For most of the history of the web, search worked as a routing system. A person entered words into a search box, the engine ranked pages, and the user decided which result to open. The search engine’s main job was to point toward information; the user did much of the reading, comparison, and judgment. That division is changing.
Google’s AI Overviews and AI Mode, along with conversational search products such as ChatGPT Search, increasingly place an AI-generated layer between a question and the underlying web. Instead of receiving only a ranked list of pages, users can receive a synthesized answer assembled from multiple sources, ask follow-up questions without starting over, and increasingly use images, files, or voice as part of the search process. Google’s current AI Mode documentation describes a system that breaks a question into subtopics, searches for information across multiple sources, and combines those results into a response.
The significance is larger than a new search interface. AI search changes what it means to “find something online.” The user may no longer begin by choosing a webpage. They may begin by evaluating an answer generated from several webpages. That creates a new set of questions about traffic, source visibility, accuracy, attribution, and the economic incentives behind publishing information on the open web.
Search Is Moving From Retrieval Toward Synthesis
Traditional search is largely built around retrieval and ranking. A query is matched against an index, and the system presents documents or other results that it considers relevant. Even when ranking algorithms are extremely sophisticated, the fundamental interaction remains recognizable: query, results, and selection. AI search adds another stage. The system can retrieve information and then use a generative model to synthesize material from multiple sources into a response.
Google calls one part of this process “query fan-out.” AI Mode can break a complex question into related searches and run them across different subtopics and sources before producing an answer. Google’s documentation says the technique can be used by both AI Overviews and AI Mode, although the exact models and methods can differ. This matters because users no longer have to formulate every intermediate search themselves.
Someone researching a new laptop, for example, might previously have searched for processor specifications, battery tests, display measurements, reviews, and compatibility information separately. An AI search system can attempt to decompose the larger question and combine evidence from those different searches into one response. The convenience is obvious. The change in information discovery is less obvious: the search engine is increasingly performing part of the research process that used to happen after the click.
The Query Itself Is Becoming More Like a Conversation
The structure of the question is changing alongside the structure of the answer. Google reported in May 2026 that AI Mode queries in the United States were, on average, three times longer than traditional search queries. Google also reported that more than one in six AI Mode searches in the U.S. used voice or images, while planning-related queries were growing faster than AI Mode queries overall. These are Google’s own measurements and should therefore be understood as company-reported usage data rather than an independent measurement of the entire search market. That difference is important because conversational search allows people to describe a problem rather than translate it into keywords.
A conventional query might be
best laptop for video editing under $1,000
A conversational query can contain the user’s constraints, existing equipment, preferences, and follow-up questions. The system can retain the context and continue from the previous answer. Google’s current AI Mode documentation supports text, voice, images, and files and allows users to ask follow-up questions while retaining conversational context. This makes search more expressive, but it also gives the search system more responsibility. The longer and more contextual the request becomes, the more the system must decide which information is relevant and how different sources should be combined.
The Search Result Is Becoming an Intermediate Layer
One of the most consequential changes is that the webpage is no longer necessarily the first destination. Pew Research Center’s analysis of U.S. browsing behavior provides evidence of this shift. In its study of 68,879 Google searches during March 2025, 18% produced an AI-generated summary. When an AI summary appeared, users clicked a traditional search-result link in 8% of visits, compared with 15% when no AI summary appeared. Only 1% of visits to pages with an AI summary resulted in a click on a link inside the summary itself.
The study has limitations. It covered Google searches in the United States during a specific month, and the AI summaries and search results were captured after the browsing data was collected. It therefore does not establish a permanent click-through rate for AI search.
But it does demonstrate something important: when an AI-generated answer is present, a meaningful portion of users can obtain enough information from the search page without following a conventional result.
That changes the economics of publishing. For years, publishers could build an audience by earning visibility in search results and converting that visibility into visits. If more users consume the answer without visiting the source, the relationship between search visibility and website traffic becomes less direct.
The Publisher’s Problem Is No Longer Simply “Rank Higher”
Google says AI features create additional opportunities for websites to be discovered. Its Search Central documentation states that AI Overviews and AI Mode surface links to supporting websites and can expose users to a wider range of pages. Google also says there are no additional technical requirements or special AI-specific markup required to appear as a supporting link; standard search eligibility and foundational SEO practices still apply.
