For twenty years, finding something online meant gathering sources and forming your own view. AI assistants now hand you a single, ready-made answer, and a harder question comes with it: who decides what we get to know, and what never makes it into the reply? Access to information is moving through a few private systems, and the internet community has a short window to keep it open and accountable.

For most of the web’s history, learning something online was a plural process. You typed a query, a search engine returned pages of links, and you did the rest. You opened several sources, compared them, noticed where they disagreed, and built your own understanding. The search engine was a gateway, not a judge.

That is changing fast. More and more, you ask a question and an AI assistant hands you the answer, having already read the sources, weighed them, and decided what matters before you see anything. For the user, this feels like a gift: you can ask longer questions, follow up, and skip the tedium of ten blue links. The search box is becoming less an index of the web and more a personal assistant.

The convenience is real. So is the shift underneath it. We are moving from a web of discovery to a web of mediation, and the step that disappears is the one where we consulted the sources ourselves. When a model composes the answer, someone, or something, has already decided which sources counted, which were set aside, and which were never surfaced at all. If the answer arrives pre-assembled, who decides what we get to know, and what stays out of view?

The funnel is narrowing to a few pipes

One caveat first. AI still drives only a small share of traffic: AI chatbots send less than 1% of the page-view referrals that reach publishers, even after ChatGPT referrals grew more than 200% in a year. This is a shift in progress, not a finished one. What matters is the direction and the concentration, and both are already clear.

By mid-2025, a single assistant, ChatGPT, accounted for roughly 89% of measurable AI referral traffic in one closely watched dataset. A year later, the field had widened only a little. ChatGPT, Claude, Gemini, and Perplexity together made up close to 99% of it. Discovery is not fragmenting across a healthy diversity of tools. It is consolidating onto a few large models.

Users are also clicking through less. By late 2025, close to seven in ten Google searches ended without a click to any external site, and that share rose when an AI summary appeared, according to industry data. Pew Research has found that people are less likely to follow a link when an AI answer sits at the top of the page. Publishers feel it unevenly: over two years, search referrals fell by about 60% for the smallest publishers against roughly 22% for the largest. Gartner has forecast that traditional search volume will fall 25% by 2026 as users move to AI chatbots. Whether or not that exact figure lands, the direction is no longer in doubt.

The sources behind the answer are concentrated too

If a few models decide how you find things, a still smaller set of sources decides what those models draw on. A 2026 index that aggregated more than 680 million citations from ChatGPT, Google AI Overviews, Perplexity, Gemini, and Claude found that just 15 domains account for 68% of all AI citations, a concentration more extreme than Google’s old PageRank ever produced. Reddit alone was cited around 40% of the time across the major engines, and Wikipedia supplied between a quarter and nearly half of ChatGPT’s top citations. The firm behind the study concluded that what these systems cite is neither open, neutral, nor stable.

So the mediation runs in both directions. A handful of models sit between you and the web, and a handful of domains sit behind those models. Much of what the public reads is shaped by private, frequently changing decisions about which few places on the internet are worth quoting.

A knowledge problem, not only a traffic problem

Read only as economics, this is a story about clicks and ad revenue. Read as a knowledge question, it matters more.

When search mediated discovery, the range of sources you might meet was wide, and the choice of which to trust was yours. When an assistant mediates discovery, the range narrows to what one model chooses to show, and part of the choice of what to trust has been made for you. Researchers have begun to measure the effect. A 2025 study testing 27 language models across 155 topics and 12 countries found that almost all of them produced less epistemically diverse answers than a basic web search. Ask the web and you meet a broader spread of claims than if you ask the model.

The economist Andrew Peterson gave this dynamic a name in an influential paper, later published in the journal AI & Society: knowledge collapse. Because language models generate toward the center of their training distribution, relying on them can pull attention toward dominant, popular ideas and push peripheral or minority knowledge out of view. His modeling suggests the risk is real: when getting an AI answer is easier than seeking out varied sources, public understanding can drift further from the truth. Other scholars call this a loss of epistemic friction. Without the push-back of views that do not fit, a prevailing idea becomes harder to question.

This is where the technical story becomes a social one. If a small number of systems frame what billions of people learn about health, history, politics, or products, then the decisions those systems make about what to include and leave out become decisions about public knowledge itself. Most of the time no one sets out to suppress a view. The narrowing happens by design, as a by-product of optimizing for one fluent answer. That makes it harder to notice, and harder to contest.

When information access has few gatekeepers

The deeper concern is where this authority sits. A few AI layers are becoming the gatekeepers of information access, and they rest on an ecosystem that was already concentrated. As the internet policy scholar Konstantinos Komaitis has argued, AI tends to reverse the internet’s original logic. Where the network distributed agency horizontally among millions of endpoints, AI’s reliance on vast proprietary datasets and compute recentralizes it vertically, into the few firms that build and run the leading models. Those firms are often the same ones that already supply the cloud, the search box, the browser, and the app store.

This is not a new fight. For two decades, the internet governance community has worked to stop other parts of the network, from cloud hosting to content delivery, from collapsing into a few dominant providers. The Internet Society has tracked consolidation as a structural risk since its 2019 Global Internet Report, and the danger became visible in late 2025, when outages at a few infrastructure providers rippled across the global web. Concentration is fragile, whether what is concentrated is content delivery or the framing of an answer, and it is easy to steer when so few hands are on the controls.

What should the internet community do?

There is no single fix, and the choice is less technical than institutional. A few priorities follow from the argument above.

Keep information access anchored in open, accountable systems rather than proprietary ones that users can neither inspect nor challenge. The way we find things online is changing, and so is the role of the domain in an AI-driven web; the parts that can stay based on open standards and shared governance should. The domain name system is a useful model. It has stayed open and globally accountable for decades because it is run through a bottom-up, multistakeholder governance process that keeps any single company or government from owning the internet’s core functions. That model is not a relic. It is the most tested tool we have for keeping decisions that affect everyone answerable to more than a few firms.

Push for transparency about how answers are built. If a source is used or ignored, the people affected, a small publisher, a researcher, a business, should be able to see why and challenge it. Today, that is almost never possible.

Defend epistemic diversity as a public value. The measurable narrowing of what AI systems surface, set against open search, is reason to keep plural sources reachable and to resist a web in which one synthesized answer is the only one most people ever see.

Keep the ordinary user in the conversation. The people with the largest stake in how AI mediates knowledge, everyone who will be told by an assistant what is true and worth knowing, are the least represented in the rooms where the defaults are being set. Digital trust and discoverability in an AI-mediated web are being defined now, and they should be defined with users in mind, not only the companies that will operate the systems.

AI-first discovery is more convenient. That was never in doubt. The question worth ending on is different. Will it give us a better internet, or a more centralized one? The answer depends on choices being made right now, and on whether the people those choices affect have any say in them.