Edition 15
AI is beginning to separate the person who asks from the people who produced the answer.
For much of the last twenty-five years, the web ran on an imperfect bargain.
Someone created something useful. A journalist reported a story, a reviewer tested a product, an engineer explained how something worked, a lawyer wrote an analysis, a traveller described a place. A hobbyist solved a problem and posted the solution. A company documented its software. Thousands of people answered questions on forums simply because they knew the answer and somebody else did not.
Search engines indexed all of it. When somebody wanted the information, search helped them find the source, and the publisher received a visit. That visit might produce advertising revenue, a subscription, a sale, a newsletter signup, a consulting enquiry, a donation, some reputation, or simply another person joining the community.
Nobody thought the arrangement was particularly fair. Publishers complained about search engines for years. Search companies complained about people trying to manipulate rankings. Users complained about advertising, search engine optimisation (SEO) pages and sites that buried a two-line answer under fifteen paragraphs of filler. It was messy, but there was an exchange. The person looking for the answer usually had to visit somebody who had produced part of it.
AI is beginning to separate those two things.
A person can now ask a question and receive an answer assembled from several sources without opening any of them. For the user, this can be excellent. For the web that produced the information, the economics are less obvious.
The Prompt
Suppose you want to know why a particular error appears in a database. A few years ago, you might search for the error message, open the official documentation, read a Stack Overflow discussion, find a developer’s blog post and perhaps discover that somebody had encountered exactly the same problem three years earlier.
Today you can paste the error into an AI assistant. If the answer is good, you may be finished in thirty seconds. You do not need to know which engineer wrote the explanation, which forum discussion exposed the unusual edge case or which documentation page supplied the exact parameter that fixed the problem.
That is much of why AI assistants are useful: they remove the labour of finding, opening and reconciling sources. The same pattern works for recipes, travel, product research, health information, legal questions, financial concepts, academic subjects, home repairs and thousands of other everyday searches. We should expect people to prefer the easier route, and there is already evidence that they do.
Pew Research Center examined Google browsing behaviour in 2025 and found that users clicked a conventional search result on 8 percent of visits where an AI summary appeared. When there was no AI summary, they clicked a result on 15 percent of visits. Links cited inside the AI summary itself received clicks in only about 1 percent of those visits.
Those numbers came from one study of one period and should not be stretched into a prediction for the entire web. They describe a fairly intuitive behaviour: if the answer on the page is sufficient, there is less reason to leave the page. The question is what happens upstream when that behaviour becomes normal.
The Mirage
It is easy to think of information on the internet as something that simply exists. There is so much of it that scarcity seems absurd. Search almost any common question and millions of pages appear. AI systems can retrieve from an enormous pool of material, and if one website disappears, ten others may contain something similar.
That abundance hides how the useful part of the web is made. A news report begins with somebody making calls, reading documents, attending an event, checking facts and deciding what deserves publication. A serious product review requires somebody to obtain the product, use it, measure it and compare it with alternatives. Technical documentation has to be written by people who understand the system. A useful forum answer may contain knowledge accumulated over twenty years of doing the job. A scientific paper represents research that might have taken months or years. A local restaurant review requires somebody to have actually eaten there, and a travel article eventually traces back to somebody who went to the place.
AI can make all of this information easier to consume. It does not remove the cost of producing the first observation. Knowing something and summarizing what other people know are different activities, and generative AI is extraordinarily good at the second. The web still needs the first.
The final answer looks self-contained. The user sees a neat explanation and has little reason to think about how many acts of human work sit underneath it.
The source becomes invisible at exactly the moment its information becomes most convenient.
The Reality Check
Publishers are already seeing parts of this change. Traffic from traditional search has been declining across many publishing businesses. AI is not responsible for all of it: search algorithms have changed, social referral traffic has fallen, audience habits are shifting and publishers themselves have changed strategy. AI-generated answers are now another factor.
Recent publisher-industry research points in the same general direction as the earlier Pew findings. When an answer is given directly, fewer users continue to the underlying websites. Traffic sent from standalone AI assistants is growing, but it remains small compared with the search traffic publishers historically received.
