Yesterday this publication posed five questions for Amazon’s third-quarter earnings call, all about the agentic commerce era Amazon is building: how much of Alexa for Shopping’s answers are sponsored, whether the company will ever audit its review corpus, how it chooses which outside agents get blocked, what Rufus does with inauthentic reviews when it summarizes them, and who is liable when a sponsored recommendation rests on manipulated ones.
Last night the results arrived. Four of the five questions were met with silence, and no analyst asked any of them. But the call did answer the third question by accident, in a sentence of striking candor, and it disclosed three numbers about Amazon’s agentic shopping machine that had never been public before.
The short version: the machine is enormous, the furnace is enormous, the trust layer still does not exist.
The Print: A Record Margin Sitting on a Record Burn
Per CNBC’s report, here is how the quarter landed against LSEG consensus:
- EPS: $1.89 vs. $1.96 estimated, a miss. Net income was $20.69 billion against $21.19 billion a year ago, though the comparison is distorted by the $9.5 billion pre-tax Anthropic gain booked in Q3 2025.
- Revenue: $201.94 billion vs. $202.31 billion estimated, up 12 percent year over year, or 13 percent excluding foreign exchange.
- AWS: $45.98 billion vs. $45.29 billion expected, up 39.3 percent, its fastest growth in nearly five years and a sixth consecutive quarter of acceleration.
- Advertising: $19.64 billion vs. $20.82 billion expected, up 11 percent. More on this below.
- Online stores: $68.47 billion vs. $69.68 billion expected, up just 2 percent, against nearly 10 percent growth a year ago.
- Operating income: $28.1 billion, up 61 percent, at a 13.9 percent operating margin that CFO Brian Olsavsky called the highest in company history.
- Free cash flow: negative $42.4 billion for the trailing twelve months, a 386 percent year-over-year swing and the second consecutive negative quarter.
- Capex: $77.3 billion in the quarter against a FactSet consensus of $58.7 billion, with the full-year forecast held at roughly $220 billion.
- Q4 guidance: $240 to $247 billion in sales, a midpoint below the $244.5 billion consensus, with operating income of $28.5 to $33.5 billion.
The stock fell more than 6 percent in extended trading. That reaction gets its own reading later in this piece: what the market punished was not the agentic strategy but the financing of it.
The Machine: AWS at 39.3 Percent and a $578 Billion Backlog
Every narrative about AI agents shopping and transacting ultimately terminates in physical infrastructure, now being installed at a pace that reorders the balance sheet. Per the call transcript summaries from TradingKey and Benzinga:
- AWS reached a $184 billion annualized run rate, with backlog at $578 billion and growing at a triple-digit rate year over year.
- Amazon is installing an AI accelerator approximately every 14 seconds.
- Bedrock token volume rose nearly 17-fold, Fortune 500 spending on Bedrock increased almost sevenfold, and inference already represents well over half of AWS AI revenue.
- Commitments for Amazon’s custom Graviton and Trainium silicon each grew more than 20-fold year over year, and customers using that silicon saved more than $2 billion in the quarter. Trainium becomes available outside AWS in late 2027.
- Jassy said on the call that AWS will be a $1 trillion revenue business within five to ten years, and in the release he projected AWS revenue growth of “over 40%” in both 2027 and 2028.
Against that, the company expects to spend roughly $220 billion in cash capital expenditures this year and, in Jassy’s words, “spend meaningfully more in 2027,” while claiming a “clear line of sight to attractive financial returns.”
Hold those two facts together and you have the agentic paradox in miniature: a record 13.9 percent operating margin and a negative $42.4 billion free cash flow, printed in the same quarter of the same company. The margin says the business works. The cash flow says it is being rebuilt underneath its customers on the assumption that the agentic future arrives on schedule. Agents that shop, compare, and buy are the demand signal that justifies a data center accelerator every fourteen seconds.
The Miss That Matters: Advertising Decelerated Exactly When the Thesis Needed It
The line that should worry anyone building on Amazon’s agentic commerce stack is the one Wall Street scrolled past fastest. Advertising revenue of $19.64 billion missed the StreetAccount consensus of $20.82 billion by roughly $1.2 billion and grew 11 percent, against 26 percent growth and $19.8 billion in the second quarter. It is the slowest ad print in years for the company’s quiet margin engine.
