Generative Search and the Platformization of Sensemaking

long note sensemaking platform power

The gate is no longer only before access. It is also inside interpretation.

00 Retrieval becomes answer

Search engines have never been neutral maps of the web. Long before generative AI entered search interfaces, search engines ranked, filtered, and ordered information in ways that shaped what users could plausibly know. The literature on algorithmic gatekeeping, platform governance, and search engine ethics has already established that visibility is a form of power: to rank is to allocate attention, to make some sources appear authoritative, and to make others practically disappear even when they remain technically accessible.

Generative search does not replace that older problem. It adds a new stage to it. When a search interface returns a synthesized answer before the user has inspected the underlying materials, the system is no longer only arranging access to sources. It is participating in the user’s first act of sensemaking.

That distinction is small at the level of interface design and large at the level of knowledge. A ranked list asks the user to move among sources and assemble an initial account. An answer-oriented interface performs part of that assembly in advance. The user may still click through, compare sources, and reject the system’s synthesis, but the sequence has changed. The first coherent account is no longer constructed primarily through source navigation; it is increasingly delivered as an answer.

The issue, then, is not whether AI summaries are useful. They often are. They reduce search costs, support cross-language access, and help users enter unfamiliar topics. The harder question is what happens when the convenience of synthesis becomes the default mode through which people encounter public knowledge. At that point, the politics of search is no longer only a politics of visibility. It becomes a politics of mediated sensemaking.

01 Gatekeeping after the ranked list

The strongest way to frame this shift is not to invent a new theory of platform power, but to connect generative search to existing work on cognitive authority, sensemaking, and algorithmic gatekeeping.

Patrick Wilson’s account of cognitive authority remains useful because most knowledge is second-hand. People rely on sources they regard as competent, trustworthy, or institutionally credible. Search engines have long participated in that process by making certain sources appear more relevant than others. Dervin’s sense-making theory adds a second point: information seeking is not merely retrieval, but the construction of meaning under conditions of uncertainty. Users search because there is a gap to be bridged, not merely because there is an object to be fetched.

Generative search intervenes precisely at that gap. It does not only present candidate sources that might help the user make sense of an issue. It offers a provisional bridge across the gap. The answer becomes a cognitive shortcut: not simply a pointer to knowledge, but an initial organization of it.

This is why the existing language of algorithmic gatekeeping needs to be extended rather than abandoned. Classic gatekeeping decides what enters circulation; algorithmic gatekeeping decides what becomes visible and in what order. Generative search preserves both functions while adding a further operation: it compresses selected materials into a coherent explanatory surface. The gate is no longer only located before access. It is also located inside interpretation.

The point is not that platforms now control meaning in any total sense. Users remain active. Sources remain available. Competing interpretations still circulate. The point is more precise: answer-oriented systems increasingly occupy the default position from which ordinary users begin interpreting a topic under conditions of limited time, limited attention, and uneven expertise.

02 The epistemic sequence

The practical difference between a result page and a generated answer is a difference in sequence.

In a conventional search workflow, the user first encounters source plurality. Even if ranking has already shaped the field, the user still sees multiple documents as separate objects. The first task is source triage: deciding which materials deserve attention. Interpretation emerges through movement across these objects, however shallow or biased that movement may be.

sequence shift In an answer-oriented workflow, the user first encounters a composed response. The first task becomes response triage: deciding whether the answer is adequate enough to continue, stop, or ask a follow-up. Source inspection becomes secondary. It is not removed, but it is displaced to a later stage, after the system has already established salience, context, and a provisional frame.

This is why the change cannot be reduced to summarization. Summarization shortens a prior text. Generative search often produces an answer from a heterogeneous field of materials, ranking some claims as central, backgrounding others, and smoothing differences into readable prose. The output is an interpretive object. It tells the user not only what information exists, but what the issue appears to amount to.