For a large newspaper, that becomes an obvious revenue problem. For the rest of the web, the effect may be less visible. Consider the independent expert who writes detailed technical articles because they bring consulting clients, the enthusiast who maintains a specialist website supported by advertising, the software company that invests in documentation because good documentation brings developers into its ecosystem, the reviewer whose site survives on affiliate revenue, and the community whose best contributors keep answering questions because reputation inside that community has value.
The arrangements differ, but each depends on some return from making useful information public. That return does not always have to be money. Recognition, audience, reputation, community and the chance that somebody will discover your work all count. If AI systems routinely consume the value while the source receives less of the relationship, some of those incentives will weaken.
Perhaps new incentives will replace them. Publishers may license content directly to AI companies. Experts may build paid communities. More information may move behind subscriptions. Businesses may publish material specifically so machines can understand their products. Creators may rely more heavily on newsletters, direct audiences and membership rather than search traffic. The web has adapted before, but adaptation can change what kind of information remains freely available.
The Leadership Question
This problem reaches beyond publishers. Any company building with AI should care about the health of the information environment it depends upon. A good answer requires good source material, and if the available sources become thinner, more commercial, more derivative or less carefully maintained, the quality of the systems built on top of them eventually changes as well.
There is an awkward possibility here. AI can reduce the incentive to produce original information while simultaneously increasing demand for original information. The better the answer, the less likely the user may be to visit the sources. The fewer people who visit the sources, the harder some sources become to fund. If fewer original sources are produced, future systems have less primary material to retrieve from.
This will not happen uniformly. Government data will still be published, companies will still document products, universities will still produce research, and people will still write because they enjoy writing. Communities will continue to form around shared interests. Commercial publishing is more exposed because its economics depend directly on audience.
There is also a difference between information created for people and information created to be consumed by machines. Once companies realize that AI assistants influence what users learn, buy and consider, they will adapt their content accordingly. We saw this with search. SEO began as a sensible effort to make websites understandable to search engines and eventually became an enormous industry devoted to influencing what search engines ranked.
The same incentive now exists around AI systems. Companies will want their information retrieved and brands will want to be recommended. Consultants will offer ways to improve AI visibility, and publishers will structure material so systems can parse it easily. Some of this will improve the quality of information. Some of it will become another form of gaming.
The web after answers may therefore contain an odd mixture: less economic support for some kinds of original work, and much more material designed specifically to influence the machines doing the answering. An abundance of information does not guarantee an abundance of knowledge.
The Closing Signal
There is no good reason to make information deliberately difficult to find simply to preserve an older internet business model. If an AI assistant can answer a straightforward question in thirty seconds, forcing somebody to visit six websites would not make the world better. Users will choose convenience when convenience works. The challenge is finding an arrangement where convenience does not slowly undermine the sources that make the convenient answer possible.
Some of that will be solved commercially. AI companies are already entering licensing agreements with publishers and other content owners. New forms of attribution, referral and compensation will develop. Publishers will find ways of building direct relationships with readers, and businesses whose information has commercial value will keep producing it because being represented accurately by AI systems matters to them.
Some parts will be harder. Nobody pays the retired engineer who writes the definitive explanation of an obscure failure on a forum, or commissions the hobbyist who spends a weekend testing five pieces of equipment and posts the results. Nobody knows in advance which obscure blog post will contain the observation that becomes invaluable three years later.
Much of the web’s usefulness came from people contributing information before anyone knew how useful it would become.
That is difficult to reproduce through licensing agreements.
It also leads to the next problem. If producing new human material becomes less rewarding while producing synthetic material becomes almost free, the composition of the web itself begins to change. Future AI systems will not be searching the same web that earlier systems learned from. There will be much more material written by machines, rewritten by machines, summarized by machines and published by people who may never have checked the original source.
AI has become very good at giving us the answer without requiring us to visit the people who helped create it. That is a real convenience. Whether it becomes a sustainable information economy depends on what happens to those people after the click disappears.
The Closing Signal
If nobody needs to visit the source to get the answer, what will keep the next source worth visiting?