Vernacular Suppression, the Four Treasuries, and the AI Search Pivot

On May 19, 2026, at the Shoreline Amphitheatre in Mountain View, Sundar Pichai walked onstage to introduce a redesigned Google search box. The same field that had organized roughly twenty-five years of web behavior now stretched into a long-form input, anticipated multi-clause questions, and routed the query into AI Mode without requiring the click that had previously separated typing from answering. Behind it sat what the company calls query fan-out: a single user question decomposed into hundreds of subqueries issued in parallel across Google’s index, then synthesized into a single response. AI Mode, launched in March 2025 and past one billion monthly active users by May 2026, had absorbed search functionality that built the enterprise and brand synonymous with finding information.

A single text field, centered on white space, retained the minimalist sobriety that has signaled algorithmic neutrality since the 1998 launch of google.com. The interface announced an enlarged input field. The interpretive function it now performed was kept off the stage. The lineage of this gesture, in design terms, is long. The plain white page of the Bauhaus printed text, the Helvetica-on-Swiss-grid neutrality of the postwar corporate identity, the deliberately unmarked terminal prompt of the Unix shell: each insists that the absence of ornament is a guarantor of objectivity. Google’s interface has, since its earliest days, traded on this same visual rhetoric. The May 2026 demonstration preserved that rhetoric intact and quietly transposed what occurs behind it.

Google’s published position is that nothing of consequence has changed. AI Overviews still send “billions of clicks” to the web. Traffic patterns will follow familiar routes. The open web remains intact. The data, reported through 2025 and into 2026 by Chartbeat, the Pew Research Center, Ahrefs, and the legal filings of Penske Media, the European Publishers Council, and Chegg, describes a different situation. Publisher referrals from Google declined globally by approximately one-third in the year to November 2025, with United States declines closer to thirty-eight percent (Newman, 2026). On informational queries, click-through rates from the position-one organic result collapsed from 1.76 percent in 2024 to 0.61 percent in 2025. The Pew Research Center’s behavioral study of 68,879 actual Google searches across 900 United States adults in March 2025 found that users clicked traditional search results in eight percent of visits when an AI summary appeared, compared to fifteen percent when no summary was present, and clicked the citation links inside the summary itself in only one percent of cases (Chapekis et al., 2025). The Planet D, a travel publication built over twenty years of original reporting, shut down after a ninety percent traffic collapse following the Overviews rollout (Allyn, 2025).

While the  interface may persist, beneath it, the indexed world’s horizons are contracting.

Historical Pivots on Information Control

THis crisis is not novel. There are some great historical analogs to tie back to, when considering the impact of narrowing information access. The first is the Vulgate regime, the centuries-long Catholic Church monopoly on scripture that suppressed vernacular translation, executed translators, and destroyed indigenous textual cultures, all while preserving mediation as the form through which laypeople would encounter the sacred text. The second, from China, is the Siku Quanshu regime, the Qianlong Emperor’s late-eighteenth-century compendium that absorbed Chinese textual culture into a single authoritative library and rendered everything outside that library effectively invisible.

Both were curation projects that presented themselves as preservation. Both produced effects that took generations to detect and longer to repair. Both presumed a particular kind of human being, the layperson or untrained reader, who needed an interpretive class to stand between them and the source. Both produced that interpretive class as the durable institutional residue of their work.

The present case carries a third overlay that neither earlier regime had access to. The apparatus now also functions as a behavioral profiling system that monetizes the user being mediated, not merely the act of mediation. Google has built, in a single product surface, the architecture of Vulgate authority, the architecture of imperial compendium, and the architecture of surveillance capitalism. The historical analogy clarifies what ordinary commentary on AI search misses. The comparison sharpens the question of what infrastructure must exist alongside the synthesis layer to keep direct access to source material available for those who go looking.

