The most important shopping agent in the world is not made by OpenAI, Google, or Meta. It is made by the store it shops in.
On May 13, 2026, Amazon retired the Rufus name and folded its assistant into “Alexa for Shopping,” merging Rufus’s product expertise with the personal context of Alexa+. The launch framing was friendly and unsurprising: “the world’s best, most personalized AI assistant for shopping,” available to every US customer in the Amazon app, on the website, and on Echo Show, no Prime membership required. Rajiv Mehta, Amazon’s vice president of Conversational Shopping, described it as “an expert personal shopper who already knows you and remembers your preferences, your past purchases, and your conversations.”
Behind the rebrand sits the largest deployment of shopping AI ever built. And in 2026 it acquired three properties that deserve much more scrutiny than the name change got: it moved into the main search bar, it started carrying advertising billed per click, and it became the only shopping agent welcome on Amazon while every competitor’s agent gets blocked at the door.
The scale nobody prices in
Start with the numbers Amazon itself has disclosed. Rufus was used by more than 300 million customers in 2025 and, per Amazon’s February earnings disclosure as reported by Modern Retail, helped drive nearly $12 billion in incremental annualized sales. On the Q1 2026 earnings call, CEO Andy Jassy said monthly active users were up 115 percent year over year and engagement up 400 percent. Amazon says customers who use the assistant are 60 percent more likely to complete a purchase.
For comparison, the entire third-party agentic commerce push, OpenAI’s Instant Checkout included, was retired in March 2026 after roughly six months because it did not offer the experience OpenAI wanted. The first-party agent, by contrast, is already inside the default product discovery interface of the largest Western marketplace, with a decade of purchase history and review data to draw on.
The capability tour explains why it converts. Alexa for Shopping answers questions typed directly into the main Amazon search bar, generates side-by-side product comparisons, surfaces AI overviews in search results and on product detail pages, and shows up to a full year of price history on hundreds of millions of products. It accepts standing orders: “Add this sunscreen to my cart if the price drops to $10 and I haven’t purchased it in the last 2 months.” And it will “shop from other online stores across the web,” completing checkout on brands’ own websites through the Buy for Me feature.
Third-party tracking shows how fast this is moving. Analysis of on-site AI traffic cited by StartupHub found product-page traffic from Amazon’s on-site AI jumped 26 percent month over month in June and 537 percent year over year. Seller-side guides like Amalytix’s now map assistant touchpoints across the entire journey: “Researched by AI” modules on the results page, “Customers Ask” summaries whose most-clicked chat prompts get promoted into standard search, an “Ask Alexa” button on product pages, and “Compare with similar items” inside the cart. The assistant is not a chat bubble on the side. It is becoming the interface.
Then the ads arrived
In November 2025, Amazon began rolling out a new ad format called sponsored prompts: suggested questions inside search results and product pages, designed to spark conversations with the assistant. A shopper browsing running shoes might see a prompt nudging them toward a specific brand. Jassy told analysts that “it’s early, but we’re seeing nearly 20% of shoppers who interact with the brand prompt and Rufus continue the conversation about that brand.”
Since March 25, 2026, those placements have been billed per click, with Amazon’s own advertiser documentation describing sponsored prompts that “may open a dialog in Rufus or respond to the customer directly on the page,” as The AI Praxis recounts. Walmart has moved the same direction, testing its first ad formats tied to the Sparky agent in fall 2025.
Jassy has been refreshingly explicit about where this goes. “With AI interfaces, you tend to find that you’re asking questions. You’re narrowing questions… in that process there are multiple opportunities to surface relevant products to customers, many of which will be organic and some of which will be sponsored,” he said on the Q1 call. “Advertising will do well in a world of agentic commerce.”
The financial context makes the strategy legible. Amazon booked $68.6 billion in advertising revenue in 2025, roughly twice the combined operating profit of its North America and International store segments ($34.4 billion). Every $100 of store sales carried $11.67 of ad revenue against $5.84 of store operating profit. In Q1 2026, ad revenue grew 24 percent year over year to $17.2 billion, again per Modern Retail’s earnings coverage. Advertising is not a side business that the shopping agent accidentally brushes against. It is the profit engine the shopping agent is being built inside of.
