What Happens to Writing When Nobody Has to Visit the Website?

A reader asks a question about something that matters only enough to solve, not enough to study. The answer appears in a pale panel before the links. It is clear, plausible and brief. There are source names at the edge of it, but nothing in the answer requires the reader to meet them. The question is closed. The page remains unopened.

Somewhere behind that answer is a piece of writing: a reported article, a guide, a study, perhaps a paragraph that took a person longer to verify than the reader took to consume. Its information has travelled. Its setting has not.

This is an old tendency in a new form. Search engines have long tried to shorten the distance between a question and its resolution. Featured snippets, knowledge panels and instant answers already allowed many searches to end without a click. Generative search changes the scale and texture of that extraction. Instead of lifting one passage, it can compose a new response from several pages, then continue the exchange without returning the reader to any of them.

For writers and publishers, the obvious concern is traffic. Yet traffic is only the visible edge of the change. The deeper question is what writing becomes when usefulness no longer carries the reader back to the place where that usefulness was made.

The answer has begun to leave the page

Traditional web search was a map with commercial interests. It arranged destinations, described them and asked the reader to choose one. The destination still carried most of the burden of explanation.

AI search increasingly behaves like a host. It gathers material, speaks in its own voice and keeps the conversation within its room. Google says AI Overviews and AI Mode may issue several related searches across different subtopics before composing a response. OpenAI describes ChatGPT Search as a way to receive a direct, current answer, with source links available alongside it. Both companies frame these links as opportunities for discovery. [2][3]

The links matter, but they are no longer the organising principle of the experience. The answer is.

That distinction changes the smallest unit of online discovery. It was once usually the page: a reader saw a title, chose a source and entered an environment created by someone else. Now the unit may be a sentence assembled across pages, delivered in the visual language of the intermediary. The reader can ask a follow-up question without learning which source carried the central fact, which supplied an example and which was merely adjacent.

A source can therefore become visible without being encountered. Its name may appear; its work may shape the answer; the reader may still never form an impression of who made it.

A visit was never only a visit

In analytics, a visit is a session. In human terms, it is an opening.

The reader sees the evidence surrounding a claim. They notice whether the author is careful or careless, whether the page distinguishes reporting from opinion, whether the argument develops or simply accumulates points. They may follow a related idea, recognise a name, remember a visual, subscribe, enquire or return without asking a search engine for permission.

For an advertising-funded publisher, the visit creates inventory. For a subscription business, it creates the possibility of conversion. For an independent writer, it may create something less easily counted: recognition. The reader begins to connect an idea to a person rather than to the internet in general.

That is why the loss of a click cannot be understood only as a smaller number in an acquisition report. An extracted answer may preserve the immediate informational value of a page while interrupting the formation of a relationship with its author.

This does not make every visit profound. Most are not. A person checking a definition or a train time is unlikely to develop affection for the page that supplied it. But the possibility of a deeper encounter existed because the reader arrived. When the answer is detached from the destination, that possibility narrows before the writer has any chance to earn it.

The evidence is real, but still untidy

The clearest behavioural evidence so far points towards fewer clicks when an AI answer is present.

Pew Research Center examined 68,879 Google searches made by 900 US adults in March 2025. When an AI summary appeared, users clicked a conventional result in 8 per cent of visits, compared with 15 per cent when no summary appeared. A link inside the summary itself was clicked in only 1 per cent of visits. The study was substantial, but it was still a snapshot of one country, one month and one search product moving quickly beneath the researchers’ feet. [1] (pewresearch.org)

A 2026 working paper used the staggered rollout of Google AI Overviews and Wikipedia’s multilingual structure as a natural experiment. Across more than 161,000 matched article-language pairs, the authors estimated that exposure to AI Overviews reduced daily traffic to English Wikipedia articles by roughly 15 per cent. The effect was larger for cultural topics and smaller for science, technology, engineering and mathematics, which suggests that substitution depends on whether a short synthesis can satisfy the query. [5] (arxiv.org)

Another working paper, published in July 2026, compared US desktop information-seeking activity in ChatGPT and Google using Comscore clickstream data. It found outbound clicks in 5.2 per cent of ChatGPT conversation sessions and estimated that wider access to ChatGPT Search reduced conventional search use by 9.4 per cent. The remaining referrals were more likely to lead to specialised destinations and less likely to lead to advertising-supported sites. These are early results, not a final measurement of a settled market, but they describe a different distribution of attention rather than a simple transfer from Google to another referrer. [4] (arxiv.org)

The scale still matters. Similarweb data reported by Axios counted six million referrals from nine major AI chatbots to the top 500 US media and news sites in February 2025. Traditional search engines sent roughly 2.26 billion referrals to that group in the same month. AI referral traffic was growing quickly, yet it remained small beside the channel it was beginning to disturb. [10] (axios.com)

Measurement is made harder by the platforms themselves. Google currently folds traffic from AI Overviews and AI Mode into the general “Web” search type in Search Console rather than exposing a separate report. It also says that people who click from search results containing AI Overviews tend to spend longer on the destination, although its public guidance does not provide the underlying methodology needed to evaluate that claim fully. [2] (developers.google.com)

