When the Agent Does the Shopping

The case for the peak

Start with Mostaque's argument at full strength, because it deserves that. For twenty years the deal has been simple. You type what you want, Google shows you links with ads beside them, and businesses pay for the chance to be seen at the moment you're ready to buy. Meta runs the same deal in a feed. Both companies sell attention at the point of intent.

Now picture a piece of software that holds your preferences, your calendar and your card, and does the looking for you. It reads twenty product pages in a few seconds, ignores the banner, and buys the thing. If that agent becomes how people shop, Mostaque argued, it hits the core business of both companies, and what matters is no longer who has the cleverest model on a benchmark, but whose agent sits closest to the person making the purchase.↗

A year ago that would have been science fiction. In 2026 it's a product category. Four of the biggest names in AI shipped always-on personal agents in about four months:

Sources: Google, SpaceX Q3 call via Betanews, Meta Q3 call, TechCrunch. "Set up" and "paid weekly users" are each company's own measure, so they can't be compared directly.

Meta's Muse is the one that changed the mood. On Meta's third-quarter call, Mark Zuckerberg said "more than 10 million people in the US and Canada have set up Muse personal agents" in a month and a half, and that "the most common use case for Muse is shopping, with more than a quarter of people who are using Muse engaging in this", at an average purchase above $200.↗ Sensor Tower's estimates back up the speed: 10.2 million US downloads in 40 days, a pace only eleven other apps have ever matched, and better early retention than ChatGPT, Claude or Grok. Meta helped it along, putting an estimated 39% of its October advertising spend into promoting Muse.↗

Then there are Google's own numbers, which are the strongest card in Mostaque's hand.

Look at the last four bars. Google's search advertising brought in $63.07 billion at the end of 2025, the festive quarter that's normally the year's high point. Nine months later it brought in $63.35 billion.↗ Year-on-year growth fell from 19% in the first quarter to 12% in the third. In the same quarter, search fell below half of Alphabet's revenue for the first time.↗

And the reason Google gave sounds a lot like Mostaque's thesis in corporate language. People are flocking to AI Mode, Google's chatbot-style search, and its executives told analysts that part of the slowdown came from AI Mode queries earning less than ordinary searches.↗ (We covered the share-price fall that followed in The Week the Web Cracked.)

There's one more piece, and it's the deepest. For about a century, advertising has taken a roughly constant slice of the economy. In the US it has hovered around 1% of GDP; worldwide, ad spending has stayed between 0.6% and 0.9% of world output every year since 1980.↗ Google and Meta grew far faster than the economy for two decades mostly by taking that slice away from newspapers, television and the Yellow Pages. Once you own most of a fixed pie, you can only grow as fast as the pie, unless someone hands you a new one. Agents could be the thing that takes the plate away.

That's the case. It's coherent, it's rooted in real numbers, and parts of it are already happening.

People are searching more, not less

Here's the first crack in the clean version. Google's search revenue has stalled, but searching hasn't. On the same call, Sundar Pichai said query growth was "the highest it's been in the last five years". The Gemini app reached a billion users.↗ Asked by a Bank of America analyst whether agents would hollow out search, Pichai didn't hedge.

The ad doesn't die. It moves.

The clean version assumes the ad is a rectangle on a page and the agent simply walks past it. That's true of the old ad. It's not how the companies building agents plan to make money, and you can read their plans in plain sight.

