On November 4, 2026, Amazon opened a door it has spent the rest of the year bolting shut on everyone else. The company launched the Shop Direct Merchant Workspace, a free, self-service portal where online retailers that do not sell on Amazon can register, submit a product feed, and have their products surface in Amazon search results and inside Alexa for Shopping, Amazon’s agentic shopping assistant. No seller account. No referral fee. No warehouse. Three steps: create an account, verify your business website, share your catalog.

The scale claims in Amazon’s own announcement are enormous. “Since introducing Shop Direct last year, tens of millions of customers have discovered hundreds of millions of unique products from other stores, and Shop Direct products are appearing in hundreds of millions of searches monthly.” Amanda Doerr, vice president of Core Shopping at Amazon, framed the pitch in one sentence: “Customers are increasingly trusting Amazon and Alexa for Shopping as the place to start any shopping journey, whether what they’re looking for is in our store or somewhere else on the web.”

Read that sentence again, because it contains the entire strategy. Amazon is positioning itself as the starting point for shopping missions that end somewhere else entirely, on storefronts Amazon does not operate, against products Amazon has never verified, recommended by an assistant whose every other input Amazon controls. The company’s stated goal is explicit: “when customers search for specific products or brands that aren’t sold in our store, we want to provide them easy ways to find and buy them on Amazon and our Alexa for Shopping agent.”

This is the largest product index on earth voluntarily ingesting inventory it does not stand behind. And the announcement that opened the floodgate says nothing, not one word, about what verifies any of it.

What actually shipped

The Merchant Workspace is the self-service front end to a program that has existed since 2025 but was previously gated behind feed partners and account managers. Per PPC Land’s documentation of the launch, the mechanics matter more than the marketing:

  • The feed format is Google’s. Amazon’s developer documentation specifies that listings are submitted through the Selling Partner API using a feed type named GOOGLE_PRODUCT_DATA_FEED, the same product data format retailers already maintain for Google Merchant Center. Variants are grouped by a shared item_group_id. A catalog built for Google’s shopping surfaces is now, with minimal rework, the input to Amazon’s assistant.
  • Four feed providers, one new. Merchants connect through Feedonomics, Salsify, CEDCommerce, or VersaFeed, the last added since the program opened to third-party feeds in March 2026, or integrate directly via the SP-API.
  • Eligibility is undefined. The Workspace serves “any eligible merchant,” in Amazon’s words, but the announcement names no minimum catalog size, no category restrictions, no revenue threshold, no excluded verticals. The only visible gate is website verification, and how that verification works is not explained.
  • The merchant keeps everything. Inventory, pricing, fulfillment, shipping, returns, exchanges, and customer service stay with the merchant. Amazon becomes a discovery and, through Buy for Me, a checkout layer for businesses it has no commercial relationship with.

Two purchase paths exist. The shopper taps “Shop Direct,” gets a notification that they are leaving Amazon, and finishes on the merchant’s site. Or, when eligible, they tap “Buy for Me”: the customer confirms delivery address, taxes, shipping fees, and payment method on Amazon’s interface, and per Amazon’s announcement, “Amazon’s AI then securely completes the purchase from the merchant’s website.”

That second path is the one to watch. It is a fully agentic purchase of an unverified product from an unaffiliated merchant, executed by AI on a website Amazon does not control, for a customer who chose the listing because Amazon’s search surfaced it.

The numbers Amazon gave, and the ones it withheld

Amazon’s announcement quantifies reach and nothing else. “Tens of millions” of customers is not defined; PPC Land notes it is unclear whether a “discovery” means a search impression, a page view, or a click. “Appearing in hundreds of millions of searches monthly” measures exposure, not traffic delivered. The March 2026 baseline was more than 100 million products from more than 400,000 merchants; “hundreds of millions of unique products” implies at least a doubling, but Amazon published no current counts.

What merchants actually need to judge the program, referral clicks, click-through rates, Buy for Me order volumes, conversion rates, average order values, is entirely absent. Amazon describes its own store in precise terms, “300 million items available with fast, free Prime delivery, across more than 35 categories,” and describes Shop Direct in atmospheric ones. Buy for Me’s growth curve is real, from roughly 65,000 products at inception to more than half a million by November 2025, per PPC Land’s earlier reporting. But a merchant signing up through the new Workspace is flying on faith.