There are other reasons to be cautious. Query types differ. Geography matters. Interfaces change. Some studies rely on third-party estimates; several of the newest are preprints. Even impressive growth numbers can conceal a rising platform tide: one 2026 field study found that ChatGPT referrals to a site grew 5.7 times, while untreated pages on the same domain grew 3.5 times over the period. Its authors described the intervention effect as suggestive rather than conclusive. [11] (arxiv.org)

The direction of travel is becoming easier to see than its final distance. Answer interfaces can reduce the need to visit informational pages. What they will do to every kind of site, and what new behaviour they may create in return, remains less certain.

When information becomes abundant, context becomes scarce

It helps to separate two kinds of value.

Answer value is what survives extraction. It is the fact, instruction, definition or comparison a reader can use after it has been lifted from the original page.

Destination value is what emerges from entering the source: the sequence of the reasoning, the evidence beneath it, the author’s voice, the visual explanation, the related work, the correction history and the sense that somebody is answerable for what has been said.

Some pages contain almost entirely answer value. A unit conversion does not become richer because the reader meets its author. A recipe temperature can travel safely on its own, provided it travels accurately.

Other writing depends on its surroundings. An essay is not simply a warehouse of sentences waiting to be retrieved. Its meaning accumulates through order, emphasis and return. A reported investigation gains force from the method and evidence around its conclusion. Original research cannot be responsibly reduced to a chart without its sample, definitions and limitations. Even a short opinion changes when the reader can see the wider body of thought from which it came.

A sentence has neighbours. Remove it and the dictionary meaning may remain, while its pressure changes.

An answer can be extracted from a page; the reasons to trust it usually cannot.

This matters because citation can create the appearance of provenance without guaranteeing it. A 2026 audit of Google AI Overviews broke responses into 98,020 individual claims and found that 11 per cent were unsupported by the pages cited. A separate large-scale experiment found that citations increased people’s trust in generative search even when the references were incorrect or hallucinated. [6][7] (arxiv.org)

The source page is therefore more than a courtesy link. It is one of the few places where a reader can inspect what the intermediary has compressed, combined or misunderstood. A small citation marker may point towards that inspection, but it cannot perform it on the reader’s behalf.

Perhaps the visit was never the point

There is a strong objection to all of this: readers do not owe publishers a click.

They arrive with a question, not a duty to support a business model. If an answer interface resolves that question accurately in twenty seconds, it may have served the reader better than a page that hides the same information beneath adverts, pop-ups, autoplaying video and six paragraphs written to delay satisfaction.

The web helped create the appetite for extraction. Too many pages treated the reader’s attention as something to be captured before it could escape. Search-led publishing often stretched small answers into large surfaces because page views rewarded the surface. Some zero-click searches are not evidence of cultural decline. They are evidence that the reader did not need a destination.

Nor is every AI-mediated referral worthless. OpenAI and Google both argue that their search products can expose people to sources they might not otherwise find. The July 2026 clickstream study found that ChatGPT’s relatively scarce outbound clicks leaned towards specialised sites. In retail, where a website is still needed to complete the transaction, Adobe Analytics reported that AI-referred US shoppers generated 53 per cent more revenue per visit and converted at a higher rate than visitors from non-AI sources in May 2026. That result cannot simply be transferred to publishing, but it suggests that fewer visits need not mean weaker intent. [2][3][4][8] (developers.google.com)

A smaller number of more deliberate readers may be valuable. A person who leaves an answer engine because the summary is insufficient has already expressed a deeper need. The visit begins later in the journey, after the basic question has been partly resolved.

The counterargument changes the shape of the problem. The goal should not be to force every answer back behind a click. It should be to ask whether the system can keep producing original, accountable work when intermediaries retain much of the attention generated by it.

Usefulness to the reader and sustainability for the source are separate questions. A good interface may answer the first while quietly worsening the second.

The website has to become more than a container

If a page can no longer compete by being the only place to obtain a fact, it has to become the best place to understand why the fact matters.

That begins with work that has a source because it came from somewhere: firsthand experience, original reporting, experiments, data, close observation, a defensible argument. Synthetic abundance raises the value of evidence that can be traced to a person, a method and a moment.

The page should make that provenance visible. A clear byline, a useful biography, primary sources, publication and revision dates, methodological notes and honest corrections are not decorative signals of credibility. They let the reader inspect the work rather than merely receive its conclusion.

Destination value also comes from connection. A strong independent site is not a pile of isolated answers. It is a body of thought in which one piece alters the meaning of another. The archive shows continuity, contradiction and development. The reader can see what the writer keeps returning to, and perhaps why.

For an organisation, the equivalent is a site that reveals its people, evidence and way of thinking. Product claims can connect to research. Advice can connect to practice. A case study can show the decisions hidden behind an outcome. The site becomes a record of responsibility rather than a distribution endpoint.

