Two different Amazons made news this week, and reading them together tells you where the trust layer of e-commerce is going.

The first Amazon was at unBoxed, the company’s annual advertising conference in San Francisco, where the keynote spent an hour on AI tools for buying media. The second Amazon was trending on Threads, where users discovered the blunt personal profiles its About You feature had inferred about them, including one shopper’s revelation that Amazon believed her “husband has skinny legs.”

One Amazon sells things. The other Amazon decides what you see, what the assistant says, and what the crowd appears to think. This week, those two businesses merged in a way that got less attention than either story alone: the launch that brand and agency buyers at unBoxed were most excited about was not an ad tool at all. It was a review machine.

The Breakout Hit Was Not an Ad Product

Kiri Masters’ report from unBoxed for The Drum is worth reading in full, because the enthusiasm gap she documents is the real story. Amazon announced plenty: a combined buying platform called Amazon Ads Agent, AI-run Full-Funnel Campaigns that the company says lifted long-term return on ad spend by 67 percent for beta advertisers, analytics you can query in plain English. Danny Hoffman, vice-president of global commerce strategy at Flywheel, noted the keynote was “an hour long, and at minute 58 was when they first said AMC,” referring to Amazon Marketing Cloud. Very little Prime Video, no live sports. In his words: “This is like an SMB pitch for 50 minutes of a keynote.”

Advertisers rated the AI buying tools coolly. One agency source called Full-Funnel Campaigns “very mid to low,” built for basic brands. But nearly everyone Masters spoke to singled out Review Requests, announced almost in passing. The tool invites verified buyers to rate or review a product from busy pages such as the Amazon homepage, with a single tap. Amazon says products in early testing received around three times as many ratings and reviews each week. It launches in open beta in the United States in late October.

Ross Walker, director of retail media at Acadia, gave it ten out of ten: “Another lever, a legal lever for brands to pull to get more reviews, has been in need for so long.” Summer Jubelirer, senior Amazon media manager at OLLY, went further: she will make space for it in her 2027 budget, “for new product launches or products that just fall below 4.2 stars.”

Masters supplies the historical framing: reviews are the oldest competitive advantage Amazon has, the thing shoppers in other retailers’ aisles used to check on their phones. Amazon has run Vine and a string of review programs, “but until now there was nothing a brand could put media budget behind.” Now there is. The review layer, the last surface on Amazon that read as organic, is on the media plan.

The 4.2-Star Confession

Read Jubelirer’s sentence again, slowly, because it is the most honest pricing document the review economy produced this year. The use case for Review Requests is “products that just fall below 4.2 stars.”

That is not a product-quality program. It is a rating-repair program. The threshold is not arbitrary: below roughly 4.2 stars, marketplace wisdom holds, conversion falls off a cliff, and a product that cannot cross the line organically has exactly one path back: more ratings. The tool does not improve the product. It improves the product’s arithmetic, by flooding the denominator with fresh, easy, positive-leaning signal until the average drifts over the line.

If that mechanism sounds familiar, it should. Three weeks ago this column walked through Singapore’s record fake-review bust, where the Reputifly dashboard included a rating calculator that computed exactly how many five-star posts a business needed to reach a target Google average. Review Requests is the legal, platform-sanctioned version of the same arithmetic. The inputs are verified buyers rather than rented gig workers, so the provenance is clean. But the optimization target is identical: a number, moved by volume, without the product changing at all.

The legitimate version and the fraudulent version even share a business model. Both exist because the star average is the single most load-bearing trust signal in commerce, and both monetize the gap between what a product is and what its rating says. One sells the gap to gig workers at five dollars a post. The other sells it to brands out of a media budget.

What a One-Tap Rating Actually Measures

Amazon’s framing of Review Requests is a supply problem solved: authentic reviews are scarce, especially for new products, and the company is using its own traffic to fix the scarcity. The early number, three times weekly ratings and reviews, will be quoted everywhere as success.

But ask what a single-tap rating from the homepage actually measures, and the picture changes.

