The Aggregator Century: A Political-Economy History of Who Controls Discovery

The pattern begins: the store, the catalog, and the directory

The pattern this study traces did not begin with a server. It began with a shopfront. Aristide Boucicaut took over a small Parisian drapery called Le Bon Marche, founded in 1838, and from 1852 rebuilt it into what business historians generally credit as the first modern department store: fixed prices instead of haggling, goods organized into departments a shopper could browse without a clerk's permission, and a mail-order catalog and home delivery service that let the store's reach extend past anyone who could physically walk through its doors. The innovation was not the goods on the shelves. It was that Boucicaut had built a single destination a buyer could default to for an entire category of need, and every supplier who wanted access to that buyer now had to come to him on his terms, stocked, priced, and displayed as he saw fit.

The mail-order catalog took that same aggregation and stretched it across a continent. Richard Sears began as a railway station agent reselling a shipment of watches in 1886, and within a decade, joined by watch repairman Alvah Roebuck, had turned that side business into a catalog reaching rural America, where roughly two-thirds of the country still lived. By the late 1890s the Sears catalog ran past 500 pages, and within the following decade it listed more than 10,000 items covering, as company lore put it, everything from birth to death. The distribution numbers alone show the scale of the aggregation: Sears mailed 318,000 catalogs in 1897 and 3.6 million by 1908, according to History.com's account of the company's archives. A rural household with no department store for a hundred miles now had one place to look, and Sears, not the thousands of small manufacturers whose goods filled its pages, decided what appeared on which page.

The Yellow Pages did something narrower but arguably closer to what would come later. Reuben H. Donnelley published the first officially branded Yellow Pages directory in 1886, the same year Sears began, and built a business model with nothing to do with selling merchandise: he sold placement. A business paid the phone company or its publishing partner for a listing, and a larger, bolder, better-placed listing cost more. That advertising-funded directory model proved extraordinarily durable, peaking at 14.7 billion dollars in annual US advertising revenue in 2005, per longstanding industry accounts of the format's history, before the same buyers who once opened the book at the kitchen counter started typing their question into a search box instead. Department store, catalog, and directory had already worked out the core mechanics decades before anyone spoke of platforms: aggregate the buyer's attention, control the field of choice presented to that buyer, and charge the seller for a place inside it.

What the internet changed: Aggregation Theory as the analytical lens

The department store, the catalog, and the directory all did real aggregation work, and all of them hit a hard ceiling that the internet removed. A department store's reach stopped at the edge of a delivery radius. A catalog's reach stopped at the cost of printing and postage. A directory's usefulness stopped at the last page of the book. Each additional buyer or supplier a physical aggregator served cost it something real, in shelf space, page count, postage, staff. The analyst Ben Thompson named the structural break this created Aggregation Theory, writing on his site Stratechery that the internet's distinguishing feature is that it drives the marginal cost of distribution and the marginal cost of a transaction toward zero. A search result, a marketplace listing, or a generated answer costs almost nothing to serve to the millionth user beyond what it cost to serve the first.

That near-zero-cost distribution changes who wins the relationship with the buyer, and Thompson's model is specific about the mechanism. An aggregator that has built a direct relationship with end users, at a customer-acquisition cost that falls rather than rises as it grows, does not need to lock up exclusive supplier relationships the way a pre-internet distributor did. It becomes the place users default to, and suppliers follow the users onto the aggregator's platform on the aggregator's terms. Thompson calls this commoditizing suppliers: a retailer on a marketplace, or a listing in a search index, competes on the aggregator's ranking and pricing logic rather than on a direct relationship with its own customer, because the aggregator, not the supplier, now owns that relationship. Thompson frames this as a genuine reversal of the pre-internet order, in which distributors competed for exclusive supplier deals and treated the end customer as an afterthought.

This is the lens the rest of this study applies. Every aggregator that follows, Google's search results, Amazon's and India's marketplaces, and the AI answer engine now emerging, is doing a version of what Boucicaut, Sears, and Donnelley did: standing between buyer and seller, capturing the buyer's default attention, and charging the seller for access to it. What changes is the cost of scaling that chokepoint. A century ago, growing an aggregator's reach meant printing more catalogs or opening more stores. Today it means adding servers, and the businesses Thompson groups among the internet's canonical aggregators reached a scale and a speed of concentration no department store chain or catalog publisher ever approached.

The auction: how Google turned a search box into the chokepoint

Search inherited the discovery chokepoint from the directory, and it did so by inventing a new way to sell the same real estate: an auction, bid in real time, for the buyer's attention at the exact moment of intent.

