From Shelf to Feed to Answer: The Three Ages of Retail Discovery

The planogram and the price of position

A retail shelf looks like neutral space. It has not been neutral for decades. By the late 1970s, US grocery chains had worked out that they could charge suppliers directly for the right to be stocked at all, a practice that came to be called a slotting allowance or slotting fee. The Federal Trade Commission's most detailed public study of the practice, a November 2003 staff report built from case studies in five product categories, fresh bread, hot dogs, ice cream and frozen novelties, shelf stable pasta, and shelf stable salad dressing, found manufacturers routinely paying between one million and two million dollars to get a new product into roughly 85 percent of US supermarkets, with fees varying widely by category and, in some cases, exceeding the product's entire first year revenue in a given metropolitan market. The FTC's own word for the practice was blunt: "widespread." A separate accounting of the underlying FTC data puts a single new item's initial slotting fee at roughly $25,000 in a regional store cluster, rising to as much as $250,000 in the most competitive markets.

Position within the shelf, once a product had a slot at all, carried its own separate price. Retailers plan every inch of a store with a planogram, a diagram that assigns each product to an exact shelf and facing, and the shelf at eye level, roughly 36 to 60 inches off the floor for an average adult, has always commanded the premium, the folk wisdom retail buyers still summarize as "eye level is buy level." The empirical case for that folk wisdom is old and specific: a 1994 field study by Xavier Drèze, Stephen Hoch and Mary Purk, published in the Journal of Retailing, measured how repositioning products on real supermarket shelves actually moved sales, and found that shelf position carried real, measurable elasticity independent of the product itself. A manufacturer that wanted the better position paid for it much as it paid for the shelf itself, through display and promotional fees layered on top of the base slotting allowance. The shift this represented was structural, not incidental: a 1990 Harvard Business Review analysis of the practice, titled plainly "The Costly Bargain of Trade Promotion," described a documented shift in retail power from manufacturer to trade over the prior decade, as chains learned they could extract payment for the shelf itself rather than compete only on what was stocked on it.

The catalog: the first remote shelf

Before there was a marketplace search box, there was the catalog, and it deserves to be read as the same kind of surface: a curated, finite list standing between a buyer and every possible seller, at a moment when the nearest full service store might be a full day's travel from a farm. Aaron Montgomery Ward issued what is generally credited as the first general mail order catalog for the public on August 18, 1872, a single sheet listing 163 items with ordering instructions, built explicitly to reach rural customers who had previously depended on the markup and limited stock of the local general store. Richard Sears followed in the 1880s, founding his company in 1886, and by 1893 had partnered with Alvah Roebuck to build the mail order catalog that customers came to call, informally, the "Big Book," and later, officially in 1968, the "Wish Book."

Two pieces of federal infrastructure did for the catalog what highways would later do for the shopping mall: Rural Free Delivery, begun in 1898, and parcel post, begun in 1913, both of which cut the cost of physically reaching a customer who lived nowhere near a store. The catalog was, in that sense, the first remote shelf, and it operated on a logic that would later govern Amazon's search results and TikTok's For You page just as directly: a limited number of pages, positions within them, and a publisher who decided what a shopper actually got shown. The Sears catalog ran for more than a century before the company shut it down on January 25, 1993, citing competition from discount chains, the same forces that were, at that moment, already building the next surface.

Amazon becomes a search engine

The clearest marker of the shelf's move online is not that people started buying on Amazon. It is that they started searching there instead of on Google. In September 2016, a survey of 2,000 US consumers by BloomReach and Survata found 55 percent starting a product search on Amazon, an 11 percentage point jump from the year before, against 28 percent who started on a search engine, a gap that had not existed a few years earlier. The pattern held into the next decade: Jungle Scout's Q2 2023 Consumer Trends Report, surveying 1,000 US adults, found 57 percent of shoppers starting product searches on Amazon versus 42 percent starting on a search engine. Both figures come from self reported panel surveys run by commerce research vendors rather than from an independently audited count, and vendor waves on this exact question have disagreed with each other by double digits in other releases, so the reading is directional rather than precise: Amazon had, by the middle of the last decade, become a genuine rival to the search engine as where a US product search begins, years before generative AI made "search engine" a contested term again.