Tyndale’s Plowboy, Landa’s Bonfire

The Vulgate regime is remembered for the bodies it left in its wake. William Tyndale, strangled and burned at Vilvoorde in October 1536. Jan Hus, burned at Constance in July 1415. John Wycliffe himself, dug up forty-four years after his death and burned posthumously by order of the Council of Constance, his ashes scattered into the River Swift. The bodies of these people travel through history more easily than the transgressions that led them to their fates.

The Constitutions of Oxford, promulgated in 1407 in response to the Wycliffite movement of the late fourteenth century, made the unauthorized translation of scripture into English a capital matter of heresy. Wycliffe’s followers, who had produced the first complete English-language version of the Bible from the Latin Vulgate in 1382, were hunted as Lollards for a century before Tyndale’s birth. Manuscript fragments were seized. Owners were forced to recant. Some were burned alongside their books. When Tyndale moved to translate directly from the Greek and Hebrew, and to use Gutenberg’s press to outrun the enlightened religious that wanted him dead, he was extending an argument the Wycliffites had begun a hundred and fifty years earlier. By the time of his execution, eighteen thousand copies of his New Testament had circulated in England. Two complete volumes survive (Britannica, n.d.-b).

The Vulgate scriptural intelligentsia preserved Latin because the laity could not read it. An interpretive class therefore had to stand between the layperson and the text. The Mass conducted in Latin, the homily delivered in the vernacular, and the scripture rendered legible only through priestly explication formed a single machine of knowledge control. To attack one piece was to attack the whole. Tyndale understood this with unusual clarity. He framed his project as the restoration of direct access; the theological corrections were beside the point. The plowboy who knows more scripture than the priest threatens the machine from outside the heresy frame entirely – he stands outside the mediation.

That same regime, transposed across an ocean and into a different cosmology, produced the auto-da-fé at Maní on July 12, 1562. The Franciscan friar Diego de Landa, having determined that the Maya codices placed before him contained “nothing in which there was not to be seen superstition and lies of the devil,” ordered the destruction of approximately twenty-seven codices and roughly five thousand sacred images in a single public ceremony. The Mayanist William Gates would later remark that “ninety-nine percent of what we today know of the Mayas, we know as the result either of what Landa has told us in the pages that follow, or have learned in the use and study of what he told,” and that, in the same sentence, “he burned ninety-nine times as much knowledge of Maya history and sciences as he has given us” (Britannica, n.d.-a). Three pre-Columbian Maya codices and a fragment of a fourth survived. The last codex destruction at Tayasal occurred 135 years later, in 1697.

The bonfire at Maní was theater. The Spanish acto de fe means an act of faith, and the act was performed for a crowd. The codices were paraded. The friar read out what they were said to contain. The fire was lit, and the indigenous laypeople who had been reading those codices without priestly mediation watched a Franciscan mediator remove their reading material from the world. The Mass kept the layperson on the outside of the text by making the text unreadable. The bonfire kept the layperson on the outside of the text by making the text gone. Either way, the priest stayed in the middle.

The case for AI Overviews repeats this argument in transposed form. The texts of the open web are complex; the layperson is ill-equipped to interpret them safely; an authorized synthesis layer can give them what they actually need to know. Google’s argument that a user need not click through to a third-party page for content “not in dispute” reproduces the argument that the layperson need not read scripture for content the priest can explicate. The interpretive class has shifted from clergy to algorithm. The function is preserved.

Qianlong’s Wager

The Qianlong Emperor’s Complete Library in Four Sections, commissioned in 1772 and completed in 1782, occupies a different position in the family tree of information control. The Vulgate regime preserved mediation by suppressing alternatives. The Siku Quanshu preserved it by absorbing everything. The project was positioned as the most ambitious act of textual preservation in Chinese history. Local officials were instructed, by imperial edict, to collect every important work in the empire for inclusion. Owners were promised that books would be returned once the compilation was complete. Three hundred and sixty-one scholars worked on the editorial board for over a decade. The completed compendium ran to roughly 36,000 volumes containing approximately 800 million words. Seven manuscript copies were produced and housed in seven specially-constructed libraries across the empire, the Wenyuan Ge in the Forbidden City being the most prestigious. The architectural commitment matched the scale: each library was custom-built to display the imperial dimension of the project (Guy, 1987).