Design belongs here too, although not as cosmetic rescue. Typography can create patience. Diagrams can preserve relationships that a summary flattens. Navigation can turn a solitary article into an invitation to explore. A quiet, fast and readable page can remind the visitor that attention is being hosted rather than harvested.

Publishers are also experimenting with control at the infrastructure and commercial levels. Cloudflare introduced tools to block AI crawlers and test a pay-per-crawl model in 2025, while some publishers have negotiated licences with AI companies. Le Monde’s agreement with Perplexity paired compensation and source links with an answer engine built into Le Monde’s own site and apps, using its archive as the bounded source. None of these models has resolved the wider bargain, but each recognises that content access, attribution, visitation and payment can no longer be treated as the same thing. [9][12] (reuters.com)

The website, then, becomes less valuable as a container and more valuable as a place: part archive, part evidence room, part expression of authorship. Its advantage is not that an AI system cannot read it. Its advantage is that a reader can encounter more there than the AI system has chosen to carry away.

Writing after the click

The change will exert pressure on the writing itself.

One response is to make every passage easier to extract: descriptive headings, compact definitions, self-contained claims and sentences shaped to survive quotation. Much of this is good practice. Clarity helps readers. Stable structure helps search engines and assistive technology. Explicit sourcing makes verification easier.

But a page can become so eager to be extracted that it stops rewarding completion. Every section begins to behave like a separate answer. The prose loses rhythm because nothing is allowed to depend on what came before. Nuance is repeated until it becomes a slogan. The essay starts to resemble the summary that may eventually replace it.

Writers now have to work for two forms of legibility. A machine should be able to identify the subject, argument, author and evidence. A human reader should still feel that the whole piece offers more than the sum of those extractable parts.

That does not require a secret discipline of AI optimisation. Google’s own guidance says the same technical and editorial foundations used for conventional search remain relevant to its AI features, and that no special schema or machine-readable “AI file” is required. Pages still need to be crawlable, internally connected, fast, accessible and honest about what their structured data describes. [2] (developers.google.com)

The harder work is editorial. State important ideas clearly without repeating them mechanically. Use headings to orient, not to pre-digest every conclusion. Make claims quotable, then give the reader reasons to inspect how they were reached. Write sentences that can travel, but build essays that reward staying.

Search optimisation should make the path visible. It should not turn the destination into a corridor.

What remains worth visiting

Return to the reader at the beginning. The answer arrived. It was enough. They never opened the page.

That is not necessarily a failure of writing. A fact can be useful at a distance. An idea can reach someone who would never have found its author. The open web has always depended, in part, on information travelling beyond the conditions of its publication.

The risk appears when travelling becomes the only mode of encounter: when words circulate but sources become interchangeable, when attribution is visible but unvisited, when influence grows less measurable while the cost of producing trustworthy work remains stubbornly real.

A website visit may no longer be the automatic reward for publishing something useful. Perhaps it never should have been. It now has to be earned through what the answer box cannot fully contain: evidence, sequence, voice, visual thought, memory and the possibility of a relationship with the person behind the words.

The answer may travel on its own. The writer’s task is to make the source worth finding.

---

Sources and further reading

1. Athena Chapekis and Anna Lieb. “Google users are less likely to click on links when an AI summary appears in the results.” Pew Research Center, 22 July 2025.

2. Google Search Central. “AI Features and Your Website.” Google for Developers, last updated 10 December 2025.

3. OpenAI. “Introducing ChatGPT search.” OpenAI, 31 October 2024; updated 5 February 2025.

4. Qiaoni Shi, Kai Zhu and Kai Gu. “Answering Without Referring: How AI Search Rewrites the Web’s Economic Bargain.” arXiv, 8 July 2026.

5. Mehrzad Khosravi and Hema Yoganarasimhan. “Impact of AI Search Summaries on Website Traffic: Evidence from Google AI Overviews and Wikipedia.” arXiv, 5 February 2026; revised 12 May 2026.

6. Haofei Xu, Umar Iqbal and Jacob M. Montgomery. “Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact.” arXiv, 13 May 2026.

7. Haiwen Li and Sinan Aral. “Human Trust in AI Search: A Large-Scale Experiment.” arXiv, 8 April 2025.

8. Arriana McLymore. “AI-referred US shoppers browse longer, spend more per visit, data shows.” Reuters, 15 June 2026.

9. Krystal Hu. “Cloudflare launches tool to help website owners monetize AI bot crawler access.” Reuters, 1 July 2025.

10. Sara Fischer. “Axios Media Trends: Search slows.” Axios, 1 April 2025.

11. Keisuke Watanabe and Kazuki Nakayashiki. “Disentangling Answer Engine Optimization from Platform Growth: A Log-Based Natural Experiment on ChatGPT Referral Traffic.” arXiv, 3 June 2026.

12. Jérôme Fenoglio and Louis Dreyfus. “Artificial intelligence: Le Monde signs partnership agreement with Perplexity.” Le Monde, 14 May 2025.