Is This a Paradigm Shift? Testing the AI-Search Anomaly Against Kuhn's Model

What Kuhn actually argued, and what he did not

In 'The Structure of Scientific Revolutions' (1962), Thomas Kuhn described science as moving in two modes. Most of the time it does 'normal science', solving puzzles inside an agreed framework, a paradigm, whose rules nobody questions. A paradigm shift happens only when that framework accumulates anomalies it cannot explain, enters a period of crisis in which its own tools stop producing trustworthy results, and is displaced by a rival framework that explains the anomalies better. The shift, in Kuhn's account, is not gradual refinement; it is a change in the questions the field considers legitimate.

Two cautions have to be stated before the analogy is used at all. First, Kuhn was writing about physics and chemistry, not commerce, and he was explicit that a paradigm shift is a rare, structural event, not a synonym for 'big change'. Applying it to digital marketing is a borrowed lens, useful for discipline, never a literal claim. Second, the value of the lens is that it can return a negative. A framework that can only ever confirm the exciting conclusion is not a test. The reason to use Kuhn here is precisely that his model gives us conditions a transition can fail to meet.

The test: four conditions, checked against real evidence

If the shift from ranked lists to synthesized answers is a genuine paradigm change rather than another format cycle, it should satisfy the conditions Kuhn set out. There should be anomalies the incumbent model cannot account for. There should be a crisis in the reigning instruments, the tools practitioners trust to measure their work. There should be a rival framework with its own methods and literature. And, for a full shift rather than a coexistence, the old paradigm should be substantially displaced rather than merely supplemented.

Taken one at a time, against primary data rather than vendor enthusiasm, these conditions do not all resolve the same way. That split is the finding, and it is more useful than a headline.

Condition one: anomalies the ranked-list model cannot explain

The classic model of search assumed a straightforward exchange: the engine ranks pages, the user clicks one, the click delivers a visit. The strongest anomalies are behaviors that break that exchange, and they are measured on independent clickstream panels rather than asserted.

Zero-click search, a query that ends without any click to an external site, is the clearest. SparkToro's 2024 study with Datos clickstream data put US zero-click at 58.5 percent; a 2026 follow-up using Similarweb panel data reported 68.01 percent for early 2026, a rise of about 9.51 points in roughly two years. The direction is consistent across two independently sourced, methodology-disclosed studies.

The behavior sharpens when a generative summary is present. The strongest primary anchor is the Pew Research Center's July 2025 study, which tracked the actual browsing of 900 consenting US adults across 68,879 Google searches in March 2025. Users clicked a traditional result in 8 percent of searches that showed an AI summary, against 15 percent for searches without one, and clicked a link inside the summary itself only about 1 percent of the time. A ranked-list model of value, in which position drives clicks, cannot easily explain a page that is 'seen' but generates no visit. On this condition the anomaly is real and independently corroborated.

Condition two: a crisis in the reigning instruments

Kuhn's crisis is not only about surprising results; it is about the trusted tools quietly ceasing to measure what practitioners think they measure. Here the digital-marketing case is genuinely unsettled, which is why this condition reads as partly met rather than cleanly met.

The dominant measurement apparatus of search descends from the link-based ranking model that Brin and Page formalized in their 1998 PageRank paper, in which a page's authority is inferred from the links pointing to it. That apparatus is very good at answering 'what position does this page hold for this query'. It has no native way to answer 'how often is this business named inside a synthesized answer', because a citation inside a generated paragraph is not a rank and is not deterministic across repeated queries. The instrument still works for the surface it was built for; it is simply blind to a surface that now carries buyer attention.

The old instruments are not broken so much as incomplete. Rank tracking still measures classic search accurately, and classic search still matters. The crisis, such as it is, lies in mistaking a rank report for a full picture of visibility. That is a real gap, but it is a gap in coverage, not yet a collapse of the reigning method.

Condition three: a rival framework with its own literature

A format change becomes a candidate paradigm when a distinct body of method grows up to explain it, with its own object of study. On this condition the evidence is unusually clear for so recent a shift.

In 2024, researchers from IIT Delhi, Princeton, and Georgia Tech published 'GEO: Generative Engine Optimization' at KDD, one of the top data-mining venues. The paper defines 'generative engines' as a genuinely new response paradigm, synthesized answers rather than ranked links, builds a benchmark, and tests interventions for whether they change how often a source is cited inside a generated answer. It reports that adding cited statistics, quotations, and authoritative sources measurably raised a source's visibility in the engines it tested.

The significance is not any single tactic; it is that the object being optimized has changed. The incumbent discipline optimizes for a position in a list. The rival framework optimizes for being named and cited inside an answer. That a peer-reviewed literature now formalizes citation, not rank, as the target is the strongest evidence that this is more than a cosmetic SERP feature. On this condition, a rival framework with its own methods demonstrably exists.

The case against: this looks a lot like another format cycle

The other side of the argument deserves the same rigor. The search results page has been re-formatted many times, universal search, featured snippets, knowledge panels, local packs, and the industry adapted each time without abandoning its underlying craft. A skeptic can reasonably read AI Overviews as the next entry in that sequence rather than a rupture, and the forecasting record gives that skeptic real ammunition.

The most-cited 'search is collapsing' prediction has not held as stated. In February 2024 Gartner forecast that traditional search-engine volume would fall 25 percent by 2026 as AI chatbots absorbed queries. Reporting that checked the prediction against 2026 reality found that Google still holds more than 90 percent of the search market and that traditional search volume had not dropped 25 percent; the observed pattern was AI answers embedded inside Google's own results rather than volume defecting to competitors. This is exactly the kind of evidence a paradigm-shift claim must survive, and the fourth Kuhn condition, wholesale displacement of the old paradigm, is not met.

There is a structural point underneath the failed forecast. The 2026 SparkToro analysis found that AI Overviews appear on more than 20 percent of Google searches and, when present, cut click-through by close to 60 percent. Both facts can be true at once: the incumbent surface persists at scale and a new surface is materially reshaping behavior on the queries it touches. A partial, coexisting shift is not the clean replacement Kuhn described in the sciences.

The verdict

Scored against the four conditions, the transition returns a genuinely mixed result. Anomalies the ranked-list model cannot explain: met. A rival framework with its own literature: met. A crisis in the reigning instruments: partly met, a gap in coverage rather than a collapse of method. Wholesale displacement of the old paradigm: not met, and the confident forecasts that predicted it have so far been wrong.

A responsible reading declines to call this a completed paradigm shift and equally declines to dismiss it as nothing. What the evidence supports is a real structural change, unevenly distributed across queries and surfaces, whose eventual magnitude is not yet knowable from the data in hand. Kuhn himself noted that participants rarely recognize a shift while they are inside it; the label is applied in retrospect, once the record is long enough to judge. That is the position the evidence supports, and it has a direct operational consequence.

Why the answer is to measure, not to assert

If the question cannot be settled by assertion today, then the only defensible response is to build the evidence that will settle it. A single reading of where a business stands across classic search, the local map pack, AI answers, and reputation is a snapshot; a disciplined series of such readings, taken the same way over time, is data. The difference between opinion and evidence in this field is a time series that was measured consistently rather than remembered selectively.

That is the reasoning behind treating visibility as something to instrument continuously rather than to declare once. The contestation documented above, forecasts that missed, vendor figures that vary by denominator, real anomalies alongside real persistence, is not a reason to wait for certainty. It is the reason to start measuring now, so that when the record is long enough to judge, it is your own record answering the question rather than someone else's headline.