The real innovation lay inside the project rather than outside it. Embedded inside the comprehensive collection effort, hidden in plain sight, sat the largest book-banning campaign in Chinese history. A separate catalog, the Siku Jinhui Shumu, recorded the works that did not survive editorial selection. Roughly 2,855 titles were destroyed outright. Another four to five hundred were edited to remove anti-Manchu content (Goodrich, 1935). Owners who failed to surrender their copies were sometimes exiled with their families. Sometimes executed. The promise that books would be returned was not honored.

R. Kent Guy spent most of The Emperor’s Four Treasuries on the publishers. They watched what happened to colleagues whose books got called in, and they stopped printing in the categories the emperor’s editors were sifting (Guy, 1987, p. 207). The censors did not have to chase every dangerous book. The publishers did the chasing for them, before publication, by deciding which books to never set in type.

In 2024, Ying Bai, Ruixue Jia, and Jiaojiao Yang put numbers on Guy’s claim. They pulled publication data for over 161,000 books spanning the 1660s to the 1940s and tracked what happened to the categories that drew imperial scrutiny. The publishers exited first. The categories went silent for seventy years. Recovery began in the 1840s, when the Opium Wars knocked the dynasty’s hand off the throat of its own publishing industry (Bai, Jia, & Yang, 2024). By the time the imperial grip weakened, the people who had known how to publish those categories were dead, retired, or in other lines of work. Bringing the categories back required bringing the publishers back, and bringing the publishers back required a generation that had not learned to be afraid of those words.

Two kinds of curatorial regimes. Two different fates for the publishers. Rank an existing field of books and the books stay on the shelf: the publisher who ranks tenth still gets the buyers looking for what tenth has, and the publisher who ranks five-hundredth still has reason to print the five-hundredth book. Compile a single authoritative collection and tell readers that what is not in the collection is not worth reading, and the publisher of the unincluded book has no buyers, no rank, no presence. The field underneath rots out, because the collection is now where the reader goes. The Qing version ran on the speed of imperial edicts and printer-publisher economics. Seven decades. The Google version runs on the speed of advertising contracts and venture-funded publisher runways. The publishers cannot afford seven decades.

PageRank Was Never Neutral

The serious counter-argument to this framing is that Google was never neutral, and the AI overlay therefore represents continuation rather than rupture. Safiya Umoja Noble’s Algorithms of Oppression demonstrated, in 2018, that the ten ranked links of the standard Google search results were already a curatorial regime. Searches for “Black girls” surfaced pornographic content above academic resources. Autocomplete reproduced racist suggestions at scale. Commercial preference was laundered as algorithmic neutrality. The structural racism of the wider information environment was reproduced, and intensified, by the apparatus that claimed to transcend it (Noble, 2018). Matteo Pasquinelli, writing from a Marxist analytic position in 2009, described PageRank itself as a machine for extracting surplus value from the linking labor of the web’s users. Each hyperlink, in Pasquinelli’s reading, constituted a unit of cognitive capital that Google harvested for ranking signal without compensating its producer (Pasquinelli, 2009). Shoshana Zuboff’s account of surveillance capitalism documented the entire apparatus of behavioral data extraction underlying the search interface long before AI Overviews arrived, and identified that apparatus as the platform’s constitutive economic logic (Zuboff, 2019).

Pasquinelli in 2009. Noble in 2018. Zuboff in 2019. Three writers, three angles, the same finding: Google’s thumb was on the scale from launch. Calling the AI Overlay a fall from neutrality requires inventing a Google that never existed. The Overlay is the older Google with the costume off.

What the Noble Critique Could Not Yet See

This position is right about the history. It does not yet have the right explanation for why this moment is different. The Siku Quanshu is where they missed it.