The asymmetry extends to the program’s history. Earlier in 2026, independent brands reported unauthorized Shop Direct listings scraped from their websites, products appearing in Amazon search that the brands never submitted. The Workspace’s website verification step is, in part, a response to a program that previously ingested catalogs somewhat faster than it checked them.

The only agent welcome everywhere is Amazon’s

The context for this launch is a year of Amazon doing the opposite to everyone else’s agents. As PPC Land summarized it: Amazon has restricted outside agents on its own store, including litigation against Perplexity and the blocking of Meta’s Muse shopping assistant in September, while independent brands were finding their catalogs scraped into Shop Direct without consent. Amazon wants to be the agent that shops other people’s stores while keeping other agents out of its own.

That position is commercially rational and it is already paying off. Every Shop Direct listing is demand captured at the search layer before a shopper can wander to Google, ChatGPT, or a rival assistant. Every Buy for Me order is a transaction Amazon mediates without carrying inventory cost. And every external merchant that joins strengthens the answer Alexa for Shopping gives when a customer asks for something Amazon does not stock, which is precisely the query where an assistant would otherwise have to admit defeat or send the shopper elsewhere.

But it also means the trust question concentrates. When Alexa for Shopping recommends a product from Amazon’s own store, the recommendation leans on Amazon’s review corpus, its purchase data, and its seller performance history, however imperfect those signals are. When Alexa recommends a Shop Direct listing, none of that machinery applies. The listing arrived through a feed pipe. The feed carries structured data: identifiers, prices, availability, shipping and return policy fields. Ratings can ride along in the same formats as optional attributes. Nothing in the pipeline verifies any of it, and Amazon’s announcement does not say how, or whether, Shop Direct product pages display review data at all. PPC Land’s coverage lists the open questions plainly: how Shop Direct listings rank against paying sellers, whether they can carry ads, and how they are measured were all left unanswered.

The feed format Amazon chose is optimized for freshness and structure, not truth. A GOOGLE_PRODUCT_DATA_FEED with a polished title, a competitive price, and a flattering rating attribute costs essentially nothing to produce. Against a ranking system, an unverifiable signal is indistinguishable from a verified one until someone checks.

Meanwhile, the rest of the web is closing its doors

The same week Amazon opened its index, an audit quantified how hostile the open web remains to the agents that might shop it. As PPC Land reported, conversion agency Blend Commerce ran an automated shopping agent against 200 UK and US retailers across 20 sectors between September 8 and 15, scoring each on 17 checks: reachable, discoverable, understandable, transactable. The results:

  • Average readiness score: 65 out of 100. Baby and nursery retailers led at 88; fashion came last at 49.
  • 29 of 200 retailers refused the agent outright, on two attempts, including Home Depot, Best Buy, Macy’s, Tesco, Argos, and Next. Saks Fifth Avenue, Bergdorf Goodman, Anthropologie, and JD Williams refused even requests for their robots.txt files and scored zero.
  • The blocker is software, not policy. Blend could read robots.txt for 24 of the refusing stores, and all 24 allowed the ChatGPT, Perplexity, and Claude crawlers. Blend’s co-founder and CEO Adam Pearce put it flatly: “Most of these retailers have likely told the AI companies their agents are welcome.”
  • Where agents got in, the data was skeletal. Product data was readable at 171 stores, but only 18 percent included a GTIN, 22 percent shipping details, and 20 percent a returns policy. Of 38 retailers with a live agent checkout endpoint, 37 ran on Shopify.

The identity layer is lagging just as badly. An Okta analysis published the same week found that only 22.4 percent of organizations building AI agents give them unique, tightly scoped, time-limited credentials, and Auth0 CPO Gareth Davies described the inversion plainly: “For years, security teams focused on keeping bots out. Now, we have to let the good ones in.”

Put the two weeks side by side and the pattern is unmistakable. Amazon’s answer to agents is to welcome them as merchants and bar them as competitors. Retailers’ answer is to welcome them in policy documents and refuse them at the WAF. The index layer is being rebuilt around agents even as the evidence layer, the part that would tell an agent whether a product is actually good, remains nobody’s project.

