In short: if more users get complete answers without leaving search, the web will not disappear, but its economics will change. The most exposed sites will be those whose main value can be compressed into a few AI-generated sentences. Sites with proprietary data, tools, services, communities and a direct audience should be more resilient. The analysis below looks at how this could reshape the web and which kinds of sites face the greatest pressure.
For years, the search-driven web operated on a fairly simple exchange. Websites made their pages available for indexing, Google helped people discover them, and a share of those users clicked through to the source. Those visits funded advertising, sales, subscriptions and the slow process of building a recognizable brand.
Generative search inserts another layer between the publisher and the reader. A search engine can read several pages, compare them and produce a usable answer on its own. That is often convenient for the user, but it creates a new problem for publishers: their information may be used even when the visit never happens.
The mechanism is already visible in user behavior. In a Pew Research Center study based on browsing data from 900 U.S. adults in March 2025, users clicked a traditional Google result in about 8% of searches that contained an AI summary, compared with 15% when no AI summary appeared. Links inside the AI summary itself received clicks in roughly 1% of those searches. This does not prove that every information site will lose half its audience, but it clearly illustrates the zero-click mechanism: once the answer is already visible, opening the source becomes less necessary.
Search is becoming an information layer of its own
Traditional Google search was built around the click. A user entered a query, scanned the results and opened one or more websites to assemble an answer. AI Overviews and AI Mode alter that sequence by trying to provide a synthesized response first and offering sources as a second step for deeper exploration.
The scale is now large enough that Google has begun treating generative visibility as a separate part of the search ecosystem. The company has reported audiences in the billions for its AI search features while also introducing new controls and measurement tools for website owners. In its explanation of these changes, Google explicitly frames the challenge as developing AI search without breaking the relationship with the open web.
Search Console now includes a separate Generative AI Performance report that can show how pages appear inside generative search experiences. This creates a new state for a webpage: it can be visible to Google and used in AI answers while receiving relatively few real visits. Visibility and traffic, which once tended to move together, increasingly need to be measured separately.
The most vulnerable website is not necessarily the smallest one
Size matters, but a more important question is how easily a page’s core value can be transferred into a short answer. A query such as “What is LiDAR?” can be answered reasonably well in a few sentences. The same is true for many basic definitions, simple device-lifespan questions and straightforward A-versus-B comparisons. If dozens of sites repeat essentially the same facts, an AI system can compress them into a summary without depriving the user of much practical value.
Pages built around original datasets, real-world testing, live databases, calculators, maps, forecast histories or active communities behave differently. An AI answer can describe them, but replacing them is harder because part of their value exists only when the user interacts with the resource itself.
| Page type | Risk of replacement by a short AI answer | What still gives users a reason to visit |
|---|---|---|
| Definition or basic reference page | Very high | Deeper database, examples, tool |
| Simple A vs. B comparison | High | Decision matrix for real-world conditions |
| How-to guide | Medium | Video, files, images, interactive steps |
| Analysis | Medium | Original causal reasoning and context |
| Forecast | Medium | Method, forecast history, resolution criteria |
| Original research | Low | Primary data and methodology |
| Live database or calculator | Low | Current data and interaction |
| Forum, store or service | Low | Action, community or transaction |
A large library of evergreen articles therefore does not guarantee resilience. A small specialist site with a useful database or tool may be better protected than a much larger publisher whose pages mostly repackage information available elsewhere.
A lost click also removes everything that could happen after it
A website visit matters for more than an analytics chart. Once a reader arrives, they may see an ad, continue to another article, join a mailing list, buy something or simply remember the brand. The economics of much of the open web is built on this sequence of opportunities rather than on the click alone.
Under the traditional model, a publisher invests in creating information, search brings an audience, and a small part of that audience funds future work. Generative search can shorten the chain: the information is indexed, AI produces an answer, and the user ends the search before reaching the source.
The size of the future impact is still uncertain, and it would be misleading to attribute every decline in Google traffic to AI. The Reuters Institute’s 2026 report cites Chartbeat data showing a substantial decline in organic Google referrals across a large sample of news sites. The researchers also note that AI Overviews cannot be cleanly separated from other algorithmic and behavioral changes. At the same time, publishing executives surveyed for the report expect search traffic to weaken further.