PageRank had a thumb on the scale and a thousand-page field underneath it. The page existed. The algorithm ranked it. The user clicked. The publisher got paid, and the payment paid for the next page. The whole arrangement ran on links from one document to another, and the links were laid down by humans who had read what they were linking to. Bias steered the reader toward the first link. Bias did not delete the four hundred and ninety-ninth. A user who knew the operators (site: for a domain, quotation marks for a phrase, minus for an exclusion, cache: for an archive, the small “Cached” link Google used to ship at the bottom of every result) could walk past the first link and find the four hundred and ninety-ninth. The bias was the front door. The back doors stayed open.

The back doors were the thing. Journalists worked them. Librarians worked them. Researchers worked them. A reporter writing about a corporate scandal could pull a 2007 SEC filing past a 2024 PR statement by knowing how to ask. A graduate student could trace a misattributed quote back to the 1920 philology journal where it started by knowing how to ask. A parent who wanted to read the actual study behind a school-board memo could find the study by knowing how to ask. The Overlay does not have back doors. The Overlay is the door. Whatever the Overlay shows you is the room you walked into. The “Further Exploration” links Google added in early 2026 to soften the publisher revolt are picked by the same machine that wrote the answer. You do not get the ranked field anymore. You get the answer, and a curated way back out of the answer into whatever the answer was assembled from.

The Siku Quanshu is the precedent Noble’s framework did not yet need. Bai, Jia, and Yang traced the seventy-year publication collapse to a single change in what the publishers could earn from publishing (Bai, Jia, & Yang, 2024). The burned books were the headlines. The dead categories were the actual story. The publishers walked away because there was nothing left to publish for, and once the publishers walked, the categories went with them. The Pew, Chartbeat, and Ahrefs numbers are showing us the same exit, run at the speed of a venture-funded media economy instead of a Qing-era one. Independent travel writing is contracting because the people who used to write it are not getting paid. The same is happening to educational content (Chegg’s traffic down forty-nine percent), to service journalism, to mid-tier news, to reference desks for facts the Overview now considers settled, and to the long tail of niche expertise that survived on organic ranking and on nothing else.

The Field Was Already Rotting

The second serious counter-argument is that the loss was already underway. The open web by 2023 had been corrupted by SEO optimization, content-farm output, and the early flood of generative slop; calling it a healthy textual ecology displaced by AI synthesis romanticizes a corpus of words already in advanced decay. Arvind Narayanan and Sayash Kapoor document, in AI Snake Oil, the extent to which the indexable web by the early 2020s had become a fertile field in which AI-generated misinformation was already endemic. News websites published “error-filled AI-generated stories on important topics such as financial advice, and then refus[ed] to stop using the technology even after the errors came to light” (Narayanan & Kapoor, 2024, p. 22). The catalog has its defenders precisely because the field it replaces was already in full of fungus and fetid water..

The defenders of the Qianlong compendium made the same arguments at the time. The Ming-dynasty derivatives, the anti-Manchu polemics, the redundant commentaries that the editors removed were corruptions of the textual tradition, and the Siku Quanshu was the moment of long-needed authoritative consolidation. The argument was true at the level of any individual editorial decision the compilers made. It failed at the level of the institution doing the deciding, because the capacity to make any of those decisions belonged to a single authorizing body. Mediation, once monopolized, becomes the substrate of everything that follows, and the question of whether any particular act of mediation was correct on its merits becomes subordinate to the question of whether the monopoly itself is endurable. The Vulgate apparatus was administered, for most of its history, by people sincerely persuaded that they were preserving souls.

The Thing No Prior Mediator Had

Both counter-arguments hold under scrutiny. The genuinely novel element of the present case shoots past the threshold either represent. Behavioral profiling, deployed at the same scale and within the same AI-enabled search as the mediation layer, was not available to any previous mediating institution. The Vulgate priest knew nothing about the layperson kneeling in the confessional that the layperson had not chosen to tell him. The Siku Quanshu editors knew nothing about the readers for whom they were curating, or the anonymous bookseekers who perused the list of banned titles. The mediator class, in both historical cases, exercised authority over the act of access while remaining ignorant of the person accessing.

