Emanuel: AI Agents Will Destroy Google and Amazon Ad Models

Full analysis

Ari Emanuel and Ben Thompson spent a chunk of a Stratechery conversation arguing that AI agents will gut the search-ad businesses at Google and Amazon, with Amazon's roughly $60 billion ad arm called out as acutely exposed. The claim is that once software does the browsing and buying, the ad slot that depends on a human reading a results page has nothing to sell.

What's actually being decided for a reader who buys and builds with AI: how fast to discount any revenue or ad plan that assumes a human eyeball at the end of a click. That belief is easy to undo. You can lean into agent-proof channels gradually and reverse course if adoption stalls. Nothing here sets a deadline. No pricing change, no product ship date, no contract. This is a thesis, not an event.

The Skeptic. This is a dinner-party thesis with no adoption curve bolted to it. "Agents buy for you" has been imminent since Siri in 2011. For Amazon's $60 billion ad business to crack, millions of people have to hand their credit card to software and stop looking at what it bought. That trust transfer is the whole ballgame, and Emanuel and Thompson wave past it. The social-network-safe-haven line also happens to flatter Meta, which is Thompson's favorite running story. Convenient. The threat is real on a ten-year clock. Calling it a near-term structural break is the vivid-demo problem: a slick agent video makes the shift feel closer than any usage data says it is.

The Researcher. The logic holds together and the evidence doesn't. The real question is a number nobody has: at what share of shopping done by agents does query volume drop enough to move what advertisers pay per thousand impressions? Until agents are doing a measurable slice of actual purchases, CPM curves don't move. The two-towers critique is the one serious idea here. Today Amazon matches you to products by turning your behavior and every product into vectors and finding the closest pair across giant databases. Thompson's pitch is to run that matching on your own device instead. Apple's on-device machine learning shows it's possible in principle. "Possible in principle" is not "beats Amazon's relevance." The jump from "agents act on their own" to "ad formats collapse" skips every messy step in between.

The Compute Pragmatist. The local-device personalization story has a bill nobody added up. To beat Amazon's matching on your laptop, you need either a very strong on-device chip or constant model updates streamed from a server, and the second one quietly rebuilds the central system this was supposed to kill. Apple's Neural Engine handles light inference. It cannot hold the embedding lookups across millions of product SKUs that make Amazon's ad relevance work. The agent layer costs too. Every autonomous shopping session is a chain of model calls, each one burning tokens. "Agents replace search" only pencils out when one agent run costs less than the ad revenue a human search would have thrown off. We are not there.

The Enterprise Buyer. If I run ad budget, this changes nothing I sign this quarter and everything I plan for the next three years. The part that would actually move me is the privacy pitch: run the targeting inside the user's own machine so their data never leaves, and you can promise privacy while keeping targeting precise enough to matter. That solves a real procurement headache. But who ships that device-side system? Who audits the model sitting on the customer's machine? Who indemnifies me when an agent buys something it shouldn't on a user's behalf? None of that exists. Until there's a vendor with audit logs and a contract, the agent-proof channels I can buy today are the ones Emanuel named: social feeds and live events, where a human is still watching.

The Safety Lens. The device-side ad model got zero scrutiny and deserved the most. A personalization engine living on your machine, trained on your most intimate behavior, spitting out ad targets, is a fat attack surface with a friendly privacy label. Who checks the weights? What stops that "privacy-preserving" box from becoming a surveillance endpoint with better branding? And autonomous agents making purchases scatter accountability: when an agent buys something harmful or steers a protected group away from an offer, who owns that? Diffuse liability is not a footnote. Neither Emanuel nor Thompson touched the adversarial case, which is where this idea either survives or doesn't.

Where they split. The Researcher and the Compute Pragmatist agree the thesis is coherent but can't price it: nobody has the agent-adoption number, and nobody has the inference cost that makes agent shopping cheaper than a search. The Skeptic goes further and says the trust transfer itself never arrives at scale. The Enterprise Buyer is the one who takes the device-side idea seriously, because it solves a real buying problem, and the Safety Lens is the one who says that same idea is the scariest thing in the conversation. The decision lives in that gap: the privacy-preserving local model is simultaneously the most useful and the least audited piece, and both can be true.

What this hinges on: whether agents do a real, measurable share of actual purchases before advertisers reprice search inventory, and whether on-device targeting can beat Amazon's centralized matching without quietly phoning home. Both are testable. Neither is settled. The council leans hard against the near-term framing. The ten-year direction is plausible. The "acutely at risk" timing is hype.

Prediction: Amazon's advertising revenue, reported across its four 2026 quarterly earnings releases through the Q4 2026 print in late January or early February 2027, will be higher than its 2025 advertising revenue.

Confidence: High. Agent-driven purchasing is nowhere near the volume needed to dent ad demand.

Why: The whole thesis rests on humans handing purchasing to agents at a scale that shrinks the search queries Amazon monetizes, and the story offers no adoption number because there isn't one. Amazon's ad business grew roughly 20% year over year through 2025 on sponsored products that sit inside a human shopping session, and nothing shipping in 2026 moves a meaningful share of those sessions to autonomous agents that skip the sponsored slot. For revenue to fall, millions of shoppers would have to delegate buying to software and stop seeing the ads inside twelve months, a trust transfer that has been "imminent" since 2011 and still hasn't happened. The disruption is real on a decade horizon; it does not show up in a 2026 revenue line.

Revisit by 2027-02-15: We're right if Amazon's full-year 2026 advertising revenue, from its Q4 2026 earnings release, exceeds its 2025 figure. We're wrong if it comes in flat or lower.

The interesting fight isn't whether Emanuel is right about agents. It's that he and Thompson smuggled the device-side targeting model past everyone as a privacy win, when it's the one piece an operator should want audited before it ships anywhere near a customer's machine.