Is Amazon Losing the Agentic Commerce Race by Choice?

Okay, I need to talk about the single most revealing thing Amazon has done this year, and it's not Rufus, it's not Alexa+, and it's not whatever internal codename they've given their latest AI project. It's an ad buy.

Amazon is blocking OpenAI's crawlers from reading its listings. It is also purchasing advertising inside ChatGPT to pull shoppers back to Amazon.com. Both things are true at once, and the combination is — I'm going to be generous here — a choice. The corporate equivalent of refusing to let someone into your house while paying to put a flyer on their windshield. Eighteen months ago, when AI-referred traffic converted like a landing page written in Wingdings, this made obvious sense. The numbers have turned.

What Prime Day Actually Showed (And It's Not the Headline Number)

US shoppers dropped a record $26.4 billion across all retail sites during the four-day Prime Day window. That's the headline. Here's the number that actually matters: shoppers arriving from AI assistants were roughly 40% more likely to complete a purchase than those coming from search, email, or social, according to Adobe's measurement.

Read that again. Forty percent.

In earlier cycles, AI-referred visitors were the least likely to convert — digital window-shoppers who'd wandered in because ChatGPT told them to look at something. The flip suggests assistants have got materially better at doing the work that used to happen across your 47 open browser tabs: narrowing a category, resolving spec questions, checking whether a product fits a constraint. By the time the shopper lands on a retail page, the deliberation is mostly finished. (Must be nice.)

Two caveats. First, base rates — agentic traffic is still under 1% of visits across every major retailer. A 40% conversion premium on a rounding error is still a rounding error in absolute revenue. Second, selection effects — high-intent shoppers might just be the people who bother using an assistant; the assistant could be sorting intent rather than creating it. Adobe's data can't separate the two.

But directionally, the point stands. Assistants have absorbed the research stage of the funnel — precisely the stage Amazon has spent two decades monetising through sponsored placements. The highest-converting shoppers are now arriving from a surface Amazon has actively chosen to exclude itself from.

Amazon's share of that agentic traffic is the lowest among major retailers, at roughly 0.4% per J.P. Morgan. That's not an accident — it's the designed result of blocking crawlers, suing Perplexity, and winning a preliminary injunction against it operating in logged-in areas of the site. Fair enough, except the shoppers who convert best during your own biggest sale of the year are arriving via the thing you blocked.

The Case That Amazon Is Fine (And It's Not a Bad Case)

The defence of Amazon's position is stronger than the hot-take merchants allow, and it rests on three things.

First, Rufus is working. Folded in May into Alexa for Shopping, it's passed 250 million users, with monthly actives up more than 115% year on year. Customers who use it are reportedly more than 60% more likely to buy, and AWS has attributed close to $12 billion in incremental sales to it last year. You can quibble with the attribution methodology — "used the assistant" versus "was already going to buy" is the same selection problem that troubles the Adobe figure — but a first-party assistant embedded in the purchase environment isn't a defensive crouch. It's a competing product with distribution most rivals can't buy.

Second, the economics. Amazon's ad business is expected to bring in around $83 billion this year, and on high margins, it accounts for roughly a third of operating income. That revenue depends on shoppers browsing an Amazon-controlled surface. Handing the discovery decision to an external assistant doesn't just risk a transaction — it disintermediates the auction that funds a third of the company's profit. No rational management team opens that door before it's built a replacement. (I wouldn't either, and I panic when someone rearranges my desk.)

Third, Andy Jassy's capability argument. Third-party agents, he's said, lack a shopper's purchase history and frequently get pricing wrong, and shoppers will gravitate toward whichever assistant knows them best. This is not obviously wrong. Amazon holds two decades of transaction history, household-level context via Alexa, and real-time inventory and pricing truth. An external agent working from crawled or fed data is operating on a lossy copy of reality.

The Case That Amazon Is on the Back Foot (And It's Also Not a Bad Case)

The problem with all three defences is they're arguments about today, and the risk is about defaults.

Walmart, Target, and others have opened their catalogues to outside assistants. They're not doing this because their AI teams are better. They're doing it because a challenger has less to lose from a channel that might cannibalise its own funnel, and more to gain from being the retailer an assistant reaches for when a shopper asks a generic question. Every one of those interactions is a repetition. Repetitions become habits. Habits become the default answer to "where do I buy this." (It's the same reason you still google things even when you know you shouldn't. Muscle memory is undefeated.)

Google's expansion of connected apps in AI Mode is the more consequential development, and it's been under-read. Users can now link services directly inside Search — adding groceries to an Instacart basket, pulling templates from Canva, saving a playlist to YouTube Music. Strip away the consumer framing and what Google is assembling is an execution layer inside the query itself. The retailers present in that layer become the fulfilment options. Those absent become a link at best. Amazon is not in that layer. By choice.

This is where the ChatGPT ad buy becomes properly revealing. Amazon has calculated that assistant surfaces are already valuable enough to pay for access to — but is renting that access as an advertiser rather than participating as inventory. It's a position with a natural expiry date. If assistant-mediated discovery keeps compounding, Amazon ends up buying its own demand back at auction prices set by someone else, having declined the option to be present organically. Jassy has acknowledged the tension, saying Amazon is in talks with AI companies to find an arrangement that works for all parties. Talks are a hedge, not a strategy. (I say this as someone who has described "we should probably look into that" as a plan on more than one occasion.)

Moonraker and the Cost of Catching Up

Which brings us to the internal picture, and honestly, it's the least reassuring part.