Personal Intelligence, the feature Google announced at I/O 2026 and is rolling out across nearly 200 countries and 98 languages, lets AI Mode draw on the user’s Gmail, Google Photos, and Google Calendar to personalize the synthesis. Each user receives an answer shaped by the contour of their own profiled past behavior, drawn from data the same company holds. The synthesis varies across users. It is computed against the individual profile and adjusted for what that user has previously demonstrated they will accept. The mediator class now interprets the reader as well as the text.

Surveillance capitalism’s critics correctly identified the architecture this convergence required, but lacked the consummating example. Zuboff named the process. Noble named the bias. Pasquinelli named the extraction. The AI overlay completes those three patterns by converging them into a single interface, one that absorbs the user’s question, knows the user’s history, and answers without ever returning the user to the ranked field of the underlying material.

What Should Cohabitate

Defeating the AI synthesis layer is not the solution. Consumer convenience pressure is moving in a single direction, and the market for AI-mediated search will continue to expand regardless of publisher harm. The question worth asking is what civic infrastructure must exist alongside the synthesis layer to keep direct, non-mediated access to source material available for the users, researchers, and producers who require it.

Two precedents exist and supply a working template. The Internet Archive, founded by Brewster Kahle in 1996, and Wikipedia, founded by Jimmy Wales and Larry Sanger in 2001, both demonstrate that civic-scale, non-extractive, openly-accessible knowledge infrastructure can be built and sustained outside the commercial platform regime. Kahle founded the Internet Archive with explicit reference to the Library of Alexandria, arguing that the digital age would lose its memory in ways comparable to the ancient cultural catastrophes if no countervailing preservation apparatus were constructed (Kahle, 1997). The Wayback Machine, the Archive’s most visible public service, now holds over one trillion archived web pages and functions as the only reliable record of significant portions of digital journalism, government documents, and personal publication from the first three decades of the web. Wikipedia, the only single text in human history larger than the Siku Quanshu, has operated for twenty-five years as a non-extractive open knowledge layer governed by community editorial process. Kahle himself described the shift his work was designed to resist with unusual precision: the digital transition, he argued, has moved from local control to central control, non-profit to for-profit, diverse to homogeneous, and from “ruled by law” to “ruled by contract.”

Neither precedent suffices on its own. Wikipedia gives the reader aggregated reference content with no search apparatus. The Wayback Machine gives the reader archived web pages with no retrieval optimized for question answering. The missing piece is a civic-scale open search and retrieval layer: a Wikipedia for knowledge access rather than knowledge content, a Wayback Machine that indexes the present instead of the past, a retrieval apparatus governed by transparent indexing and ranking rules instead of opaque algorithmic synthesis, owned by no commercial entity, and constitutionally incapable of the profiling overlay that turns each user into the substance being mediated. The technical components mostly exist already. Open web crawlers, federated indexes, semantic embedding models that run on consumer hardware, and the entire stack of open-source search tooling developed before the commercial platforms absorbed the consumer-facing interface remain available, mostly unused, in the open-source ecosystem. The institution that would assemble them and operate them at civic scale does not yet exist.

That institution cannot and will not be built by venture-funded startups. The economic logic of any extractive platform pulls toward mediation and profiling, because mediation and profiling are how the platform monetizes the act of access. It will not be built by Google or the other dominant platforms, which are restructuring their products precisely to capture the value the open web previously distributed. The institutional model that fits the requirement is the one already demonstrated by the Internet Archive and Wikipedia: a non-profit operating under a charter that legally forecloses profit-seeking, funded by a coalition of foundations, libraries, universities, public broadcasters, and individual donors, governed by a board with substantial representation from librarians, archivists, journalists, and information scientists who treat preservation and direct access as constitutive obligations rather than business decisions. Such an institution will require a sustained capital commitment on the order of hundreds of millions of dollars per year, indexed against the operating costs of running a web-scale crawler and retrieval system. The model exists. The institutional commitment does not yet.