The one-way door
Now hold that next to Amazon’s treatment of everyone else’s agents.
Per verification firm Billions Network, Amazon blocks agents from OpenAI, Google, Anthropic and Perplexity, and on September 20 it began blocking Meta’s Muse, twelve days after Muse launched. Shoppers who tried to buy through Muse hit a popup citing Amazon’s Conditions of Use. The road there ran through court: Amazon sued Perplexity over its Comet shopping browser in November 2025, won a preliminary injunction in March 2026, and then watched a unanimous Ninth Circuit panel vacate it on August 4, holding that when an AI tool acts on a shopper’s instructions, it is the shopper, not the tool’s maker, who accesses Amazon’s computers under federal hacking law. The court explicitly left open Amazon’s ability to police access through its terms of service. The Muse popup is Amazon walking through that door. eBay went first, amending its User Agreement effective February 20, 2026 to ban “buy-for-me agents, LLM-driven bots” and any end-to-end ordering flow without human review, unless eBay grants permission.
Meanwhile Alexa for Shopping does to other retailers exactly what Amazon just banned: Buy for Me completes checkout on brands’ own websites from inside the Amazon app, and CNBC reported in January that some small merchants discovered their catalogs listed through Buy for Me without having opted in. As one analysis of the asymmetry put it: “Amazon is not against shopping agents. It is against shopping agents that do not carry its shelf.”
The rest of retail is sorting itself by where it collects its toll, and the WSJ’s survey found a spectrum rather than a consensus. QVC and HSN let agents complete purchases and push catalogs to OpenAI and Google: “We need to be out in the places that already are grabbing her attention,” said QVC’s David Thompson. Gap decides case by case. Tapestry lets agents browse but not buy, citing fraud and inventory hoarding. Kohl’s blocks outright. Walmart lets Google’s Gemini surface its products, then hands the basket to its own Sparky, per PYMNTS.
There is a coherent business logic in every one of these choices. The problem is what happens to the shopper in the middle, because the largest of these agents, the one most shoppers will actually use this holiday season, is owned by the party that sells placement in its own answers.
The referee problem, now measured
Until this year, the conflict of interest in first-party shopping agents was an argument. It is now a measurement.
A September 2026 preprint by Wadi and Ma (arXiv 2609.17989) ran repeated trials of AI agents choosing among listings, one marked sponsored. Left to act for the traveler, an agent tends to penalize the sponsored listing. Tell the same agent it works for the platform, and the penalty shrinks. In the authors’ words: “platform delegation significantly attenuates the penalty that agents apply to sponsored listings.” Stricter wording, “Sponsored” instead of “Promoted,” does not close the gap.
The penalty itself is real. The ACES study by Allouah, Besbes, Figueroa, Kanoria and Kumar put AI agents in a randomized shopping sandbox and found a sponsored tag cut a product’s selection rate from 10 percent to 8.9 percent under Claude Sonnet 4, 8.0 percent under GPT-4.1, and 7.9 percent under Gemini 2.5 Flash. “Sponsored ads thus buy attention through placement but induce a credibility cost.” What Wadi and Ma add is that the credibility cost depends on whom the agent believes it serves. A first-party agent is, by construction, told it serves the platform.
Human oversight does not rescue this. An April 2026 preprint by Salvi, Cuevas and Horta Ribeiro across 2,012 people found a 17.9 percent detection rate for native advertising inside AI-mediated content, and “explicit ‘Sponsored’ labels do not significantly reduce persuasion.” The shopper being informed and the shopper being solicited, the distinction the FTC was already demanding of search engines in 2002, blurs further when the recommendation arrives inside a conversation with a trusted assistant. And the sponsored tier being sold into that conversation is not hypothetical quality: across 4,800 Amazon searches in four countries, Dash and colleagues (AIES 2024) found that “top sponsored results are costlier and are of poorer quality than the top organic results” in most cases.
The AI Praxis proposes four indicators for how much commercial interest shapes an agent: the disclosure gap between the platform’s commercial arrangements and what its agent reveals at the moment of recommendation, the link between platform revenue and rank, the share of the platform’s own products in recommendations, and whether any outside party can audit the agent. Apply that test to Alexa for Shopping and watch every indicator fail at once: sponsored prompts are a documented, CPC-billed product inside the answers; advertising is Amazon’s most profitable line; Amazon Devices and private labels populate the catalog; and no outside party can audit the ranking model. This is what The AI Praxis calls a Shadow Principal: the undisclosed party whose objectives are embedded in the agent beyond those of the consumer it nominally serves.