Amazon is developing an unreported Alexa project codenamed Moonraker — yes, like the Bond film, and about as smoothly executed so far — designed to handle multi-step requests like "book me a ride and text my friend" rather than single commands. This is Amazon building a genuine agent rather than a conversational search box, and it's the correct ambition. It's also, according to internal planning documents, the highest-cost new initiative in the Alexa+ overhaul, projected to exceed $100 million in GPU costs this year, with at least one document floating delay or scale-back as a way to relieve budget pressure.

Two things follow. First, Amazon is treating agentic capability as a line item to be managed rather than a land-grab to be won, at a moment when its competitors are treating it as existential. The company that spends over $100 billion a year on capex is reviewing a $100 million GPU budget for the product that determines whether it owns the interface to its own store. That's a signal about internal conviction, not affordability.

Second — and this is the bit that made me do a proper double-take — planning documents show Amazon using an Anthropic model for advanced reasoning and visual functions within Moonraker. Amazon's proprietary assistant, the moat against external agents, is partly built on a model from a company that also ships a consumer assistant capable of shopping. The vertical control of the interface does not extend to vertical control of the intelligence powering it. That's like building a fortress and renting the walls.

Add the operational history: repeated delays to the Alexa+ rollout, beta testing that surfaced hallucinations and inconsistent responses, and internal concern the team has overspent on the underlying models. This is a company building the right thing, slowly, while arguing publicly that everyone else's version isn't good enough yet.

Is It Too Late?

No — but the window is narrower than the balance sheet suggests, and the risk isn't the one Amazon appears to be managing.

Amazon can almost certainly build a better shopping agent than anyone else, because the data required to do it lives inside Amazon. What it cannot manufacture is the reflex to ask. The threat is not that ChatGPT or Gemini out-recommends Rufus on an Amazon product page. It's that the question never reaches an Amazon surface at all, because the shopper asked an assistant, the assistant answered, and a retailer with an open catalogue completed the sale. Amazon's assistant only wins the interactions that start on Amazon. That's a shrinking denominator problem, and no amount of first-party personalisation solves it.

The strategically coherent move is the one Jassy has hinted at without committing to: negotiated participation on Amazon's terms — structured feeds, authenticated agent access, defined commercial arrangements — rather than the current binary of blocked crawlers plus paid placement on rival surfaces. Walmart and Shopify have shown that controlled integration is possible without ceding checkout. Amazon has more leverage to negotiate such terms than any retailer alive. It's spending that leverage on litigation and ad buys instead.

What This Means for Sellers Like Us

The practical read hasn't changed much, but the urgency has.

Amazon remains where the money is, and will for years. But if you're optimising exclusively for Amazon's internal surfaces, you're optimising for the share of demand that still originates there — and the highest-converting cohort in the market right now is arriving from somewhere else entirely. Amazon-only presence increasingly means invisibility in the layer where purchase consideration is actually being resolved.

That argues for the same dual-surface discipline we've talked about before, now with better evidence behind it. Structured, complete, machine-legible product data everywhere you sell, not just in Seller Central. A presence on at least one catalogue that assistants can actually read. Content that answers comparative and constraint-based questions, because that's the form the query now takes. And an honest accounting of how much of your revenue depends on a discovery mechanism that a single platform is currently choosing to fence off from the rest of the internet.

Amazon is betting it can be the destination and the assistant simultaneously. It may well be right. But the shoppers who converted best during its own biggest sales event of the year got there by asking something else first.

• eBay adds eBay Live to its Partner Network, letting affiliates earn commissions on livestream shopping events, with new Live tracking fields in the Transaction Detail Report. The move comes as Whatnot and TikTok pressure eBay Live.

• OpenAI adds custom audience uploads and bid adjustments to ChatGPT Ads, moving closer to Google and Meta style first party targeting. Match rates run partial, and AI generated ad suggestions push whole ad groups back into review.

• Radial finds 44% of US consumers first buy a new brand on marketplaces versus 10% on retailer sites and 3% on social. TikTok Shop posts the worst delivery reliability at 16%, and 53% ditch a brand when paid expedited shipping fails.

• Walmart says Sparky users spend 35% more and build baskets 35% larger, with weekly active users up 110% quarter over quarter and Sparky attributed goods sold up 150% in Q1. Units sold via Sparky quadruple as replenishment rolls out.

• Dana Suskind's new book warns AI toys and tutors risk becoming the ultra processed food of childhood, splitting kids into those with human attention and those with artificial substitutes. No federal law currently regulates AI toys.

• Ozone data shows publisher ad request volumes fall up to 41% year over year in Q2 as zero click search strangles referral traffic. Rising eCPMs, up 30% in the UK, absorb some of the hit, but US programmatic spend still drops 44%.

• Long horizon agent prompting hinges on a pseudo formal brief: exact success predicate, enumerated non counting outcomes, and named failure modes for auditors. Persistence without a verification gate yields confident non solutions.

• OpenAI tells enterprise buyers to judge AI on useful work per dollar, not token price, citing GPT-5.6 hitting better coding scores with 54% fewer output tokens. New Admin Console analytics track credit spend by user and model.

📸 Kondor for iOS - ChatGPT for your phone camera

🖥️ Respell - Combine no-code workflows, agent chat, and AI suggestions to automate work.

🌐 Permar - Generate a conversion-rate-optimized landing page with a single prompt.

🖌️ Mockey - Generate high-quality product mockups with 1000+ templates.

📃 Huntr - Helps you organize your job search, create tailored resumes and cover letters in seconds and fill out application forms in one click.

😍It was great, all I needed

🙂Good, not great

🙄It sucked