It measures ease, not experience. A one-tap rating is given at the moment of browsing, not the moment of use. The buyer is reacting to the fact of ownership and a prompt, not to a product they have lived with. The information content per rating, how it performs, what broke, what surprised, approaches zero. You can triple the volume of a corpus while adding almost nothing to what a careful shopper, or a careful machine, can learn from it.

It selects for the casual middle. Solicited single-tap ratings are answered by whoever happens to be on the homepage and willing to tap. Dissatisfied customers with a story to tell tend to write reviews anyway; delighted customers tap; the mildly disappointed, the ones whose complaints carry the most decision value, neither tap nor write. Prompt-based collection reshapes the distribution before a single rating is cast.

It launders volume into credibility. A product with 40,000 ratings reads as more trustworthy than one with 400, even when most of the 40,000 carry no information. And when those ratings carry a “verified buyer” flag, the weakest signal gets dressed in the strongest provenance. The badge certifies that a purchase happened. It says nothing about whether the rating reflects anything.

None of this is illegal, and that is the point Walker was making with “a legal lever.” Amazon simultaneously blocks hundreds of millions of suspected fake reviews, this column covered the company’s own enforcement numbers last month, and now sells brands a compliant way to generate rating volume. Enforcement against the black market, productization of the white market. Both treat review quantity as the metric to manage. Neither touches quality, because quantity is what the machine reads first and quality is what the platform cannot sell.

The Narrative Layer Went on Sale in the Same Keynote

Review Requests was not even the only trust surface Amazon monetized at unBoxed. The same event shipped Branded Conversations, a Sponsored Brands unit inside Alexa for Shopping, the assistant formerly known as Rufus, since combined with Alexa+ and renamed. Brands supply product information, what sets their products apart, which ones work together, and the assistant draws on that supplied material when answering shoppers’ questions in conversation.

The framing from Muthu Muthukrishnan, vice-president of sponsored ads at Amazon Ads: “With Branded Conversations, we are giving brands the ability to bring their product expertise and storytelling into those conversations, at the moment of decision.”

Set the two launches side by side and the architecture is unmistakable. A purchase decision on Amazon now runs on three evidentiary layers: what the crowd says (reviews), what the assistant says (the conversational layer), and what the shopper wants (personalization). At this unBoxed, layer one gained a volume-stimulation product and layer two became directly authorable by advertisers. Two of the three inputs to trust are now payable by the seller.

There is context worth holding. Sponsored Prompts, which first surfaced at last year’s unBoxed as a mention of “sponsored questions” in Rufus, was the first ad unit inside any retailer’s AI shopping assistant, and Amazon is already on its second format. An agency executive who tested it told The Drum it casts “a wide net that’s hard to aim,” with reporting that amounts to a list of the prompts an ad appeared on. Branded Conversations, in that light, reads as much as a fix for the first format’s targeting and brand-safety problems as a new invention. The direction of travel is the story: the conversational layer is inventory, and inventory gets productized.

The Mirror Went Viral

The third layer, what the shopper wants, is the one Amazon made visible this week. About You, live since May, is an editable dossier of the preferences Amazon has inferred from your purchase history, your searches, your saved Lists, the product reviews you have authored, and your conversations with Alexa for Shopping. You can ask the assistant, directly: “What do you know about me?”

This week the answers went viral on Threads, as Business Insider documented: the shopper told she “has flat buttocks” (“I’m literally speechless”), the husband with skinny legs, the user who “Uses phone in the shower to read Kindle books,” the profile that listed “does not trust people easily” under Interests and Hobbies. Users described the feature as “cosplaying a judgmental aunt.” Amazon’s own positioning is gentler: the company says it aims to personalize “nearly every shopping experience across Amazon” for “hundreds of millions of customers.”

Credit where due: About You is the genuinely good trust feature in this whole story. It shows the shopper the inference, lets them correct or delete it, and is honest about which signals feed it. That is transparency and control, done properly, for the demand side.