The consequences are strongest for queries that are exploratory, ambiguous, contested, or value-laden: public affairs, health, education, consumer advice, migration, work, cultural interpretation, and personal decision-making. In these domains, the first answer does not merely save time. It can determine which distinctions are noticed, which uncertainties are preserved, and which forms of disagreement remain visible.

The behavioral evidence is still emerging, but the pattern is already clear enough to support a bounded claim. AI summaries are not simply placed beside traditional results; they alter what users do next.

Pew Research Center’s browsing-data study of Google Search in March 2025 examined 68,879 searches from 900 U.S. adults. When an AI summary appeared, users clicked a traditional search result in 8 percent of visits, compared with 15 percent when no AI summary appeared. They clicked a link inside the AI summary in only 1 percent of visits. They were also more likely to end the browsing session after encountering an AI summary: 26 percent of pages with an AI summary, compared with 16 percent of pages with only traditional results. Those figures do not prove that users trust AI summaries uncritically. They show something narrower and more important: answer surfaces can become sufficient for the immediate task often enough to reduce movement into the source field.

This behavioral tendency matters because the reliability of the answer layer remains fragile. The Tow Center for Digital Journalism tested eight generative search tools on 1,600 news queries designed around direct excerpts from real articles. The tools collectively provided incorrect answers to more than 60 percent of queries, often with confident phrasing. The study’s narrow significance lies in testing whether systems can perform a task that ordinary search could handle relatively straightforwardly, namely identifying the source of a known excerpt. When systems fail at this task while maintaining an authoritative conversational tone, the issue is not simply error. It is the production of apparent grounding without reliable attribution.

The BBC and European Broadcasting Union reached a related conclusion at a larger cross-market scale. In their 2025 study, 22 public service media organizations across 18 countries and 14 languages evaluated more than 3,000 responses from leading AI assistants on news and current affairs. Almost half of the evaluated answers contained at least one significant issue, and sourcing was the most common major problem. This matters because news knowledge is not only a matter of isolated facts. It depends on attribution, context, chronology, and the distinction between reporting, opinion, and institutional claims. A system that blurs those distinctions can distort public knowledge even when parts of the response are factually plausible.

Together, these studies do not justify a sweeping claim that generative search always misleads users or that ordinary search was epistemically clean. They justify a more defensible claim: answer-oriented systems are increasingly used as initial knowledge interfaces, while their practices of attribution and contextualization remain unreliable enough to make first-pass interpretation a governance problem.

04 Transparency is not contestability

Many AI search interfaces attempt to solve the problem by adding citations, source cards, or links. This is necessary, but not sufficient. A citation is not the same as a traceable evidential relation. A link may indicate that some source is associated with the answer without showing which claim it supports, how strongly it supports it, or whether the source has been represented accurately.

This is where work on transparency and accountability is especially relevant. Ananny and Crawford warn that visibility does not automatically produce accountability, because what becomes visible is itself structured by institutions and interfaces. Michael Power’s account of the audit society makes a related point: verification can become ritualized. Systems may generate the appearance of accountability without creating the conditions for meaningful contestation.

In generative search, apparent transparency can therefore become part of the problem. A response with citations may look more accountable than an uncited answer, but if the citations are weakly connected to the claims, the interface can borrow legitimacy from sources it has not faithfully represented. The burden of verification is shifted back to the user, while the interface retains the benefit of appearing grounded.

For search to remain contestable, the user must be able to move from claim to source, from source to context, and from context to disagreement. If the system compresses these relations into a fluent answer and then attaches generalized citations, it has not preserved contestability. It has aestheticized it.

05 Coherence can close disagreement

Generative systems are optimized for helpfulness, fluency, and coherence. These qualities are valuable, but they also create a structural risk. Many questions do not have a single stable synthesis. They involve disagreement among experts, conflicts between institutional perspectives, changing evidence, jurisdictional variation, or unresolved moral and political stakes.