That is the platform’s position. Independent research has produced a more complicated picture.
A February 2026 study examining 161,382 matched Wikipedia article-language pairs estimated that exposure to Google AI Overviews reduced daily traffic to affected English-language Wikipedia articles by approximately 15%, with larger effects for some types of informational content. The authors used a difference-in-differences design and described the findings as evidence that generative answers can redirect attention away from source publishers.
Another preregistered field experiment published as an August 2026 research preprint, involving 1,100 participants, reported that removing AI Overviews and AI Mode increased publisher click-through rates. The study also reported lower click-through and weaker measures of user experience and trust in an AI mode-only condition. Because this is a research preprint rather than a settled industry standard, its findings should be treated as evidence from a particular experiment, not a universal measurement of all AI search behavior.
The disagreement is therefore not simply about whether AI search changes behavior. There is growing evidence that it does. The unresolved issue is how those changes vary by query type, source, search product, and user intent.
Information Discovery Is Becoming More Selective
Traditional search could expose a user to ten blue links, videos, news stories, forums, government pages, and other results. The user had to make choices. An AI answer can reduce that initial choice set. That can be useful when the question has a straightforward answer. It can also become problematic when the question depends on competing interpretations, incomplete evidence, or rapidly changing information.
Pew’s 2025 analysis found that Wikipedia, YouTube, and Reddit were among the most frequently cited sources in both traditional results and AI summaries. It also found that government websites appeared more frequently among sources linked in AI summaries than in standard search results in its sample. The implication is not that AI search always hides diverse sources. Rather, synthesis changes how users encounter them. A person may see five sources represented inside one generated explanation instead of opening five pages and comparing them directly. That makes source selection an increasingly important part of the search engine’s internal process.
The New Gatekeeper Is the Retrieval-and-Synthesis System
Traditional search ranking already gave search engines considerable influence over what users saw. AI search adds another layer of selection.
The system must decide:
- which sources to retrieve;
- which passages are relevant;
- how conflicting information should be handled;
- which evidence belongs in the answer;
- how the information should be summarized;
- Which sources should be cited
- and when the available evidence is insufficient.
Google explicitly warns that AI Mode can make mistakes, misinterpret web content, or miss context, and advises users to check important information against multiple sources. This is a different failure mode from a conventional search result being poorly ranked. With traditional search, a bad result can often be recognized as a bad result before it influences the user’s conclusion. With an AI-generated synthesis, several individually reasonable pieces of information can be combined into an answer that sounds coherent but contains an incorrect inference. The problem is therefore not only retrieval accuracy. It is synthesis accuracy.
AI Search Also Changes What a Website Needs to Be
The implications for publishers extend beyond traffic. A webpage can now serve at least three different functions: it can be the destination a user visits, a source used to construct an AI answer, or a brand/entity that the AI system recognizes and cites. These functions do not necessarily produce the same commercial value.
A website might receive an impression in an AI answer without receiving a visit. It might receive a small number of highly motivated visitors from an answer citation. Or its information could influence the generated response while the user never sees the original page. Google has begun giving publishers more visibility into this new environment. In June 2026, Google announced dedicated Search Console reports for generative AI features and said that by August 31 the insights had been rolled out worldwide. The reports show impressions, pages, countries, devices, and dates for visibility within AI features in Search and Discover.
That is significant because conventional search analytics were built around clicks and page visits. AI discovery creates situations where visibility can matter even when the user does not click. The industry is consequently starting to measure a new layer between ranking and traffic.
ChatGPT and Other AI Search Systems Add Another Route to the Web
Google is not the only company changing information discovery. OpenAI’s current publisher documentation says public websites can appear in ChatGPT search and that publishers should allow OAI-SearchBot if they want their content to be discoverable and surfaced with citations and links. OpenAI also says referral URLs from ChatGPT search include a UTM parameter that allows publishers to track inbound traffic.
The technical architecture varies between products, but the broader pattern is similar: an AI system retrieves information from the web, interprets it in the context of a user’s request, and presents an answer that can include links back to sources. That means the web is becoming a source layer for multiple answer engines rather than only a destination reached through conventional search rankings. For users, the difference can be difficult to see. For publishers, it can be substantial.
