The AI that shops for you threatens the advertising business of Google, Meta and Amazon

The article, written by Alex Salkever for Forward Future, opens with a recent anecdote: Amazon has thrown Meta's new shopping agent, Muse, out of its store, barring it from browsing the catalogue or even crossing the platform's 'virtual threshold'. From there, the author builds a thesis: the rise of AI agents that shop on the user's behalf could dismantle the advertising model that has made Google, Meta and Amazon fabulously profitable over the past twenty-five years.

The piece traces the model's origins to the famous turn-of-the-millennium 'monkey' banner ('Punch the monkey and win'), which already sensed that the internet's real product was never the banners but the user's attention. On that idea, the article argues, was built the most sophisticated advertising machine in history: Google waiting for someone to type 'cheap flight to Paris' so it could sell them to an airline, and Meta going one step further by manufacturing desire before it existed, showing an attractive person in nice shoes in Tuscany.

The figures the article offers for 2025 are striking: Google's advertising generated $294.7 billion, almost three quarters of Alphabet's revenue, while the company produced $164.7 billion in operating cash flow. At Meta, advertising accounted for $196.2 billion of $201 billion in total revenue. Amazon took in $68.6 billion from advertising, 22% more than the previous year ($56.2 billion), compared with $46.9 billion the year before that. Walmart, with a far younger advertising business, reached almost $6.4 billion, up 46%. Across the whole market, digital advertising in the US hit $294.6 billion in 2025 (+13.9%), of which $63.4 billion came from 'commerce media' alone.

The article identifies a pattern: companies that originally sold physical products have gradually shifted their business towards renting out their own digital shelf space, because it is far more profitable than manufacturing, shipping or taking back merchandise. Walmart is the chosen example: besides selling detergent, it charges brands such as Procter & Gamble to appear in front of the user searching for 'laundry detergent'. Apple, for its part, does not break out its advertising figure — it is folded into 'Services' — but acknowledges that advertising helped drive growth in that division, which billed $109.2 billion with a gross margin of $82.3 billion, above 75%. Amazon, with its logistics machinery of warehouses and vans, has on top of it a far cleaner advertising business and, as the company itself tells investors, that business helps improve its operating income.

The core of the argument comes with the example of an AI agent asked to 'get me the same pair of black shoes, size 10½, that I bought last year, delivered by Friday and for no more than $160'. That agent does not need to see sponsored results, 'customers also bought' carousels or the 'Amazon's Choice' badge: it can go straight to the exact product. The author describes it as 'an ad blocker with a checking account', able to complete the purchase without ever displaying the page it has blocked. That is why he reads Amazon's ban on agents such as Muse or Perplexity — which, he stresses, were acting with the explicit permission of legitimate users, not as malicious bots — as a fundamental defence of its $68.6 billion advertising business, which depends on controlling the final metres between 'I want something' and 'I've bought it'.

From there, the article explores possible business models for an agent economy. One is for advertisers to pay agents commissions for being selected (Sony offers $12, Bose $15, Apple $20), which would in practice reinvent search advertising, with the added twist that the sponsored result can now handle the user's credit card directly. The author raises the uncomfortable question of who the agent then serves, especially if the same company operating the agent also collects the commissions. He cites Robert Sapolsky in passing to note that this would mean outsourcing one's own dopamine reward circuit to a machine.

Another alternative the article weighs is that agents refuse advertising for recommendations and instead solicit competing offers from retailers: a reverse-auction model in which different stores compete on price, shipping or loyalty points against a request that has already been decided ('I want some Jordans'). This would be good for the consumer, but it would turn the agent into a relentless, immortal 'head of purchasing' that negotiates with every supplier at once without embarrassment — something, the author warns, retailers would welcome with the same enthusiasm taxi drivers showed Uber.

The article distinguishes between demand that has already formed (procurement, where advertising loses its purpose) and demand that does not yet exist. That is where it sees the great strategic value of Meta and Instagram: they do not wait for someone to search for 'best hotel in Kyoto', they generate the desire by showing a photograph that triggers a buying impulse the user did not know they had. Meta acknowledges that online commerce was the main driver of growth in its $196.2 billion of advertising. According to the author, Muse is especially interesting because it would fuse that desire-creation engine with purchase execution in a single product, running from 'what jacket is that?' to 'buy it for me in my size, under $300' without passing through an intermediate web page, allowing Meta to 'eat' the traditional advertising funnel rather than merely compete with it. The piece recalls that Meta has been preparing for this for years: Apple's privacy changes (ATT) cost it close to $10 billion in 2022, forcing it to improve its recommendations with less third-party tracking data — a capability that could now prove decisive.

The article also stresses that agents do not have to wipe out all advertising to damage these companies' profitability: it is enough for them to eliminate part of the sponsored results, which tend to be among the most profitable revenue. Amazon does not break out its advertising margin, but tells investors that the business contributes to the growth of its operating income, and Walmart notes that advertising improves its business mix — which explains, in the author's view, why Amazon resists opening the door to autonomous agents while other companies such as Google or Walmart do the opposite, in a dynamic the piece compares to the 'complement versus disruption' strategy Google used with Kubernetes against Amazon's early dominance in the cloud.

Finally, the article points to the technical infrastructure layer emerging behind these agents: MCP as the standard for agents to connect to tools and services, Google's Universal Commerce Protocol (UCP) — developed with Shopify, Etsy, Wayfair, Target and Walmart, and backed by more than twenty companies — to standardise discovery, payment and order management, and AP2, the agent payments protocol, also from Google. Even Amazon has reportedly adopted MCP and similar protocols, acknowledging that the future runs through autonomous transactions between agents. The author connects this to Ben Thompson's aggregation theory: if enough retailers expose their catalogue and terms directly through these protocols, an agent would no longer need to rely on Amazon's home page to find what the user is looking for, but could go straight to whoever is willing to sell it on those terms.

It should be noted that the portion of the article available for this analysis cuts off abruptly, mid-sentence, just as the author begins to describe how an agent could 'traverse MCP, UCP, A2A and whatever protocol is needed to find' a product or supplier. There is therefore no record of the article's final conclusion or of whether the author closes with a specific forecast about the outcome of this clash between advertising and agentic commerce; this summary is limited strictly to what appears verbatim in the text received.