attributed with clear links in AI-generated search results and must allow publishers to opt out of having their content used for model fine-tuning.

Read that alongside the no-AI button, and the shape of the new web comes into focus.

A reader wants to choose how much synthesis stands between them and the source. A publisher wants to choose how much of its work becomes material for a machine-written answer. A regulator wants the new answer layer to preserve some bargaining power, some attribution, and some trust.

All three are reaching for the same kind of control, just from different sides of the transaction.

The old web bargain was already uneven. Search engines crawled. Publishers optimized. Readers clicked. Platforms ranked. Nobody loved the arrangement, but the roles were legible enough to argue about.

AI search makes the argument stranger.

The page is no longer only a destination. It can become an ingredient, a citation, training material, summary fuel, and invisible structural support. Present in the answer without a visit ever arriving.

The dashboard exists because that presence finally needed a number attached to it. The opt-out exists because the number alone is not enough.

The wrong question is whether AI search is good or bad

The easiest version of this essay would be the least useful one: AI search is bad. No-AI search is pure. The old web was human. The new web is synthetic.

That version would travel well. It would also be lazy.

The old web had spam, plagiarism, content farms, SEO parasites, paywalled expertise, affiliate traps, and human-written garbage at industrial scale. A blue link was not a moral guarantee. A ranked page was not a truth machine.

But classic search did preserve a visible contest over sources. AI search changes where that contest happens. The better question is what kind of inspection the interface still gives you.

Can you see the sources? Can you compare them? Can you understand why they appeared? Can a publisher measure its presence and decline certain uses? Can the reader choose speed without surrendering every trace of the route?

These questions are not anti-AI, they are pro-judgment.

A fast answer is useful. A fast answer that shows no route asks for a different kind of faith. That distinction — between accepting an answer and being able to inspect one — is what the no-AI button makes visible.

Not a refusal.

One of the first honest labels in the new search era: This experience includes AI. This one does not.

That should not feel radical. It should feel basic.

The web is becoming configurable

The search box used to look simple because its complexity was hidden behind the ranking.

Now the complexity is moving toward the user. AI Mode, AI Overviews, no-AI search, publisher opt-outs, generative AI visibility reports, attribution requirements, fine-tuning controls, answer cards, source links that may not correspond to the old first page.

The web is not just being searched anymore. It is being summarized, routed, grounded, measured, negotiated, and turned into product settings.

Some people will want maximum synthesis and will happily let the machine compare ten pages and return with one answer. Others will want the messy source trail. Some publishers will want inclusion in AI search because invisibility is worse; others will want control because being used without a visit feels like slow erasure.

Most of us will want different versions of this bargain depending on the moment.

If I am checking tomorrow’s weather, synthesize away.

If I am researching a medical decision, a legal question, a political claim, or a technical issue that might break production, I want the route. I want the ugly pages. I want the date, the domain, the primary source, the dissenting result, the thing that does not fit neatly into the answer card.

I want to see where the machine walked. That may become the real divide in search. Not people who use AI versus people who refuse it. People who accept the answer and people who still ask for the trail.

Return to the button

The no-AI button is small. That is part of what makes it work.

It does not ask for a theory of the internet. It does not require a policy paper. It sits there, plain and almost awkward, doing the work of an honest label.

This experience uses AI. This one does not. The old search bar asked what you were looking for. The new one asks something quieter first. How much of the web do you still want to see?