The algorithmic shelf: relevance, velocity and the sponsored auction

What replaced the planogram was a ranking algorithm, never officially published in full but reverse engineered in detail by the sellers, agencies and researchers who watch it for a living. Amazon's earlier ranking system, known informally as A9, is described as having weighted keyword relevance and short term sales spikes; the system consultants now call A10, in wide use since roughly 2020, is described as weighting sustained conversion rate, longer run sales velocity, click through rate and even traffic a seller drives in from outside Amazon. Because Amazon has never confirmed the formula, every account of A9 or A10 should be read as informed reconstruction, not company disclosure, though the reconstructions converge closely enough, across competing seller tool vendors with no obvious reason to agree, to be treated as a reasonable account of what the algorithm actually rewards.

Layered on top of that organic ranking is a paid one. Amazon launched Sponsored Products in 2012, a pay per click auction letting sellers bid on keywords to appear inside search results, structurally the same transaction as a slotting fee, rendered as a real time bid instead of a flat payment. The scale it reached is now enormous: Amazon's advertising business, anchored by Sponsored Products, generated an estimated $56.2 billion in 2024, up nearly 20 percent year over year, tracking firm Marketplace Pulse has found, making Amazon one of the largest advertising businesses in the world before most shoppers would think to call it one. The effect on the page itself has been direct and measured over time. Marketplace Pulse documented in March 2022 that only four of the first twenty products shown in a typical Amazon search were organic results, the rest sponsored placements or brand ads, meaning a shopper scrolled past the equivalent of three browser windows of paid listings before reaching a fifth unpaid one, against an ad business worth an estimated $31 billion at that point, a figure that would nearly double within two years. A 2024 study by researchers Abhisek Dash, Saptarshi Ghosh, Animesh Mukherjee, Abhijnan Chakraborty and Krishna Gummadi, presented at the AAAI and ACM Conference on AI, Ethics and Society and covering roughly 4,800 searches and 2.6 million results across four countries, found that top sponsored results were, in the majority of cases studied, both costlier and of measurably lower quality than the top organic results they displaced, evidence that the auction was not simply surfacing the best product faster, it was substituting for a ranking a shopper would otherwise have seen. The slotting fee had not disappeared. It had become a bid, and the FTC's own description of the old practice as "widespread" applies just as well to the new one.

The buy box and the private label question

A second, quieter mechanism decides who actually gets the sale once a shopper has found the product page: the Buy Box, the single "Add to Cart" button Amazon shows for a listing that may have dozens of competing sellers behind it. Industry trackers, including the e-commerce agency Tinuiti, commonly put the share of Amazon purchases completed through the Buy Box above 80 percent, with a seller who loses it reporting sales drops of 50 to 80 percent overnight, though Amazon has not published an official figure, and the number should be read as an oft repeated industry estimate rather than an audited statistic. Amazon's own account of what wins the Buy Box points to price, fulfillment speed and seller performance metrics rather than a pure auction, but regulators have not taken that account on faith.