Discovery Is Becoming Multimodal
Another change is that “search” no longer has to begin with text. Google’s AI Mode accepts images and files in addition to typed questions and voice. This allows a person to begin with an object, document, screenshot, or visual problem rather than a carefully constructed keyword query. That matters because many real-world information needs are difficult to express with keywords. A user looking at an unfamiliar connector, a damaged component, a chart, or a document can provide the thing itself and ask what it means.
AI systems can interpret the input and then search for supporting information. The result is a shift from “What words should I type?” toward “What information can I give the system?”
The Risk of Losing the Original Context
The convenience of synthesis has a corresponding cost: context can disappear. A source article may contain qualifications, dates, methodological details, and exceptions that are difficult to preserve in a short generated answer. A research paper may distinguish between correlation and causation. A government document may describe a rule that applies only to a particular jurisdiction. A product specification may have conditions that are easy to omit during summarization.
The more aggressively information is compressed, the greater the importance of source inspection. This is why citations are becoming more than a convenience. They provide a path back to the evidence from which the answer was constructed. But citation presence is not the same as citation quality. A source can be relevant to a response without supporting every statement the generated answer makes. Users still need to inspect important claims, particularly when decisions depend on them.
What the Change Means for the Open Web
The biggest unresolved issue is economic. The open web depends on organizations and individuals continuing to create useful information. Many of those publishers depend, directly or indirectly, on audience traffic, advertising, subscriptions, affiliate revenue, or other forms of monetization. If AI systems can answer more questions without sending users to the original sources, the incentive structure can change.
This creates a difficult feedback problem. AI search needs high-quality web information to produce useful answers. But if answering questions reduces the traffic and revenue available to information producers, fewer organizations may have the economic incentive to create expensive original material.
That does not prove that the web will experience a collapse in original publishing. Nor does it mean AI search necessarily harms every publisher. Some sites may gain visibility from citations, attract higher-intent visitors, or benefit from being surfaced in new discovery channels. The important point is that the old relationship—search result visibility leading predictably to a website visit—is becoming less reliable.
What Is Confirmed, and What Is Still Uncertain?
By September 27, 2026, several developments are clear.
| Development | Current status |
|---|---|
| AI-generated answers inside Google Search | Established and expanding |
| Google AI Mode | Established search experience with conversational follow-up and multimodal input |
| Query fan-out | Documented by Google as part of AI search systems |
| AI search affecting click behavior | Supported by observational and experimental research, with results varying by study |
| Dedicated Google reporting for generative AI visibility | Rolled out worldwide by August 31, 2026 |
| ChatGPT web search for publishers | Established, with publisher crawling and referral mechanisms documented by OpenAI |
| Long-term effect on web publishing economics | Still uncertain |
The evidence therefore supports a narrower conclusion than some industry commentary suggests. AI search is clearly changing how users interact with search engines. People can ask longer questions, maintain conversational context, use multimodal inputs, and receive synthesized information without opening every underlying source. Google itself reports substantial growth in AI Mode usage, while independent studies show changes in publisher click behavior. What remains uncertain is the long-term balance between convenience and source discovery.
The Search Engine Is Becoming an Information Interpreter
The web has not disappeared behind AI search. It has become less visible during some searches. That distinction matters. The underlying pages, databases, research papers, government documents, forums, and news organizations still provide much of the information that AI systems retrieve. But the user increasingly encounters that information through an intermediary that selects, combines, and explains it. The result is a new information-discovery model: search is moving from a system that primarily helps people find documents toward one that increasingly helps them construct answers.
That can reduce the effort required to research a subject, especially when the question spans several sources or requires multiple searches. It can also introduce new dependencies on the AI system’s retrieval choices, synthesis quality, and citation behavior. For publishers, the change is equally fundamental. Being present on the web is no longer only about earning a position that produces a click. Visibility can occur inside an AI-generated response, while the measurement of that visibility is still developing. What happens next will depend on whether AI search can maintain reliable connections to original sources while preserving incentives for those sources to keep producing information. The technology has already changed the path people take to answers. The unresolved question is how that new path will reshape the web that supplies the answers in the first place.