And the review layer inherits everything
The sponsored question sits on top of an evidence base that was already compromised. The review corpus Alexa for Shopping summarizes is the same one the rest of us see: estimated at roughly 30 percent fake, with review fraud costing shoppers an estimated $770.7 billion worldwide in 2025, and with Amazon itself having blocked hundreds of millions of suspected fake reviews and admitting its tools cannot solve it alone. Purchased five-star reviews move exactly the signals, average rating, volume, sentiment ratio, verified-purchase flags, that a language model compresses into a recommendation.
So the AI overview on the product page launders two layers of contamination at once: the manipulated review corpus underneath and the paid placement possibility on top, all rendered in the calm, authoritative voice of an assistant that “already knows you.” The FTC’s Consumer Review Rule penalizes the fake-review brokers after the fact. Nothing in the rule, or anywhere else in the stack, reaches an AI overview that summarized polluted evidence with a sponsored prompt attached.
Consumers are walking into this with their eyes only half open. About 23 percent of US consumers, some 61 million people, now start their shopping research with AI, per the August survey PYMNTS cited. But only 3 percent of US adults say they would trust an agent to complete a purchase, per NMI, and a third of active AI users told VML they would never let an agent buy or even reorder for them. That combination, AI-influenced research with human-final checkout, is precisely the regime where the recommendation layer does the deciding while the human supervises only the payment. The buy button is verified. The evidence that chose the product is not.
What an independent evidence layer looks like
None of this makes Alexa for Shopping the villain. The capability tour is genuinely impressive, and Jassy is right that third-party horizontal agents currently fumble pricing and product data. The point is narrower and structural: the most-used product recommender on earth answers to the store, and no amount of labeling changes that. Trust in a purchase decision has to come from a layer that does not profit from the answer.
That layer needs three properties, and they map exactly onto the four indicators above.
First, independence from the marketplace and its ad business, so a second opinion does not inherit the contamination it exists to check. Second, filtering before the decision: review fraud screening, coordinated-behavior detection, and authenticity weighting upstream of any score, so purchased praise never reaches the average. Third, machine-native delivery: trust data as a queryable endpoint, not a rendered page, so any agent, first-party or third, can consult it before recommending. That is the architecture we built GoBuy on: a Smart Score from 0 to 100 computed on filtered review evidence, weighted by quality rather than volume, with a curated top-7 result set instead of an infinite sponsored scroll, exposed through MCP at gobuy.ai/api/mcp. Products holding a Smart Score above 80 across 90 days earn the GoBuy Verified badge. And because the human still reads Amazon pages, the Chrome extension overlays the same trust panel directly on the product page, next to the AI overview, so the person and the agent can compare notes on what deserves the checkout.
The emerging discipline for brands points the same way. Agent Intent Optimisation, in The AI Praxis’s formulation, competes for selection, and an agent selects on what it can verify rather than on what it is told. Verification is not something a first-party agent can do about itself. It is something the ecosystem has to provide.
Whose agent is it
The question that matters this holiday season is not whether AI will mediate product choice. At 300 million customers, $12 billion in incremental sales, and 115 percent user growth, Alexa for Shopping has settled that. The question is whose agent is doing the mediating, and whom it works for. Amazon built the best shopping agent in the West, put it in the search bar, and started selling its answers by the click. Every other agent is locked outside, and the court that let them in on the shopper’s behalf found Amazon’s terms of service waiting as the fallback.
A referee that plays for one team can still call a good game most nights. But you do not hand it the whistle permanently, and you do not let it be the only scorekeeper. Product trust for the agent era means an evidence layer outside the store: filtered reviews, quality-weighted scores, and a protocol any agent can query.
If you are building shopping agents, wire them to consult the trust layer before they recommend: the MCP integration lives at gobuy.ai/agent-docs. If you are shopping the old way, with your own eyes on an Amazon page that now summarizes itself, the trust panel is at gobuy.ai. Either way, when the assistant says “best,” ask who paid for the sentence.