The marketplace: Amazon, and India's parallel directories

A marketplace collapses the aggregation one step further than a search results page does. It does not just rank a seller among competitors, it hosts the transaction itself, so discovery and the sale happen inside the same owned system, and the toll can be collected on both. Amazon's rise illustrates the shift plainly: multiple survey series tracking where US shoppers begin a product search show Google and Amazon roughly swapping positions between 2015 and 2018, and subsequent surveys through the early 2020s, run by different firms with different methodologies, have put Amazon's share of that starting point anywhere from about half to two-thirds of US shoppers, Jungle Scout's Q2 2023 tracking recording 57 percent, for instance. The range reflects real methodological differences more than a single settled number, but every version of the data agrees on the direction: for a large share of American product searches, Amazon, not a search engine, is now the first stop.

The reckoning: antitrust catches up with the aggregators

By the middle of the 2020s, three of the largest jurisdictions in the world had reached broadly the same conclusion about the search and marketplace aggregators this study has traced, arrived at along three different legal paths. The United States pursued its case through courtroom litigation under a century-old antitrust statute. The European Union built a new regulatory category specifically to avoid re-litigating the same fight case by case. India ran its competition law through the administrative-penalty mechanism it uses across industries. What follows is less three separate stories than one finding, reached three times.

Disintermediation by design: ONDC and the Beckn protocol

Every aggregator this study has traced so far shares one design choice: buyers and sellers meet inside a system the aggregator owns and can change the rules of at will. India's Open Network for Digital Commerce was built to test whether that design choice is actually necessary. ONDC is not a marketplace competing with Amazon or Flipkart for the same buyers. It is a protocol, built on the open-source Beckn specification, that lets any compliant buyer-side app discover and transact against any compliant seller's catalog across the network, regardless of which company built either app, so that discovery happens at the network layer rather than inside a single company's catalog. The effort was convened through discussions initiated by India's Department for Promotion of Industry and Internal Trade, with the aim, described consistently across accounts of its founding, of letting small sellers reach buyers without depending on a single dominant platform.

The network's first order was delivered on April 29, 2022, in Bengaluru, following a pilot that had gone live that March. Growth from there was real, if concentrated in specific categories. By ONDC's own reporting, monthly transactions across the network grew from about 1 million in early 2023 to more than 15 million by the end of 2024, and monthly retail purchases specifically grew sixfold in six months, from roughly 600,000 in September 2023 to 3.6 million in March 2024, helped by the arrival of larger participating brands and by user-facing discounts and seller incentive schemes ONDC has documented itself.

That last detail deserves a sober reading rather than a celebratory one. Volume-linked financial incentives paid to network participants were cut by up to 75 percent starting in the first quarter of fiscal year 2025, a deliberate test of whether transaction volume built partly on subsidy can hold once the subsidy is withdrawn. ONDC is the clearest real-world experiment available on whether an open, government-convened protocol can actually disintermediate a discovery chokepoint rather than simply becoming a new one. It has not run long enough, or at a scale close to Amazon's or Flipkart's, to answer that question. It has run long enough to be worth watching as this study turns, in its final sections, to the same question in a different form.

The new chokepoint: discovery as a scarce citation

A ranked search results page, whatever else was wrong with it, gave a seller a position to occupy. Page one, page two, position eleven, all of it measurable, all of it in principle reachable with enough work. A single answer produced by an AI system does not have a page two. It has a paragraph, sometimes a short list of named sources, and beyond that, nothing. A business that would once have ranked twelfth for a given query and still captured some fraction of searchers willing to scroll now either appears in the system's answer or does not appear at all. That binary is the sharpest version yet of the pattern this study has traced: an intermediary aggregating the buyer's attention completely enough that the seller has no fallback position left to occupy.

The shift toward that binary is already visible in how often a search ends without any click at all. SparkToro's tracking of Google search behavior found that in 2024, for every 1,000 US searches, only about 360 produced a click through to the open web, with a similar 374-per-1,000 figure in the European Union, the remainder resolving directly on the results page without the user leaving it. Google's AI Overviews, which the company rolled out broadly across US search results in May 2024, appear to be accelerating that pattern rather than reversing it. Pew Research Center tracked the browsing behavior of 900 US adults across 68,879 individual Google searches in March 2025 and found that when an AI summary appeared on the results page, users clicked through to a traditional result link only 8 percent of the time, against 15 percent when no summary appeared, and clicked a link inside the summary itself in just 1 percent of visits.