Management attributed the shortfall primarily to Prime Day moving from the third quarter into the second to avoid the FIFA World Cup, a shift it said created an 850-basis-point impact on the quarter’s advertising growth. CFO commentary in the TradingKey transcript summary promises a return to “high teens” advertising growth in the fourth quarter, alongside newly launched AI-driven full-funnel campaigns spanning sponsored ads, display, video, and streaming TV.
Two observations. First, even granting the Prime Day distortion in full, the trajectory now depends on holiday performance landing exactly as guided, in a quarter where the sales midpoint itself missed. Second, and more important for this publication’s argument: last quarter Amazon proudly disclosed that shoppers who click a Sponsored Prompt convert 48 percent more often and spend 21 percent more, per the company’s own Q2 commentary. This quarter, with the ads line actually missing, the Sponsored Prompt disclosure went silent. The stat appears when it flatters. It disappears when the segment wobbles. That is what a metric managed as marketing rather than disclosure behaves like, and it is why no outside party should build a trust model on the marketplace’s accounting of itself.
The Three Numbers Nobody Expected
Buried in the retail section of prepared remarks, per the Benzinga transcript, were the first real KPIs Amazon has ever given for its agentic shopping surface:
- More than 400 million customers have used Alexa for Shopping in the last twelve months.
- Customer interactions with it are up 570 percent year over year.
- US customers using Alexa for Shopping spend 50 percent more per order than customers who do not.
The assistant, per the same remarks, researches products, compares options, checks a full year of price history, and tracks deals and price drops. Then came the sentence that accidentally answered yesterday’s third question:
“Though we anticipate working with horizontal agents, we believe Alexa for Shopping is going to be the primary shopping agent customers use to shop Amazon.”
Read that against the quarter’s context. Amazon blocked Meta’s Muse from its retail site in September, went to court with Perplexity, and shut out independent buying agents while opening an advertising MCP server so outside agents can query its ad inventory. The posture, now on the record from Amazon’s own earnings call: third-party agents will be tolerated as discovery surfaces while purchase intent is funneled through Amazon’s own assistant, the one embedded in the marketplace that also sells the ads, ranks the products, and owns the review corpus.
Notice also what the assistant verifies and what it does not. Alexa for Shopping checks a full year of price history on the customer’s behalf. Price is fully instrumented. Review authenticity is not. The assistant that spends 50 percent more per order is doing so inside a product evidence layer that Fakespot-derived research has estimated is more than 40 percent inauthentic, a figure Amazon disputes and has never countered with an audited integrity metric of its own. Higher spend per order through an unverified evidence base is not a trust milestone. It is monetization ahead of verification, which is the entire agentic trust problem compressed into a single KPI.
The Quarter’s Agentic Data Amazon Did Not Report
While Amazon kept its agentic disclosure to three flattering numbers, the outside world spent the quarter measuring the phenomenon in detail.
FirstPageSage’s October dataset found that agent-completed orders rose from 0.27 percent to 0.68 percent of US ecommerce orders between the second and third quarters of 2026, the largest quarterly jump in their dataset. A doubling-plus in one quarter off a small base: the classic bend of an S-curve. The same research found 61.4 percent of shoppers now use AI for product research while only 6.8 percent let an agent finish a purchase: the funnel narrows sharply exactly where trust has to be established, between choosing and paying.
Forter’s agentic order data, published October 23, fills in the mechanism. Agentic orders grew 24 percent month over month from August to September, and 48 percent counting newly detected Instinct and Grok Bot. Screen-scraping agents, which browse the open web like humans, went from 1 percent of agentic orders in late July to 70 percent in the week of October 12, overtaking the official commerce protocols because they can shop anywhere those protocols have not reached. The average agentic order is $226, and agents skew toward considered purchases: only 27 percent of agentic orders are $50 or less.
The rails kept arriving in October, none of them Amazon’s. Walmart introduced Agent Profiles and an Agent Gateway, giving customers control over which agents may shop for them. Mastercard opened early access to Agent Pay for Machines after demonstrating agent-initiated payments, adding a score estimating how likely a transaction was started by an agent. Visa launched an invoice-payment pilot with BILL and Billtrust. Anthropic’s new Claude Duet handles business purchases. OpenAI, having shelved commission-bearing Instant Checkout in March, now charges merchants nothing and routes purchases to merchant-owned checkout: discovery, not checkout, is where platform value sits. Google’s Universal Cart rolled out across traditional search and AI Mode through the month, per Retailgentic’s October roundup.