A result page can distort such questions through ranking, but it often leaves disagreement materially visible as separate source objects. A generated answer tends to absorb disagreement into a single surface. The danger is not only that the answer may be false. It may be too coherent.

Premature coherence changes the user’s relation to uncertainty. Instead of encountering a field in which disagreement must be navigated, the user encounters a response in which disagreement has already been managed. The system may mention that views differ, but the form of the answer still imposes an order: this is the main account, this is the context, this is the relevant caveat. Such ordering is unavoidable in any act of synthesis. The question is whether the system makes that ordering inspectable.

This is why the central standard should not be neutrality. No interface is neutral. The more realistic standard is recoverability: can a user recover the structure of evidence, disagreement, and uncertainty that the answer has compressed? When recoverability is weak, the answer layer becomes an authority not because users are forced to believe it, but because the cost of reconstructing what it has done is too high for ordinary use.

06 Cultural standardization

The same logic extends beyond factual search. Generative systems also shape the cultural forms through which issues become intelligible. They tend to privilege expression that is portable, orderly, explicit, and easily summarized. These are not inherently bad qualities, but they are not culturally neutral.

Research on LLM summarization and writing assistance supports this concern. Subbiah and colleagues, working with professional writers, found that models could produce coherent short-story summaries while still making faithfulness errors and struggling with specificity and difficult subtext. Agarwal, Naaman, and Vashistha show that AI writing suggestions can push users toward more generic and Westernized styles, diminishing culturally specific nuance. These findings do not mean that every generated answer erases culture. They show that fluency and efficiency can come with a standardizing pressure.

For search and public knowledge, the implication is straightforward. If answer systems repeatedly make some forms of explanation easier to retrieve and others harder to formulate, they influence not only what users know, but what kinds of articulation feel natural. Minority idioms, local categories, ambiguous literary forms, non-Western rhetorical structures, and unresolved experiential knowledge may remain present in the source field while becoming less likely to appear as the first frame.

The pattern is closer to classification and standardization, the terrain Bowker and Star analyzed in their work on infrastructures of categorization. What matters is not only what is excluded, but what becomes easy, default, and administratively legible. Generative search inherits this infrastructural politics and expresses it through language.

07 Personal orientation

The stakes are not confined to public affairs. Users increasingly ask AI systems for help with careers, education, health, migration, relationships, creativity, and identity. Such queries are informational, but they are also interpretive. They ask the system to turn uncertainty into a usable description of a situation.

Here again, cognitive authority matters. When a system offers a fluent account of what a user’s problem is and what options are reasonable, it supplies categories through which the user may understand themselves. This can be helpful, especially when the user lacks access to expert advice. But it can also over-format experience. Problems rooted in institutions may be translated into personal optimization. Political constraints may be redescribed as planning challenges. Cultural or existential conflicts may be rendered as productivity, mindset, or therapeutic language.

The issue is not that these framings are always wrong. The issue is that they often arrive first, in a polished and actionable form. Alternative framings may require more expertise, more precise prompting, or more resistance from the user. The answer does not determine the future, but it can narrow the set of futures that appear immediately intelligible.

08 Design and governance

If generative search is understood as AI-mediated sensemaking, then the relevant design question is not simply how to make answers more accurate. Accuracy is necessary, but insufficient. A system can be accurate in fragments while still misrepresenting uncertainty, flattening disagreement, or obscuring the evidential status of its claims.

The stronger standard is contestable synthesis. A contestable answer would not merely cite sources; it would make the relation between claims and sources visible. It would distinguish primary evidence from commentary, reporting from opinion, consensus from dispute, and stable knowledge from provisional interpretation. It would preserve enough of the underlying source structure for the user to understand what has been compressed.

The standard is precise: concise answers can remain concise, while systems must stop treating fluency as a substitute for epistemic accountability. In low-stakes factual queries, a direct answer may be adequate. In contested or consequential domains, the answer should expose rather than conceal the structure of uncertainty.