The question is who builds it, on what timeline, and with what guarantees of independence from the commercial platforms whose synthesis layer it is meant to provide an alternative to.

The Plowboy’s Gambit

Tyndale wagered that scripture in the layperson’s own language would be more dangerous to the Church’s monopoly than any heresy the Church could prosecute. The Vulgate apparatus was correct to perceive this as an existential threat to its mediation, and the Vulgate apparatus lost. The Wycliffite manuscripts were burned, the Tyndale Bibles were burned, Tyndale himself was burned, and the printing press kept printing, and the plowboy eventually read the scriptures, and the priestly monopoly on interpretation came apart over the next two centuries.

The Siku Quanshu wager was the opposite. Comprehensive imperial preservation, presented as scholarship, would do the work of suppression more durably than the bonfire. The Qianlong apparatus was correct to perceive this as more effective than overt censorship. It worked. The publication collapse lasted seven decades. The recovery required the dynasty’s own collapse. For the intervening generations, the catalog was the field.

The present case combines the worst elements of both. Comprehensive preservation, marketed as preservation. Mediation that is also profiling. A synthesis layer that the technically literate cannot defeat by routing around it, because the synthesis has absorbed the surface that previously allowed routing. The bonfire is now economic rather than literal, but the chilling effect on the intermediaries is producing the same pattern of publisher exits, category silences, and slow collapse that took seven decades to detect in the Qing case. There is no obvious reason to expect faster recovery from this one.

The plowboy was the point. The plowboy is still the point. The question is what tools the plowboy has, and who is building them while there is still a field for the plowboy to read.

References

Allyn, B. (2025, July 31). Online news publishers face ‘extinction-level event’ from Google’s AI-powered search. NPR. https://www.npr.org/2025/07/31/nx-s1-5484118/google-ai-overview-online-publishers

Bai, Y., Jia, R., and Yang, J. (2024). Knowledge suppression and resilience under censorship: Three-century book publications in China (NBER Working Paper No. 33258). National Bureau of Economic Research. https://www.nber.org/papers/w33258

Britannica. (n.d.-a). Diego de Landa. In Encyclopaedia Britannica. https://www.britannica.com/biography/Diego-de-Landa

Britannica. (n.d.-b). William Tyndale. In Encyclopaedia Britannica. https://www.britannica.com/biography/William-Tyndale

Chapekis, A., Bestvater, S., and Remy, E. (2025, July 22). Google users are less likely to click on links when an AI summary appears in the results. Pew Research Center. https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/

Goodrich, L. C. (1935). The literary inquisition of Ch’ien-Lung. Waverly Press.

Guy, R. K. (1987). The emperor’s four treasuries: Scholars and the state in the late Ch’ien-lung era. Council on East Asian Studies, Harvard University.

Kahle, B. (1997, March). Preserving the internet. Scientific American, 276(3), 82-83. Republished commentary at Internet Archive Blog. https://blog.archive.org/2025/09/02/looking-back-on-preserving-the-internet-from-1996/

Narayanan, A., and Kapoor, S. (2024). AI snake oil: What artificial intelligence can do, what it can’t, and how to tell the difference. Princeton University Press.

Newman, N. (2026, January 16). Global publisher Google traffic dropped by a third in 2025. Press Gazette. https://pressgazette.co.uk/media-audience-and-business-data/google-traffic-down-2025-trends-report-2026/

Noble, S. U. (2018). Algorithms of oppression: How search engines reinforce racism. NYU Press.

Pasquinelli, M. (2009). Google’s PageRank algorithm: A diagram of cognitive capitalism and the rentier of the common intellect. In K. Becker and F. Stalder (Eds.), Deep search: The politics of search beyond Google (pp. 152-162). Studienverlag.

Zuboff, S. (2019). The age of surveillance capitalism: The fight for a human future at the new frontier of power. PublicAffairs.