And the trust deficit these rails inherit is measurable too: a Wildfire survey found only 5 percent of US consumers would trust an AI platform itself to process an agent payment. Every horizontal player is building authenticated buying. None of them, including Amazon, is building verified products. The identity of the buyer is becoming the best-instrumented fact in commerce. The authenticity of what is being bought remains the least.
Scorecard: Five Questions, One Accidental Answer, Four Silences
Yesterday’s five questions, scored against the actual call:
- What share of Alexa for Shopping answers include a Sponsored Prompt, and what is the disclosure? Not addressed. The 48 percent conversion lift from Q2 was not updated, and no disclosure policy was announced.
- Will Amazon publish an audited review integrity rate? Not addressed. Again.
- How does Amazon choose between agents blocked at checkout and its own MCP ad partners? Answered by accident. Horizontal agents will be worked with; Alexa for Shopping “is going to be the primary shopping agent customers use to shop Amazon.” The two-tier system is now official company language.
- What does Rufus do with incentivized or inauthentic reviews when summarizing? Not addressed.
- What is the liability posture when a sponsored agentic recommendation rests on manipulated reviews? Not addressed, weeks before the first holiday season in which it will happen at measurable scale.
The analyst Q&A, per Investing.com’s transcript coverage, spent its time where the money visibly moves: AWS capacity reservations into 2029, whether Bedrock can be as large as EC2, margins, capex, fuel costs, grocery. Fair questions all. But an earnings call cannot price integrity risk, because integrity has no line item, no consensus estimate, no analyst model. What is not on the spreadsheet does not get asked, and what does not get asked does not get disclosed.
What the Market Punished Instead
The after-hours selloff of more than 6 percent was not a referendum on agentic commerce. It was arithmetic: a seven-cent EPS miss, a capex print of $77.3 billion against $58.7 billion expected, a fourth-quarter sales midpoint below consensus, higher fuel and linehaul costs guided into the peak quarter, and a second consecutive quarter of negative free cash flow now at negative $42.4 billion trailing. Add Amazon expensing most Leo satellite costs in Q4 as capitalization slips to 2027, and the market had its narrative: the AI buildout’s bill is arriving faster than its revenue.
Beneath the headline, retail told its own quiet story. Worldwide paid units grew 12 percent while online store revenue grew 2 percent: customers are buying far more, far cheaper things. Grocery and everyday essentials grew roughly three times faster than the rest of the store, approached $250 billion in trailing-twelve-month gross sales, and now represent about half of units sold. The mix is racing toward high-frequency, low-ticket essentials, compressing revenue per unit and raising the strategic value of everything that monetizes the decision rather than the item: advertising, sponsored placements, and agentic surfaces like Alexa for Shopping, where the disclosed 50 percent higher spend per order is exactly the monetization answer a cheap-unit mix demands.
That is the strategic logic of the Sponsored Prompt era. It works only as long as the recommendation layer is trusted, and the company that owns the layer reports nothing that would let you check.
The Gap Is Still the Product
Strip the quarter down and the structure of the agentic economy is now explicit. The infrastructure layer is being financed at a $220 billion annual pace on the expectation that agents will mediate commerce. The interface layer is consolidating: Amazon has declared its own assistant primary, Walmart built a gateway for everyone else, the payment networks are authenticating the buyer, and screen-scraping agents are already placing 70 percent of agentic orders wherever the protocols have not reached. The evidence layer, the part that determines whether any of these systems are pointing at products whose ratings, reviews, and listings deserve the money moving through them, exists nowhere in any filing, any call, or any KPI.
That is the layer GoBuy builds. Fake reviews are filtered out before any score is computed, authentic substantive reviews are weighted up, and every product gets a Smart Score from 0 to 100 based on review quality, not volume, so a listing with 40,000 reviews of which a third are fake cannot outrank one with 900 verified honest ones. Agents and humans get a curated top 7 rather than a firehose, because a shortlist you can verify beats ten thousand results you cannot. The scoring is identical for every product: no sponsored prompts, no pay-to-rank. Products holding a score of 80 or higher for 90 days earn the GoBuy Verified badge, an independent signal that does not depend on the marketplace grading its own homework. Any AI agent can consult it live through the MCP server at gobuy.ai/api/mcp, and the Chrome extension puts the trust panel directly on the Amazon product page, referee-independent.
Amazon just told you its machine will be primary. Before you let any machine, theirs, Meta’s, OpenAI’s, or your own, spend your money, check what it is spending it on.
Verify before anything buys for you: gobuy.ai. Building a shopping agent? Wire in the trust layer first: gobuy.ai/agent-docs.