But notice the asymmetry. The buyer’s dossier is visible, editable, and free. The product’s dossier, the review corpus, is now a media product, and the assistant’s answers are partially authored by advertisers. The human in the transaction gets to see what Amazon thinks of them. They still cannot see what advertisers paid to place in front of the decision, or which review volume was stimulated by a brand with a sub-4.2 problem.

Everyone Is Retreating to Onsite

The monetization pressure behind all this is structural, not incidental. Bain & Company and eMarketer surveyed 61 US retail media network leaders in July about which ad formats will drive the most revenue growth over the next 12 months. Last year the top three spanned onsite, in-store, and offsite. This year, per The Drum’s writeup, all three are onsite: search and sponsored product ads at 43 percent, native AI ads at 39 percent, display at 36 percent, with offsite formats including social, CTV, and influencers last at 20 percent. eMarketer’s Sarah Marzano called it “a retreat or retrenching toward what we [in retail media] do best.” Masters’ own read of analyst estimates is that around 80 percent of Amazon’s ad spend already sits in onsite, performance-oriented ads.

When the growth engine is onsite inventory, every surface that influences an onsite decision becomes candidate inventory. Search results already were. The AI assistant already is, twice over. Reviews were the last holdout, and Review Requests is the breach. A retail media network that needs onsite growth will keep finding onsite surfaces, and the ones left are exactly the ones shoppers still trust precisely because they were never for sale.

The Reader of This Corpus Is Increasingly a Machine

There is one more actor in this story, and it is the one that changes the stakes: the reader. For thirty years the review corpus had one consumer, a human skimming stars. Now it has two. HUMAN Security measured 84 percent of Meta Muse agent requests hitting product and search pages. Outside agents, Claude, ChatGPT, Perplexity, read the review corpus as structured evidence, and Amazon’s own Alexa for Shopping summarizes it back to shoppers as conversation. When reviews were for humans, a stimulated rating inflated an impression. When reviews are machine-read, a stimulated rating becomes a poisoned feature in someone else’s inference: a weight, with a verified-buyer provenance flag attached, that the model has no way to discount.

This is why volume products are more dangerous in 2026 than they would have been in 2016. The corpus being stimulated is the same corpus agentic commerce treats as ground truth. The FTC can penalize fake reviews, Singapore can shame a hundred buyers, Amazon can block the black market at scale. None of that addresses the legal layer: authentic humans, one tap each, strategically harvested to move a number, read downstream by machines as if it meant something.

The Trust Layer That Is Not Inventory

The lesson of unBoxed 2026 is not that Amazon did something wrong. It is that every input to a purchase decision inside a marketplace is now, or is becoming, monetizable by that marketplace. The only trust signal that stays trustworthy is the one computed outside the marketplace’s economics, on evidence that cannot be bought by the seller or stimulated by a media plan.

That is the design brief GoBuy was built against, deliberately:

  • Filter before aggregating. Fake, incentivized, and low-information reviews are removed before any average or score exists, so a burst of one-tap volume changes the count, not the verdict.
  • Score quality, not quantity. The Smart Score, 0 to 100, is built on the information content of authentic reviews, not their number. A media budget can buy taps. It cannot buy substance.
  • Require persistence. The GoBuy Verified badge demands a filtered score above 80 sustained over 90 days. A campaign timed to rescue a sub-4.2 product cannot manufacture three months of history; time is the input budgets cannot compress.
  • Shortlist, don’t shelf. Seven products per category, each slot surviving filtering, instead of forty thousand results where stimulated volume can bury judgment.
  • Serve the humans and the machines. The Chrome extension puts the trust panel directly on the Amazon page, next to the rating being managed. Agents get the same filtered evidence through MCP at gobuy.ai/api/mcp, so the machine consults an independent score before checkout rather than inheriting a repaired average.

Amazon spent this week making every layer of its decision stack payable or visible on its own terms. That is what a retail media network in retreat to onsite does. The shopper, and the agent acting for them, still needs one number that nobody’s media plan can move.

Check what you are about to buy at gobuy.ai. If you build shopping agents, wire them to filtered product evidence at gobuy.ai/agent-docs.