The demand arriving through these doors

None of this is happening ahead of demand. Adobe’s September 28 holiday forecast, resurfaced this month, projects AI traffic to US retail sites to rise 130 percent year over year between November 1 and December 31, following a 693.4 percent increase in the same period last year, with the biggest single jump on Thanksgiving at 159 percent, Black Friday at 95 percent, and Cyber Monday at 88 percent. The forecast draws on more than one trillion visits across 100 million SKUs. Google, for its part, opened Universal Cart in the US on October 15, a cross-retailer cart in Search and the Gemini app built on more than 60 billion continually refreshed listings, with checkout through Google Pay at participating merchants.

The timing of Amazon’s Workspace launch is not subtle. The holiday season is the first in which AI-mediated discovery is a measurable, double-digit-percentage traffic channel. Amazon is stocking its index with external selection before the surge, so that the assistant a shopper asks first has an answer for every query, including the ones Amazon’s warehouse cannot fill.

What an open index is missing

Here is the structural problem the industry keeps stepping over. Every door that opened this month is a data door. Amazon’s Workspace ingests product feeds. Google’s Shopping Graph refreshes 60 billion listings. Meta, in the week’s strangest index story, was found to have more than six million AI-generated Facebook Answers pages indexed and ranking in Google, cited inside AI Overviews and Gemini, which the SEO consultant Glenn Gabe labeled scaled content abuse under Google’s own spam policies. Indexes everywhere are filling with machine-generated, unverified content, and the answer engines are citing it.

A feed is a claim. Price is a claim. Availability is a claim. A rating attribute is a claim. The entire agentic commerce stack being assembled at breakneck speed, payment protocols, agent identity, spend authorization, is designed to verify one party’s claims to another party’s money. Nothing in it verifies the product. When Visa benchmarked 22 shopping models from 12 labs this month, agents failed 42.8 percent of runs, and in 54 percent of failed runs they still bought something. Agents under uncertainty reach for the most legible signal available, and the most legible signal in every index being opened this month is feed-supplied and unverified.

That is the gap an evidence layer fills, and it has to sit outside the index doing the ranking, because an index operator’s incentives and a shopper’s incentives are not aligned. Amazon will not verify external merchants against its own review corpus; the whole point of Shop Direct is scale without that burden. Google will not verify merchant ratings across 60 billion listings; its graph is built on the feeds merchants supply. The verification has to come from somewhere whose only product is the verification.

That is what GoBuy builds. Before any buy decision, human or agentic:

  • A Smart Score from 0 to 100 computed from review quality, not review quantity, so a rating attribute inflated by a thousand purchased five-star reviews scores lower than a smaller corpus of substantive, authentic ones.
  • Fake review filtering that suppresses incentivized, coordinated, and AI-generated reviews before they reach the score, so the legible signal an agent grabs is one that survives manipulation.
  • A shortlist of the top 7 products per category instead of hundreds of millions of listings, which maps to what agents actually need: a defensible answer, not an infinite shelf.
  • The GoBuy Verified badge, reserved for products holding a Smart Score above 80 across 90 days, a durable input an agent can check before it commits a customer’s money to a stranger’s storefront.

All of it is exposed over MCP at gobuy.ai/api/mcp, so whether the agent lives in Alexa, Gemini, ChatGPT, or a retailer’s own stack, it can consult a trust signal that did not arrive through the same feed pipe as the listing it is evaluating.

The question every open index now has to answer

Amazon’s Workspace launch will succeed on its own terms. Free distribution to tens of millions of high-intent shoppers is an offer few merchants refuse, and the Google-format feed removes the last technical excuse. Expect the external catalog to double again, expect Buy for Me to spread, and expect Google and OpenAI to widen their own intake pipes in response.

But the definition of trust in commerce is quietly changing underneath the launch. For twenty years, “trusted marketplace” meant a corpus of reviews accumulated inside a walled garden. The index era breaks that model: the garden now contains products the garden’s owner has never seen, submitted through a feed format designed for advertising distribution, ranked by systems that treat a verified attribute and an unverifiable one as the same field. The question is no longer whether an index is comprehensive. It is whether anything behind the listing is true.

Amazon answered the comprehensiveness question on November 4. Nobody has answered the truth question. If you are building a shopping agent, an agentic checkout flow, or a merchant integration into any of these programs, wire verification in before the transaction: read the agent docs, connect to the MCP server at gobuy.ai/api/mcp, and make sure the answer your agent gives is backed by evidence that did not come from the listing itself.