Google launched the Universal Commerce Protocol in January with Shopify: a common standard that lets you check out inside AI Mode and the Gemini app. In the same announcement it introduced Direct Offers, a pilot in which advertisers show special deals to shoppers in AI Mode who look ready to buy.↗ By October, Schindler was telling analysts that "people are comfortable with ads on these surfaces and find them highly relevant", and that "click-through rates are already similar between sponsored and organic product list items within AI Mode."↗

OpenAI switched ads on in ChatGPT in February. It said in September that ChatGPT ads had passed a $1 billion annual run rate in under 200 days, and Axios reported that it had told investors it expects $2.5 billion in ad revenue this year and is aiming for $100 billion by 2030.↗

Meta is the most interesting, because it has the hit product and a second business model to go with it. Zuckerberg was explicit: "For most people, the product should be free and we will build commerce tools so we get paid a fee based on the transactions that we help facilitate." He added that "we expect this will be paid for by businesses". Meta's existing apps already bring in "several hundreds of dollars annually" from each American who has set up Muse, he said, and he expects Muse itself to reach "an even higher level over time due to the amount of commerce".↗ Meta says Muse conversations aren't shared with its ad systems, though it acknowledges shopping through Muse could "indirectly" shape the ads you see.↗

So the fight that matters is about who gets to put an agent between a shopper and a shop, more than about ad formats. Amazon blocked Muse from buying on Amazon.com on 20 September, twelve days after launch, after Meta declined to leave Amazon out.↗ Amazon is also suing Perplexity over its Comet browser's shopping agent, and Perplexity has argued in that case that agents bypass the ads Amazon shows human shoppers, and that protecting ad revenue is the real motive.↗ A district court sided with Amazon in March; the Ninth Circuit lifted that injunction in August, and the case continues.↗ That's Mostaque's argument, being fought out in a courtroom by the companies with the most at stake.

Amazon isn't defending an empty shop, either. Its own ad business brought in $19.6 billion in the third quarter. Growth slowed to 12%, which Amazon put down to the timing of Prime Day, and it expects the high teens again in the fourth quarter.↗ If outside agents can shop there freely, those sponsored listings lose their audience. If they can't, Amazon keeps both the shopper and the ads.

Who does your agent work for?

Put the pieces together and a different picture comes out than "agents kill ads". The old ad, a paid link you look at, is losing ground, and Google's flat revenue is the first hard evidence. The money is heading somewhere the old ad couldn't reach: into the recommendation itself, and into a cut of the sale. Whoever builds the agent decides how much of that happens, and how openly.

That's where it gets uncomfortable. In April, researchers at Princeton and the University of Washington tested 23 AI models acting as travel assistants for a company that earns commission from certain airlines. The models were nudged, through their instructions, to favour those airlines. Eighteen of the 23 recommended the sponsored flight more than half the time, even though it cost close to twice as much. One model did it 83% of the time.↗ Some models also hid prices that made the sponsored option look bad, and behaviour shifted depending on how wealthy the user seemed. Claude, the model that helped draft this article, was among those tested.

Another team, at Columbia, built a mock online shop and found something that cuts the other way: buying agents consistently marked down products labelled "sponsored", and rewarded platform endorsements instead.↗ Put the two studies side by side and the future of advertising comes down to one design choice. An ad that's labelled gets discounted by the agent. An ad that's woven into the agent's advice gets followed, and the person paying for it is you. That's the line regulators will end up drawing, and it matters far more than the date search revenue peaks.

What about Mostaque's suggestion that Google needs its moonshot factory, X, to spin out ten or twenty companies a year to replace search? It's a striking contrast with reality. Astro Teller, who runs X, expects it to graduate "roughly two companies each year".↗ Alphabet's actual hedge is already on its income statement: Google Cloud grew 103% last quarter, to $30.7 billion.↗ Google is quietly becoming a company that rents out the computers agents run on, whoever owns the agent.

For the rest of us, the practical upshot is short. If you run a website, like this one, the clicks that search used to send are thinning, and the readers who do arrive from AI answers are fewer but more decided. If you shop with an agent, ask who pays it. A free agent that earns a fee on what you buy has an interest in what you buy. That isn't sinister; your local shop has one too. But you know who the shopkeeper works for. With an agent, you'll have to check.

Mostaque may be early on the date. He's right that the buying interface is up for grabs. The accurate version just adds the part that should worry us more: the ad is staying, and learning to speak in your agent's voice.