A simple stress test for search dependence
For an individual website, the starting level of dependence on Google may matter as much as the size of any future drop in search clicks. A simple sensitivity test can be calculated by multiplying the share of total traffic that comes from search by an assumed decline in search referrals. This is not a forecast; it only shows how exposed the business model is.
| Share of traffic from search | If search referrals fall 25% | Fall 50% | Fall 75% |
|---|---|---|---|
| 20% | −5% of total traffic | −10% | −15% |
| 40% | −10% | −20% | −30% |
| 60% | −15% | −30% | −45% |
| 80% | −20% | −40% | −60% |
If Google supplies 60% of a site’s visits and those referrals are cut in half, total traffic falls by roughly 30% if everything else stays unchanged. A site that gets only 20% of its audience from search would lose about 10% under the same shock. The same change in search can therefore be a channel problem for one publisher and a business-model problem for another.
A four-layer model for the future web
It is more useful to think about the emerging internet not as a simple fight between Google and websites, but as a division of functions across several layers. The first layer is the source layer, where information is actually created: public records, academic research, company filings, product documentation, journalism, testing and statistical databases. AI can summarize these sources, but somebody still has to produce the measurement, document or report in the first place.
The second layer is the answer layer, where systems such as generative search compare sources and assemble a response for a specific question. The third layer consists of destination sites where the user still needs to do something: stores, calculators, maps, editors, learning platforms, forums, booking systems and searchable databases. The fourth layer is the direct relationship with the audience — users who already know the resource, search for it by name, have an account, subscribe to a newsletter or participate in its community.
The weakest position in this structure may be the middle ground: a site that creates little original data, owns no tool or product, has no strong direct audience and depends mainly on explaining information that is already widely available elsewhere.
Data and direct audience relationships may become more valuable than article volume
Classic SEO rewarded scale for a long time. A thousand strong long-tail pages could cover a huge range of small queries and collectively produce dependable traffic. Generative AI reduces the scarcity of text itself: creating one more clear explanation is cheaper, and summarizing existing explanations is cheaper still.
What remains expensive is accumulated value — years of real price history, hundreds of proprietary tests, a structured decision database, a track record of forecasts and their outcomes, or an active community. A website can therefore build value not only through published articles but also through data, methods, tools, author reputation and direct relationships with readers. Articles still matter, but increasingly they may function as entry points into a deeper asset rather than as the entire product.
The problem eventually comes back to AI itself
Generative systems are good at transforming information that already exists, but most primary facts originate elsewhere. A company has to publish the financial filing, researchers have to run the experiment, a reporter has to gather facts, and a laboratory or reviewer has to test the device. If fewer visits significantly reduce the economic return on that work, some producers of original information will have to find other funding models.
This does not mean the internet is destined to run out of new information. Governments, universities, companies, software developers and users publish enormous amounts of material for reasons that do not depend on advertising traffic. But expensive forms of content — investigations, field reporting, testing and deep analysis — require resources. The long-term question is therefore not whether information will remain online, but how much of the highest-value information stays openly accessible and how economic value flows back to its creators.
How to recognize the problem on your own site
For a website owner, the most useful signal is not the mere appearance of AI in search but a widening gap between visibility and visits. If impressions stay stable or rise, the page starts appearing regularly in generative search, and clicks barely move, the next question should be what part of the user’s need Google is already satisfying before the visit.
In that situation, automatically rewriting the title or adding more keywords may solve the wrong problem. If the AI result already captures nearly all of the page’s value, the issue is deeper than CTR. One article may need a current data table, another an original calculation or local comparison, while a forecast page may benefit from a visible history of probability changes and the evidence behind them.
The point is not to add tables or charts mechanically. It is to identify the part of the subject that remains useful after the user already knows the short answer.
The internet is more likely to change than to shrink
There is little reason to expect users to stop visiting websites altogether. Stores still need catalogs and checkout, governments need official documents, companies need owned pages, services need interfaces, and creators need a place to work directly with their audience. The information economy can still change substantially without clicks disappearing completely; it is enough for a large share of simple searches to end inside the search experience.
That would weaken the model in which almost every useful answer could become a standalone SEO page with a predictable stream of readers. For information publishers, proprietary data, methods, tools, live databases, communities and audience trust may become more important because their value cannot be exhausted by a short summary in a search box.