The governance implication is equally direct. Regulators, researchers, and civil society should evaluate answer-oriented systems not only through hallucination rates, but through source attribution, claim-source alignment, disagreement preservation, and framing variance across languages, regions, and user profiles. These are empirically testable properties. They move the critique away from general suspicion and toward observable failures of mediation.

09 First account as platform function

The web remains abundant. More people can publish than in earlier media environments, and original sources often remain available. Yet abundance at the level of publication does not guarantee plurality at the level of first interpretation.

Generative search concentrates a crucial moment in the knowledge process: the moment when scattered materials become a first usable account. That concentration is not absolute, and it does not make users passive. But it does shift platform power from the ranking of sources toward the mediation of sensemaking.

The central question is therefore not whether AI systems should answer questions. They already do, and often they do so usefully. The question is whether the answers preserve the conditions under which users can understand, contest, and revise the synthesis they have been given.

A healthy information environment does not merely provide answers more quickly. It keeps open the path from answer back to evidence, from evidence back to disagreement, and from disagreement back to judgment. Without that path, generative search risks turning the first account of the world into a platform-managed surface: fluent, convenient, and difficult to dispute at the moment when dispute matters most.

References

Agarwal, D., Naaman, M., & Vashistha, A. (2025). AI Suggestions Homogenize Writing Toward Western Styles and Diminish Cultural Nuances. Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems.

Ananny, M., & Crawford, K. (2018). Seeing without knowing: Limitations of the transparency ideal and its application to algorithmic accountability. New Media & Society, 20(3), 973–989.

Bowker, G. C., & Star, S. L. (1999). Sorting Things Out: Classification and Its Consequences. MIT Press.

Bucher, T. (2012). Want to be on the top? Algorithmic power and the threat of invisibility on Facebook. New Media & Society, 14(7), 1164–1180.

Dervin, B. (1983). An overview of sense-making research: Concepts, methods, and results to date. International Communication Association Annual Meeting.

European Broadcasting Union & BBC. (2025). News Integrity in AI Assistants: An International PSM Study.

Gillespie, T. (2014). The relevance of algorithms. In T. Gillespie, P. Boczkowski, & K. Foot (Eds.), Media Technologies: Essays on Communication, Materiality, and Society. MIT Press.

Introna, L. D., & Nissenbaum, H. (2000). Shaping the web: Why the politics of search engines matters. The Information Society, 16(3), 169–185.

Jaźwińska, K., & Chandrasekar, A. (2025). AI Search Has a Citation Problem. Columbia Journalism Review / Tow Center for Digital Journalism.

Memon, S. A., & West, J. D. (2024). Search engines post-ChatGPT: How generative artificial intelligence could make search less reliable. arXiv:2402.11707.

Napoli, P. M. (2015). Social media and the public interest: Governance of news platforms in the realm of individual and algorithmic gatekeepers. Telecommunications Policy, 39(9), 751–760.

Noble, S. U. (2018). Algorithms of Oppression: How Search Engines Reinforce Racism. NYU Press.

Pasquale, F. (2015). The Black Box Society: The Secret Algorithms That Control Money and Information. Harvard University Press.

Pew Research Center. (2025). Google users are less likely to click on links when an AI summary appears in the results.

Poell, T., Nieborg, D., & Duffy, B. (2021). Platforms and Cultural Production. Polity.

Power, M. (1997). The Audit Society: Rituals of Verification. Oxford University Press.

Striphas, T. (2015). Algorithmic culture. European Journal of Cultural Studies, 18(4–5), 395–412.

Subbiah, M., Zhang, S., Chilton, L. B., & McKeown, K. (2024). Reading Subtext: Evaluating Large Language Models on Short Story Summarization with Writers. Transactions of the Association for Computational Linguistics, 12, 1290–1310.

Wilson, P. (1983). Second-Hand Knowledge: An Inquiry into Cognitive Authority. Greenwood Press.