Ranking well in classic search, in other words, is no longer the same task as being named in an AI system's answer, and the two are turning out to be only loosely related. An Ahrefs analysis of 15,000 long-tail queries, checked against both Google's rankings and the citations returned by four AI assistants, found that only 12 percent of the URLs an AI assistant cited also appeared anywhere in Google's own top 10 results for the identical query, a figure that ranged from about 8 percent of citations for ChatGPT up to 28.6 percent for Perplexity, the assistant Ahrefs' researchers noted was built to cite sources more directly than the others. A business optimized for a search results page, in this reading, is optimized for a chokepoint that increasingly sits beside the one that actually decides whether it gets named.

Two different resolutions to that new chokepoint are being tested in parallel, and neither is settled. The New York Times sued Microsoft and OpenAI in the Southern District of New York on December 27, 2023, arguing that training language models on its journalism, and reproducing that journalism in generated answers, infringed its copyright rather than qualifying as fair use, a case in which the presiding judge denied most of the defendants' motions to dismiss in 2025, letting the bulk of the claims proceed. Perplexity has pursued a commercial answer alongside the legal one: its Comet Plus program, first announced in July 2024 and launched in a formal subscription form in January 2026, allocates a 42.5 million dollar pool to publishers on roughly an 80 to 20 split favoring the publisher, tied to citations, direct visits, and AI-agent actions attributable to their content, according to the company's own account of the program, while OpenAI has taken a different position, offering publishers flat licensing fees rather than a share of usage-based revenue. Whether payment for being cited becomes standard, stays contested in court, or never arrives at scale is, as of this writing, an open question rather than a settled market.

The open question: super-aggregation, or an open protocol layer

Every aggregator this study has examined, the department store, the catalog, the directory, the search auction, the marketplace, inserted itself once between a buyer and a seller and then held that position for decades. The AI answer engine raises a question none of its predecessors needed to answer, because none of them had the option: does a system that answers a question directly, in a single generated response, represent the final and most complete form of aggregation this history has traced, an intermediary that no longer even needs to show the buyer a list of sellers to choose from? Or does the same technology that makes such a system possible, structured, machine-readable data that software can query directly rather than a results page built for a human eye, open a genuine path around the aggregator, the way ONDC and the Beckn protocol were built to open one around the marketplace?

The case for further concentration is not hard to make from the record this study has assembled. A handful of firms, OpenAI, Google, Perplexity, and a small number of others, are building the systems that decide what gets cited, and the citation data available so far shows those systems drawing repeatedly on a narrow set of sources, Wikipedia, Reddit, YouTube, and major outlets recur across multiple studies of what gets named. Regulators, on the record assembled here, took roughly eight years to move from Google's first EU antitrust fine in 2017 to a finalized US remedies ruling in 2025, and that ruling itself was confined to search rather than the generated-answer products now shaping the next chokepoint. If discovery migrates from a ranked page, which at least exposed its own mechanics to outside scrutiny, to a generated paragraph, which does not, the historical pattern of an intermediary building years of unchallenged rent extraction before any regulator catches up could simply repeat itself in a form harder to audit.

The case against repetition rests on a genuine difference between how a human shopper and a piece of software look for something. A person, by habit and convenience, tends to default to one app or one search box. Software does not need that habit. An agent acting on a buyer's behalf can, in principle, query many structured sources directly, a business's own schema markup, a government-backed protocol like Beckn, a supplier's own machine-readable catalog, rather than defaulting to whichever single platform happened to win the buyer's attention first. ONDC is the only large-scale test of that proposition running today, and its growth, sixfold in six months for retail purchases on the network's own accounting, is real even if it remains smaller than Amazon's or Flipkart's volume and still partly dependent on incentives the network is now trying to wean itself off. If machine-readability turns out to matter more than platform habit once buyers start delegating search to software, the aggregator's oldest advantage, being the place a human defaults to, stops being decisive.

Both readings are consistent with everything documented in this study, and the evidence assembled here does not settle which one the next decade will bear out. What the history does establish, from Boucicaut's shop floor through Sears' catalog, the Yellow Pages' listing fee, Google's auction, and Amazon's and India's marketplace take rates, is that the chokepoint has moved before, and that whoever built it captured the largest rent in the years immediately after building it and before regulators, rivals, or an open protocol reached it. Whether the current moment shortens that interval or repeats it at a new layer is the question this study leaves open, not a forecast it makes.