Two proceedings tested it directly. In the European Union, the European Commission opened formal investigations into whether Amazon used non-public data from independent sellers to advantage its own private label products, and separately into how it awarded Buy Box access. On December 20, 2022, the Commission accepted a binding settlement in which Amazon committed, for seven years across most of the European Economic Area, to stop using non-public seller data for its own retail and private label business, to treat all sellers equally in Buy Box eligibility, and to show a second, differently priced or shipped offer when one exists. In the United States, the Wall Street Journal reported in April 2020, based on internal documents and more than 20 former employees of Amazon's private label unit, that staff had used individual sellers' sales data, in violation of Amazon's own stated policy, to identify which products to copy and how to price them; Amazon disputed the characterization, and the report triggered a multi year SEC inquiry into the company's disclosures. And on September 26, 2023, the FTC and seventeen state attorneys general sued Amazon for illegal monopoly maintenance, alleging among other things that Amazon's search results were "stacked" to bury better reviewed competitor products beneath widgets promoting Amazon's own private label lines. The case survived a motion to dismiss in September 2024 and is scheduled for trial in October 2026. None of these three proceedings, taken alone, proves the marketplace favors itself in every case. Taken together, they are the clearest evidence available that the operator of the modern algorithmic shelf faces a structural temptation the operator of a physical shelf never had to resist quite so precisely: it can see exactly which of its tenants' products sell, and it competes with some of them directly.

The agent that narrows the shelf to a few names

The newest surface does not show a shelf at all. It shows an answer, usually three to five named options, sometimes one, generated by an AI system reading the retailer's or the open web's product data on the shopper's behalf. Amazon moved first among the major platforms, launching its generative shopping assistant Rufus on February 1, 2024, trained on Amazon's catalog, customer reviews and outside web content to answer product questions conversationally inside the app. On May 13, 2026, Amazon retired the Rufus name and folded its function into a broader assistant called Alexa for Shopping, available free to any signed in US customer without a Prime subscription or an Echo device required, and, for the first time in an Amazon branded tool, able to complete a purchase on a shopper's behalf, including outside Amazon's own marketplace. By multiple accounts the timing was as defensive as it was ambitious: the change arrived as ChatGPT, Gemini and Perplexity were all shipping their own shopping and checkout features within the same several months.

Google and OpenAI built comparable ideas on more open rails. OpenAI, together with Stripe, launched Instant Checkout inside ChatGPT on September 29, 2025, running on the Agentic Commerce Protocol, an open specification the two companies published under the Apache 2.0 license the same day, intended to standardize how any AI agent hands a scoped, limited payment token to any compliant merchant. Google answered on May 19, 2026 with Universal Cart, an AI shopping hub spanning Search, Gemini, YouTube and Gmail, built on Google's Shopping Graph of more than 60 billion product listings and its own Universal Commerce Protocol, letting a shopper complete a purchase through Google Pay in a few taps without leaving the conversation. Traffic evidence suggests shoppers are following the agents: Adobe Analytics, tracking more than a trillion visits to US retail sites, measured AI referred traffic up 138 percent year over year in May 2026 alone, and up more than thirteenfold since October 2024, with that AI referred traffic now converting 54 percent better than traffic from other sources, a reversal from a year earlier when AI referrals converted worse than everything else.

The agent that buys: a contested frontier

Whether an agent should be allowed to buy, not just recommend, is not yet settled, and the clearest evidence of that is in court rather than in a product announcement. OpenAI's own Instant Checkout, roughly six months after launch, was reported in March 2026 to be losing ground to a redesigned approach that routes purchases through merchant owned apps, such as Instacart, Target and Booking.com, rather than completing the transaction inside ChatGPT itself, a shift several outlets read as evidence that agent mediated checkout had not converted the way OpenAI had hoped. Separately, and more sharply, Amazon sued Perplexity in November 2025, alleging that Perplexity's Comet browser used its AI shopping agent to access Amazon customer accounts while disguising itself as an ordinary browser, in violation of the federal Computer Fraud and Abuse Act and its California state equivalent. A federal district court granted Amazon a preliminary injunction in March 2026 blocking the Comet agent from shopping on Amazon; the Ninth Circuit Court of Appeals later vacated that injunction, finding Amazon unlikely to succeed on the claims as pleaded, and returned the case to the district court for further proceedings. The dispute is not really about one browser. It is the first serious court test of whether the operator of a shelf gets to decide which agents are allowed to shop it on a customer's behalf, the AI era version of the same question the European Commission asked about Buy Box access in 2022, and as of this writing it remains open.